nature neuroscience

Article                                                                                             https://doi.org/10.1038/s41593-026-02345-6


Human hippocampal ripples tune cortical
responses based on predicted uncertainty

Received: 11 February 2026                         Darya Frank 1,2 , Stephan Moratti 3, Robin Hellerstedt 1,
                                                   Johannes Sarnthein 4,5, Ningfei Li 6, Andreas Horn 6,7,8, Lukas Imbach5,9,
Accepted: 20 May 2026
                                                   Lennart Stieglitz 4, Antonio Gil-Nagel 10, Rafael Toledano10,
Published online: xx xx xxxx                       Karl J. Friston 11 & Bryan A. Strange 1,12

    Check for updates
                                                   To encode information efficiently, our perceptual system should detect
                                                   when situations are unpredictable (that is, informative) and modulate
                                                   brain dynamics to prepare for encoding. Under uncertainty, there is an
                                                   increased need to generate predictions about upcoming information, a
                                                   process that has been proposed to require coordinated activity between the
                                                   hippocampus and neocortex. Here we show, with direct recordings from the
                                                   human hippocampus and visual cortex, that after exposure to unpredictable
                                                   visual stimulus streams, hippocampal ripple activity increases in frequency
                                                   and duration before stimulus presentation. Prestimulus hippocampal
                                                   ripples suppress changes in visual cortex gamma activity associated with
                                                   uncertainty and modulate poststimulus prediction error gamma responses
                                                   in higher-level visual cortex to surprising stimuli. We reveal a function
                                                   of hippocampal ripples in facilitating the propagation of visual stimuli
                                                   based on the expected information gain. These results, therefore, link
                                                   hippocampal ripples with predictive coding accounts of neuronal message
                                                   passing and precision-weighted prediction errors, revealing a mechanism
                                                   relevant for perceptual synthesis and subsequent memory encoding.


An efficient recognition system should be able to prioritize the infor-     there is a mismatch between the predicted and the observed input, a
mation that will constrain or inform perceptual representations1 and,       bottom-up prediction error is returned to update or revise the source
eventually, be accumulated or retained. One way to facilitate this is       of predictions at a higher hierarchical level4,5. Crucially, prediction
to continuously generate predictions about upcoming inputs and to           errors are weighted by the level of uncertainty (that is, their precision)
retain information—that is, revise beliefs—when these predictions are       associated with the given context6, balancing top-down and bottom-up
violated. Prediction is considered central in this account of the brain,    information streams to scale the influence of prior predictions and
by building a generative model of the world to minimize prediction          sensory evidence, respectively7. This is sometimes framed in terms of
error when sampling the sensorium2,3. In predictive coding formula-         precision-weighted prediction errors that instantiate the Kalman gain
tions, predictions are transmitted in a top-down manner, and, when          in Bayesian filtering formulations of predictive coding1,4. The implicit

1
 Laboratory for Clinical Neuroscience, Centre for Biomedical Technology, Universidad Politécnica de Madrid, IdISSC, Madrid, Spain. 2Andrew Mayes
Centre for Cognitive Neuroscience, University of Manchester, Manchester, UK. 3Department of Experimental Psychology, Complutense University
of Madrid, Madrid, Spain. 4Department of Neurosurgery, University Hospital and University of Zurich, Zurich, Switzerland. 5Neuroscience Center
Zurich, University of Zurich and ETH Zurich, Zurich, Switzerland. 6Movement Disorders and Neuromodulation Unit, Department of Neurology,
Charité—Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Berlin, Germany. 7Department
of Neurosurgery, Harvard Medical School, Massachusetts General Hospital, Boston, MA, USA. 8Brigham & Women’s Hospital, Center for Brain Circuit
Therapeutics, Boston, MA, USA. 9Swiss Epilepsy Center, Klinik Lengg, Zurich, Switzerland. 10Epilepsy Unit, Department of Neurology, Hospital Ruber
Internacional, Madrid, Spain. 11Wellcome Trust Centre for Neuroimaging, Institute of Neurology, University College London, London, UK. 12IDG/McGovern
Institute for Brain Research, Peking University, Beijing, China.  e-mail: darya.frank@manchester.ac.uk; bryan.strange@upm.es


Nature Neuroscience
Article                                                                                                https://doi.org/10.1038/s41593-026-02345-6

encoding of uncertainty lends a dual aspect to predictive processing           While ripples represent possible outcomes, there is mixed evidence
that encompasses both the predictions of a particular sensation and            regarding whether they influence subsequent behavior48–50. Notably,
predictions of its predictability (that is, precision) that modulate the       the functional role of ripples is often examined by using navigational
influence of the ensuing prediction error8–13. Notably, these genera-          tasks in rodents, which heavily tap episodic memory and are biased
tive properties of predictive processing rely on ongoing integration           by the provision of reward at the goal location. In humans, although
of sensory inputs with internally generated, experience-dependent              there is emerging evidence for ripples supporting memory recall
sequences, and are therefore thought to involve hippocampal–                   processes51–53, it remains unclear to what extent they have a role in
neocortical interactions14–16.                                                 prediction, in the absence of memory demands.
      A hippocampal role in prediction is likely related to its function             We hypothesized that high-order predictions—namely, predic-
in extracting statistical regularities17–20 that can be applied to new         tions of predictability or precision—would be generated by the hip-
situations21,22. Therefore, the hippocampus should represent the               pocampus before stimulus presentation, as a function of uncertainty
expected information gain of an event before it occurs—as a function           (that is, predictability), and subsequently modulate cortical process-
of predictability16,20—and estimate the validity of the prediction upon        ing, or lower-level prediction errors. Ripples are a promising candidate
observation of the event23. Indeed, the hippocampus has long been pos-         to mediate the requisite precision weighting of prediction errors. We
tulated to hold a cognitive map that is used to form predictions about         tested this hypothesis using intracranial local field potential recordings
upcoming inputs24–27. Such predictive information follow successor-like        from patients with epilepsy to examine how the hippocampus and ven-
representation28,29, for example, in place cell firing30. The cortex may       tral visual stream regions (the occipital cortex and the fusiform gyrus)
also have its own predictive role through communication across deep            implement predictive processing. We used a simple paradigm—free
layers feeding-back predictions to the superficial layers of preceding         of any explicit demand on episodic memory—in which participants
regions in the processing hierarchy29,30. Mechanistically, predictions         were presented with a sequence of colored shapes and performed a
and their violation have been associated with spiking activity31 and           visuomotor selection task, given a target stimulus and four options
oscillatory dynamics in the cortex32,33. Specifically, gamma-band activ-       presented on-screen simultaneously. The probability distribution
ity has been associated with bottom-up prediction errors (reflecting           of stimulus presentation varied across task blocks, allowing for the
surprise signals from primary sensory cortices) and alpha/beta oscil-          quantification of stimulus-bound information-theoretic measures of
lations with top-down predictions (from higher-level regions such as           entropy (that is, uncertainty or unpredictability of an outcome before
prefrontal cortex)34–36.                                                       it occurs) and self-information (that is, surprise or violation of predic-
      However, the mechanism through which predictions of future               tions reflecting the improbability of a particular event) within each
sensory inputs are generated in the hippocampus and communicated               block. One might ask, how one can disambiguate the effects of surprise
to the cortex is still unclear16. Irrespective of these mechanisms, they       and entropy if—in the absence of entropy—there can be no surprise?
should manifest as a differential modulation of prediction–error               The answer is straightforward—surprise is an attribute of a particular
responses in the visual cortex, depending on the predictability of the         event, whereas entropy is an attribute of events that could happen. In
current context, which we hypothesize is itself recognized and broad-          other words, the surprise or self-information of an outcome scores the
cast through hippocampal processing. Specifically, when upcoming               degree to which the event was ‘predicted’, whereas entropy scores the
stimuli are unpredictable, they are inherently informative in the sense        ‘predictability’ of an event before it is observed. Technically, entropy
that they resolve uncertainty when observed (technically, they have            is the time average or an integral of surprise. This means entropy is
a greater expected information gain). This leads to the hypothesis             the change in surprise and, therefore, over time, surprise and entropy
that the hippocampus has a role in precision weighting by modulat-             are decorrelated (Fig. 1b). In the context of the current task, they can
ing the electrophysiological correlates of prediction errors—that is,          therefore be dissociated with respect to stimulus onset.
event-related gamma activity in the visual hierarchy—as a function of
predictability (or entropy).                                                   Results
      Hippocampal sharp-wave ripples (SWRs; ripples henceforth) are            Seventeen participants successfully completed the task, with
found in every investigated mammalian brain and consist of sharp               trial-by-trial measures of entropy and surprise modulating reaction
waves (large-amplitude, negative-polarity activity resulting from              time (RT) in accordance with Hick’s law54. Replicating results from this
synchrony in the apical dendritic layer of CA1 pyramidal neurons)              task in healthy adults20, participants’ RTs increased significantly per
and ripples (~140 to 200 s−1 in rodents, resulting from interactions           bit of surprise (t(16) = 10.61, P < 0.001, Cohen’s d = 2.57) and entropy
between excitatory and inhibitory neurons in CA3 (refs. 37,38)). They          (t(16) = 4.86, P < 0.001, Cohen’s d = 1.18; Fig. 1c).
are viewed as a preconscious mechanism to explore the organism’s
options, searching for past experiences to extrapolate and predict             Prestimulus ripples increase with entropy
future outcomes27,39. This is in addition to their role in replay of past      In view of a possible generative role for hippocampal ripples in predict-
experiences38,40. As prospection relies on past experiences and is medi-       ing future events, we tested whether ripple occurrence was associ-
ated by the hippocampus25, ripples may also underlie anticipation of           ated with increased uncertainty. Using a previously established ripple
future outcomes to guide subsequent behavior. Indeed, ripples have             detection method52, a total of 2,728 hippocampal ripples were detected
been shown to portend behavior in the immediate future, in the form of         in all participants while they performed the visuomotor task (Fig. 2a
experienced trajectories41 and new paths42,43, as well as preplay of future    and Extended Data Fig. 1). On examining the two-dimensional (2D)
events during sleep44. Specifically, the sequential firing pattern that        distribution of ripple peak time as a function of peri-stimulus time and
occurs during rodent SWRs, in addition to ‘replay’ of spatial trajecto-        entropy (Fig. 2b), we found a large proportion of ripples occurred in
ries, has been shown to reflect all physically available trajectories within   high entropy (1.7–1.9 bits) trials between −800 and −400 ms prestimu-
the environment not realized in prior behavior43, and trajectories taken       lus. There was also a peak in ripple probability just before the stimulus
by participants in subsequent goal-directed navigation42. Furthermore,         onset (−200 to 0 ms) for entropy values from 1.6–1.7 to 1.8–1.9 bits.
there is evidence that cortical activity is modulated in a peri-ripple               We tested for the significance of this observation in several ways.
manner, both through enhancement and inhibition of activity45,46, and          First, a two-sample Kolmogorov–Smirnov test showed a significant
as fluctuations in resting-state networks47. This is in line with predictive   difference between the ripple distribution and a uniform distribution
processing frameworks in which predictions and prediction errors are           with the same minimum and maximum counts (k = 0.241, P = 0.001).
exchanged across levels in cortical hierarchies, with a special role for       Next, we performed a Spearman’s correlation between the observed
regions such as the hippocampus in contextualizing this exchange5,16.          and permuted distributions to ensure that ripple distribution was not


Nature Neuroscience
Article                                                                                                                                                                                                         https://doi.org/10.1038/s41593-026-02345-6


    a                         2
                                                                                                                                                                                                                   d
Entropy (bits)




                          1.5



                               1
                                                  50               100              150                200           250                    300             350           400           450

                              2
Mean entropy (bits)




                          1.9

                          1.8

                          1.7

                          1.6
                                                  50               100              150                200           250                    300             350           400           450

                           6
                                                                                                                                                                                                                                     x = –40
        Surprise (bits)




                           4

                              2                                                                                                                                                                                    e

                           0
                                                  50               100              150                200           250                    300             350           400           450

                                                                                                             Trial number


                                                                                                                                                  Entropy Surprise
                      Entropy Surprise                                      Entropy Surprise
                                                                                      2                                                                 2          2
                                2           2                                                    2
                                                                                                                                                  1
                      1                                                         1                                                                           1.92          2.32
                                    1.92           2.32                                   1.92           2.32
                                                                                                                                                                   1.92         2.58
                                           1.91          1.58                                    1.92          2.58
                                                                                                                                                                       1.84            1.22
                                                  1.84             1.81                                 1.84             1.22
                                                                                    Block 4                                                             Block 9             1.75              1
                                   Block 1                               1.41                                 1.75                    2                  trials
                                                       1.81                          trials
                                    trials                                                                                                                                       1.65             3.17
                                                              1.75              1.16                                1.75                  2.17                                                                                   y = –47; z = –24
                                                                                                                                                                                         1.76            2.32
                                                                    1.68                  1                                 1.84                 1.73                            ...
                                                             ...                                                   ...




   b                                                                                                                        c
                          5                                                                                                               300

                                                                                                                                          250
                          4
                                                                                                                                          200
  Surprise (bits)




                          3                                                                                                                150
                                                                                                                          ms bits–1




                                                                                                                                          100
                          2
                                                                                                                                           50

                           1                                                                                                                0

                                                                                                                                          –50                                                                                       x = 46
                          0
                           0.8              1          1.2           1.4            1.6          1.8           2                          –100
                                                                                                                                                            Entropy                               Surprise
                                                          Entropy (bits)

Fig. 1 | Task design and electrode contact localization. a, Distribution of                                                                                                  For a comparison of model fits showing a linear correlation (r = 0.3) is inferior
entropy, mean entropy and surprise values across trials for one example                                                                                                      to nonlinear models, see Supplementary Note—Relationship between entropy
participant. Bottom, example trials from 3 of 12 different blocks, and their                                                                                                 and surprise. c, Group average (n = 17 patients) RT as an increase in ms per bit
associated entropy (blue) and surprise (red) values. Please note that the four                                                                                               of information-theoretic measure. Error bars reflect 95% CI. d,e, Illustration of
colored shapes presented were unique to each block, with the probability of                                                                                                  contacts in the hippocampus (smoothed hippocampus mask from the AAL atlas
occurrence of these four stimuli varying between blocks. Participants were asked                                                                                             for visualization; blue), fusiform (red) and occipital cortex (green) overlaid
to perform a visuomotor selection task, choosing the corresponding colored                                                                                                   on a 100-μm T1 scan of an ex vivo human brain, acquired on a 7T MRI scanner
shape from four alternatives with a button press, as depicted for block 4 in the                                                                                             (https://openneuro.org/datasets/ds002179/versions/1.1.0); each color
current example. b, Correlation between entropy and surprise values across all                                                                                               represents a patient. See Supplementary Fig. 1 for individual patients’ contacts.
patients. Each patient is represented by a color. By design, entropy and surprise                                                                                            Unless otherwise stated, error bars represent the s.e.m. AAL, Automated
are decorrelated, with high surprise occurring in the context of low entropy.                                                                                                Anatomical Labeling.


correlated with random noise. The correlation between the two was                                                                                                           on the normalized count values to identify the bins showing the
computed in every permutation (1,000 permutations in total) per                                                                                                             largest number of ripples, using entropy and time bin as predictors.
participant, and the mean correlation across participants was com-                                                                                                          We found a significant interaction between entropy and ripple peak
pared to 0 using a one-sample t test. The observed 2D distribution was                                                                                                      time (χ2(99) = 207.6, P < 0.001), with the five largest estimated marginal
not correlated with the random permutations (t(16) = 1.24, P = 0.23).                                                                                                       means identified around −1,000 to –400 ms and 1.6 to 1.8 entropy
Third, we used a mixed-effects gamma generalized linear model (GLM)                                                                                                         bins, as well as just before stimulus onset (−0.2 to 0, entropy values of


Nature Neuroscience
Article                                                                                                                                                                                                                       https://doi.org/10.1038/s41593-026-02345-6


 a                                                                                                                                                               b
                                300
                                200
                                100
   Raw LFP




                                  0
     (uV)




                               –100                                                                                                                                                                                                                                                25
                               –200                                                                                                                                                                                                             6
                               –300                                                                                                                                                                                                                                                20
                                   –1                –0.8          –0.6             –0.4          –0.2          0            0.2        0.4   0.6   0.8   1
                                                                                                                                                                                    150                                                         5
                                    10
                                                                                                                                                                                                                                                                                   15




                                                                                                                                                              Frequency (Hz)
                                     5                                                                                                                                                                                                          4




                                                                                                                                                                                                                                                                    Voltage (uV)
     Ripple band
      80–120 Hz




                                                                                                                                                                                                                                                            Power
                                    0                                                                                                                                               100
                                                                                                                                                                                                                                                3                                  10
                                   –5
                                   –10                                                                                                                                                                                                          2                                   5
                                      –1             –0.8          –0.6             –0.4          –0.2          0            0.2        0.4   0.6   0.8   1
                                                                                                                                                                                    50
                                    4                                                                                                                                                                                                           1                                  0
Standardized

 amplitude




                                     2
   Hilbert




                                                                                                                                                                                     –400          –200        0       200      400                                                –5
                                    0                                                                                                                                                                                                                                               –400          –200                      0              200        400
                                                                                                                                                                                            Time relative to ripple peak (ms)                                                                  Time relative to ripple peak (ms)
                                    –2
                                      –1             –0.8          –0.6             –0.4          –0.2          0            0.2        0.4   0.6   0.8   1
                                                                                                           Time (s)

          c                                                                                                                                                       d
                                                                                                                                                                                                                                                                                                               600
                                                                                                                                                                                                                                                                                                               500




                                                                                                                                                                                                                                                                                                   Frequency
                                                                                                                                                                                                                                                                                                               400
                                                                                                                                                                                                                                                                                                               300
                                                                                                                                                                                                                Ripple peak time      (–0.2,0]                                          (0,0.2]
                                                                                                                                                                                                                                                                                                               200

                                                     1.9–2                                                                                                                                                                                                                                                     100

                                                    1.8–1.9                                                                                                                         0.045                                                                    0.045                                              0
                                                                                                                                                                                                                   *
     Normalized ripple count




                                                                                                                                                                                                                                                                                                                     20   40    60   80 100 120 140
                               3                                                                                                                                                                                                                                                                                          Ripple duration (ms)
                                                    1.7–1.8
                                                    1.6–1.7



                                                                                                                                                              Ripple duration (S)




                                                                                                                                                                                                                                      Ripple duration (S)
                                                    1.5–1.6                                                                                                                         0.040                                                                    0.040
                                         Entropy




                               2                    1.4–1.5
                                                    1.3–1.4
                                                    1.2–1.3                                                                                                                         0.035                                                                    0.035

                               1                    1.1–1.2
                                                      1–1.1
                                                     0.9–1                                                                                                                          0.030                                                                    0.030
                                                   0.8–0.9
                               0
                                                                                                  0
                                                              8




                                                                                                                              8
                                                                       6




                                                                                         2




                                                                                                                     6
                                                                                                         2
                                                                                4




                                                                                                               4




                                                                                                                                       1




                                                              –        –        –        –                                                                                          0.025                                                                    0.025
                                                                                      0.




                                                                                                       0.
                                                              0.




                                                                                                                             0.
                                                                   0.




                                                                                                                     0.
                                                                            0.




                                                                                                               0.




                                                                                                                                   to
                                                                                                 to

                                                                                                      to




                                                                                                                         to
                                                                                                                    to
                                                                                                           to




                                                                                                                                   8
                                                                                             2
                                                                                    to
                                                        o

                                                                  to

                                                                           to




                                                                                                                               0.




                                                                                             –
                                                                                             0.




                                                        –
                                                       1t




                                                                                                  0




                                                                                                                         6
                                                                                                                4
                                                                                                           2




                                                                                                                                                                                            1.00      1.25     1.50    1.75   2.00                                                         1              2                3           4
                                                                                4
                                                              8

                                                                       6




                                                                                                         0.




                                                                                                                     0.
                                                                                                               0.




                                                              –        –        –
                                                                            0.
                                                            0.

                                                                   0.




                                                                                                                                                                                                             Entropy                                                                              Surprise
                                                                                      Ripple peak time

Fig. 2 | The frequency and duration of hippocampal ripples increase with                                                                                                              occurred prestimulus and in high (but not highest) entropy levels (see Extended
uncertainty or expected information gain. a, An example ripple detected using                                                                                                         Data Fig. 3 for peri-stimulus ripple distribution). d, Ripple duration as a function
the method discussed in ref. 52, showing the raw LFP trace (top), the ripple band                                                                                                     of peri-stimulus time and information-theoretic measures. Ripple duration
signal (middle) and the standardized envelope (bottom). b, Grand average peri-                                                                                                        increased with entropy (that is, expected information gain for prestimulus
ripple wavelet spectrogram (left) and raw field potential centered on ripple peak                                                                                                     ripples), but decreased for poststimulus ripples; result from mixed-effects
(n = 2,728 ripple events from 17 participants). c, Normalized ripple distribution                                                                                                     gamma regression and subsequent omnibus χ2 Wald test for the interaction
across peri-stimulus time and entropy levels, accounting for the total number of                                                                                                      P = 0.0416). Ripple duration histogram, overall recorded ripples, is shown above
trials in each entropy bin. Normalized ripple count for each time and entropy bin                                                                                                     right. Lines represent the mean and shaded areas represent the 95% CI. *P < 0.05
is color coded; average counts for each time and entropy bin are plotted as bars                                                                                                      (two tailed). See Extended Data Fig. 1 for replication using a different ripple
above and to the right of the 2D distribution, respectively. The majority of ripples                                                                                                  detection method51. LFP, local field potential.



1.8–1.9) and just after stimulus onset (0–0.2, entropy values of 1.8–1.9).                                                                                                           effect, we applied a Poisson GLM predicting ripple counts as a func-
This is in line with the visual representation of the 2D distribution                                                                                                                tion of entropy, surprise and mean entropy for each time bin. The β
shown in Fig. 2c (all estimated marginal mean values are shown in                                                                                                                    values across participants were then tested against zero in a two-tailed
Extended Data Fig. 2). Next, we examined ripple frequency as a function                                                                                                              t test using a cluster-based Monte Carlo permutation to correct for
of time bin and peri-stimulus time using a two (pre/post) by five (time                                                                                                              multiple comparisons as implemented in FieldTrip. In line with the
bin) analysis of variance. We found a main effect of peri-stimulus time                                                                                                              analyses reported above, we found a significant effect in the −800 to
(F(1.16) = 6.02, P = 0.026), with more ripples observed prestimulus                                                                                                                  −600 ms time bin such that ripple count increased with entropy (clus-
(Extended Data Fig. 3c), whereas the main effect of bins and the interac-                                                                                                            ter t = 2.66, P = 0.009, 95% confidence interval (CI) = 0.002). No other
tion were not significant (all Ps > 0.073). Finally, a threshold-free cluster                                                                                                        significant effects were observed for entropy. For surprise, we found
enhancement (TFCE) analysis of the 2D histogram showed significant                                                                                                                   a negative association with ripple count in the same time bin (−800 to
effects again primarily in the entropy range of 1.8–1.9 bits, albeit for a                                                                                                           −600 ms; t(16) = −1.89, P = 0.039, uncorrected), indicating that as sur-
longer time period (Extended Data Fig. 3e).                                                                                                                                          prise increases, there is a decrease in ripple count for this time window
     In view of a proposal55 that awake replay during SWRs increases                                                                                                                 (note the effect of surprise did not survive a correction for multiple
following prediction error (that is, surprise), we then provide evi-                                                                                                                 comparisons). This analysis confirms that the increased prestimulus
dence that human hippocampal ripple frequency is increased by                                                                                                                        ripple count is related to entropy, as opposed to the effect of surprise,
unpredictability and not surprise. If ripples reflected processing                                                                                                                   and demonstrates statistical independence—we observed significant
of prediction errors postpresentation of surprising stimuli, rather                                                                                                                  effects of entropy on ripple counts, but no significant effects of surprise
than the predictability of upcoming stimuli, one would expect to                                                                                                                     across all time bins (Extended Data Fig. 2). Finally, we analyzed data
see an increased ripple rate with surprise. However, the 2D distribu-                                                                                                                from an independent control experiment56 using a classical ‘oddball’
tion of ripple peak time as a function of peri-stimulus time and sur-                                                                                                                paradigm (Supplementary Note (Independent data from oddball task)
prise (Extended Data Fig. 3d) shows cluster of ripples at low surprise                                                                                                               and Supplementary Table 2). Participants were presented with lists
(at 1–1.5 bits). Next, to further demonstrate that the increased ripple                                                                                                              of words where occasional perceptually distinct stimuli (perceptual
count observed in the prestimulus time bin is not driven by a surprise                                                                                                               oddballs) appeared in a different font among standard control words.


Nature Neuroscience
Article                                                                                                  https://doi.org/10.1038/s41593-026-02345-6

These oddball stimuli were highly surprising by design (that is, this is a      cortex, lower down in the visual cortical hierarchy than the fusiform
standard paradigm for eliciting prediction errors), yet we observed no          and putatively supplying it with bottom-up information (for example,
increased ripple count for oddball stimuli compared to control words            prediction errors). Hippocampal activity in the gamma range (57.5–
(Supplementary Fig. 2 and Supplementary Table 3). The dissocia-                 97.5 Hz; Fig. 3a) was negatively associated with entropy in the pres-
tion between oddball surprise (no ripple increase) and entropy-based            timulus period, around −990 to −670 ms prestimulus (summed cluster
uncertainty (clear ripple increase) supports our interpretation that            t = −762.8, P = 0.0218; smallest t value in cluster = −4.57, P = 0.0218,
ripples encode predictions about predictability rather than responses           Cohen’s d = 1.08), showing reduced gamma power preceding more
to unpredicted events.                                                          uncertain outcomes (see Extended Data Fig. 4 for split by the presence
     The duration of ripples has been shown to be functionally relevant         of prestimulus ripple). This time window partly overlaps with peaks in
for memory consolidation in rodents57. In predictive coding formula-            ripple occurrence, although, notably, the gamma power around this
tions, this rests on the augmentation of associative (that is, activity or      time window was reduced under high entropy. Further examination of
experience dependent) plasticity by precision weighting that increases          the negative association with entropy, as a function of trial in block,
presynaptic (prediction error) afferents4,58. To test for changes in rip-       showed that the reduction in gamma power was mostly concentrated
ple duration with uncertainty, we then examined the relationship                in the first few trials of the block (Extended Data Fig. 5), before the
between ripple duration and the information-theoretic measures.                 emergence of prestimulus ripples.
As ripple duration followed a right-skewed distribution (Fig. 2d), a                 In the occipital cortex, on the other hand, a positive associa-
mixed-effects gamma regression was used. First, we tested whether               tion between gamma activity (35–117.5 Hz; Fig. 3c) and entropy was
ripple duration was modulated by peri-stimulus time, entropy and                observed around −510 to −130 ms prestimulus (summed t = 1,694.1,
surprise. We did not observe any significant main effects (all Ps > 0.381)      P = 0.0078; largest t value in cluster = 10.67, P = 0.0078, Cohen’s
or interactions (time window by entropy—χ2(1) = 0.665, P = 0.414; time          d = 3.77). No significant responses to entropy were found in the fusi-
window by surprise—χ2(1) = 0.171, P = 0.679; Extended Data Fig. 3j).            form, or in the poststimulus time period for all three areas (Fig. 3).
However, comparing ripples occurring just before (−200 to 0 ms, where           Under the simplifying assumption that gamma activity reflects the
there was an increase in ripple frequency; Fig. 2b) and the correspond-         amplitude of precision-weighted prediction errors, these results are
ing period just after the stimulus (0 to 200 ms) revealed a significant         consistent with an increase in the precision of visual prediction errors,
main effect of peri-stimulus time window (χ2(1) = 5.04, P = 0.0248), as         with a concomitant decrease in the precision of hippocampal predic-
well as an interaction between time window and entropy (χ2(1) = 4.15,           tion errors, before stimuli with higher expected information gain. This
P = 0.0416; Fig. 2d). This interaction suggests that as entropy increased,      could be read as instantiating the right kind of attentional set, when
prestimulus ripple duration increased, but decreased for poststimulus           salient information can be anticipated in advance59.
ripples. Ripple duration was again not modulated by surprise (main
effect—χ2(1) = 0.279, P = 0.597; interaction—χ2(1) = 0.31, P = 0.577), sug-     Ripple-triggered cortical responses
gesting that ripple duration reports predictability (that is, expected          Next, we investigated whether ripple occurrence instates a tempo-
information gain) as opposed to violations of predictions (that is,             ral relationship between the hippocampus and visual processing
observed information gain).                                                     regions in which prestimulus modulations of gamma activity were
                                                                                also observed as a function of entropy. First, we compared peri-ripple
RT as a function of ripples                                                     cortical time–frequency (TF) responses, dichotomized by whether the
Given that there is a relationship between entropy and RT20,54, as well         hippocampal ripple occurred before or after stimulus onset (Fig. 4a,b).
as between ripples and entropy, we examined whether the presence                We found an overall increase in high gamma power in the fusiform
of ripples modulates the relationship between RT and entropy—and                (−200 to 200 ms peri-ripple, 105–160 Hz, summed cluster t = 2,171.1,
surprise—on a trial-by-trial basis. To ensure that sufficient prestimulus       P = 0.0156; largest t value in cluster = 4.44, P = 0.0156, Cohen’s d = 1.67)
ripple trials were included, this analysis was restricted to trials with        as well as in occipital cortex (−180 to 200 ms peri-ripple, 95–160 Hz,
1.6–1.9 bits of entropy, where most prestimulus ripples were observed           summed cluster t = 1,474, P = 0.0039; largest t value in cluster = 9.15,
(50% prestimulus ripple trials and 50% no ripple trials are within these        P = 0.0039, Cohen’s d = 3.27) locked to hippocampal ripples. Further-
values). We found a significant main effect of ripple status, with faster       more, in occipital cortex, prestimulus and poststimulus ripples modu-
responses in trials with a prestimulus ripple (χ2(1) = 4.33, P = 0.037). This   lated activity differently (−10 to 200 ms peri-ripple, 35–160 Hz; Fig. 4c;
is an important observation because the precision of prediction errors          summed t = −1,094.5, P = 0.0039; smallest t value in cluster = −8.99,
corresponds to the rate of evidence accumulation that underwrites               P = 0.0039, Cohen’s d = 3.17). Specifically, occipital gamma activity
reaction speed. In other words, prediction errors that are afforded             was reduced after prestimulus ripples compared to poststimulus rip-
more precision exert their effects on belief updating more rapidly.             ples. This effect was partly driven by a power suppression, compared
There was also a main effect of surprise, with faster responses for             to a baseline period, of occipital gamma around prestimulus ripples
lower levels of surprise (χ2(1) = 91.3, P < 0.001), as per Hick’s law. The      (Fig. 4d shows raw power suppression in relation to baseline). That is,
interaction between ripple status and surprise, however, was not sig-           when examining occipital gamma activity in the prestimulus period
nificant (χ2(1) = 2.58, P = 0.107), indicating that ripples did not modu-       in which a positive association with entropy was found (indicating an
late the relationship between speed and surprise. There was a trend             overall power increase), there was reduced gamma power in trials with
toward an interaction between entropy and ripple status (χ2(1) = 3.41,          a hippocampal ripple in the period before the significant cluster com-
P = 0.064), with simple slopes of entropy in trials without ripples (esti-      pared to trials with no ripples (t(7) = 2.51, P = 0.040, Cohen’s d = 0.889;
mate = 0.2, t(16) = 1.55, P = 0.12) and in trials with prestimulus ripples      Extended Data Fig. 6a). Taken together, these findings are indicative of
(estimate = 0.59, t(16) = 3.28, P = 0.001), indicating a steeper positive       a hippocampal prestimulus ripple-induced suppression of prestimulus
slope in trials with prestimulus ripples.                                       gamma power in the occipital cortex (see Extended Data Fig. 7 for repli-
                                                                                cation using a different ripple detection method). No differences were
Region-specific responses to uncertainty                                        observed between prestimulus and poststimulus ripple modulation
To further characterize hippocampal and cortical correlates of entropy          of fusiform activity, or in lower frequencies in either cortical region.
and surprise, we then examined time-resolved spectral responses as                   Next, we examined whether ripple-modulated cortical activ-
a function of these measures. The focus was on two cortical regions—            ity was correlated with entropy or surprise. There were no signifi-
the fusiform gyrus, previously shown with functional magnetic reso-             cant associations between peri-ripple cortical activity and either
nance imaging (fMRI) to respond to surprise in this task20, and occipital       information-theoretic measure, suggesting that the hippocampal


Nature Neuroscience
Article                                                                                                                                                             https://doi.org/10.1038/s41593-026-02345-6


                a                                                     Hippocampus
                              160                                                                                     4

                                                                                                                      3
                              140                                                                                                                                  0.15
                                                                                                                          2                                         0.1
                              120
             Frequency (Hz)
                                                                                                                                                                  0.05




                                                                                                                                            Parameter estimates
                                                                                                                      1




                                                                                                                                  T value
                                                                                                                                                                     0
                              100                                                                                     0                                           –0.05

                              80                                                                                      –1                                           –0.1
                                                                                                                                                                  –0.15
                                                                                                                      –2
                              60                                                                                                                                   –0.2
                                                                                                                      –3                                          –0.25
                              40
                                                                                                                      –4                                           –0.3
                                    –1   –0.8   –0.6   –0.4    –0.2         0       0.2   0.4   0.6    0.8                                                        –0.35
                                                                        Time (s)                                                                                          Entropy        Surprise

               b                                                 Occipital cortex
                              160                                                                                     4                                             0.2

                                                                                                                      3
                              140                                                                                                                                  0.15
                                                                                                                      2




                                                                                                                                            Parameter estimates
             Frequency (Hz)




                              120                                                                                                                                   0.1
                                                                                                                      1




                                                                                                                                T value
                              100                                                                                     0                                           0.05

                              80                                                                                      –1
                                                                                                                                                                     0
                                                                                                                      –2
                              60
                                                                                                                      –3                                          –0.05
                              40
                                                                                                                      –4
                                                                                                                                                                   –0.1
                                    –1   –0.8   –0.6    –0.4     –0.2       0       0.2   0.4   0.6   0.8
                                                                                                                                                                           Entropy         Surprise
                                                                         Time (s)

                c                                                       Fusiform
                              160                                                                                 4

                                                                                                                  3
                              140
                                                                                                                  2
             Frequency (Hz)




                              120
                                                                                                                  1
                                                                                                                              T value




                              100                                                                                 0

                              80                                                                                  –1

                                                                                                                  –2
                              60
                                                                                                                  –3
                              40
                                                                                                                  –4
                                    –1   –0.8   –0.6   –0.4    -0.2        0        0.2   0.4   0.6   0.8
                                                                        Time (s)

Fig. 3 | Hippocampal and occipital cortical gamma activity is modulated by                             Right, parameter estimates for each predictor within the significant cluster;
entropy ‘before’ stimuli are presented. a, Left, time-resolved spectral power                          each point pertains to one patient, horizontal bars represent the group mean
in the hippocampus (n = 17 patients) shows a significant negative association                          and vertical error bars represent 95% CI. b, In occipital cortex (n = 8), there was
between gamma power and entropy around 800 ms prestimulus (significant                                 an increase in gamma power as entropy increased, around 400 ms prestimulus.
permutation-based cluster from one-sample two-tailed t test against a value of 0                       Again, entropy was not associated with changes in the poststimulus time
outlined in black here and in Figs. 4 and 5). During highest entropy levels, gamma                     windows. Each point pertains to one patient, horizontal bars represent the group
power was lowest (see Extended Data Fig. 5 for gamma power as a function of trial                      mean and vertical error bars represent 95% CI. c, In the fusiform cortex (n = 7), no
in block). No significant clusters were found in the poststimulus time window.                         effects of entropy were observed either prestimulus or poststimulus.



ripple modulation of cortical response was not affected by uncertainty                                 observations23. In the fusiform, there was a positive association between
or surprise, and thus ripples may act as a general mechanism for hip-                                  surprise and gamma activity (40–95 Hz; Fig. 5b) around 200–510
pocampal modulation of cortical activity, with context dependence of                                   poststimulus (summed cluster t = 741.9, P = 0.0234; largest t value in
this modulation affected by ripple rate and/or duration.                                               cluster = 6.79, P = 0.0234, Cohen’s d = 2.56), and a negative association
                                                                                                       between surprise and alpha/beta power (10–15 Hz; Fig. 5c) around
Hippocampal and fusiform gamma to surprise                                                             510–840 ms poststimulus (summed cluster t = −199.5, P < 0.001; larg-
We expected that prediction error, in the form of surprise, would elicit                               est negative t value in cluster = −5.41, P < 0.001, Cohen’s d = 2.05).
hippocampal and cortical responses. In line with previous findings,                                    No significant associations between surprise and occipital activity
a positive association between surprise and hippocampal activity in                                    were observed.
the gamma range (52.5–90 Hz; Fig. 5a and see Extended Data Fig. 8 for
replication using a different ripple detection method) was observed                                    Ripple-modulated responses to surprise
around 160–540 ms poststimulus (summed cluster t = 888.2, P = 0.016;                                   Under the predictive processing framework, a cardinal function of
largest t value in cluster = 4.91, P = 0.016, Cohen’s d = 1.2) and a nega-                             neural activity is to minimize precision-weighted prediction error5.
tive association between surprise and hippocampal activity in theta                                    A possible role for hippocampal ripples is to broadcast the predicted
band (5–12.5 Hz; Extended Data Fig. 9a) later in the trial (530–810 ms                                 precision of forthcoming prediction errors. Specifically, we hypoth-
poststimulus; summed cluster t = −169.4, P = 0.011; smallest t value                                   esized that prestimulus ripples would amplify precision-weighted
in cluster = −3.19, P = 0.011, Cohen’s d = 0.773) in line with previous                                prediction errors, evoked by stimuli in an unpredictable (high entropy)


Nature Neuroscience
Article                                                                                                                                                                                                                                             https://doi.org/10.1038/s41593-026-02345-6


                                         a                                     Fusiform                                                                                                b                                            Occipital cortex
                                                                               all ripples                                                                                                                                            all ripples
                                                        160                                                                                  4                                                        160                                                                        4




                                                                                                                                                                                                                                                                                      Peri-ripple OCC power (t value)
                                                                                                                                                   Peri-ripple OCC power (t value)
                                                                                                                                             3                                                                                                                                   3
                                                        140                                                                                                                                           140
                                                                                                                                             2                                                                                                                                   2




                                                                                                                                                                                     Frequency (Hz)
                                                        120                                                                                                                                           120

                                      Frequency (Hz)
                                                                                                                                             1                                                                                                                                   1

                                                        100                                                                                  0                                                        100                                                                        0

                                                         80                                                                                  –1                                                       80                                                                         –1

                                                                                                                                             –2                                                                                                                                  –2
                                                         60                                                                                                                                           60
                                                                                                                                             –3                                                                                                                                  –3
                                                         40                                                                                                                                           40
                                                                                                                                             –4                                                                                                                                  –4
                                                           –0.2         –0.1          0         0.1                 0.2                                                                                 –0.2            –0.1                    0          0.1       0.2
                                                                Time relative to HPC ripple (s)                                                                                                                Time relative to HPC ripple (s)



                  c                                    Occipital cortex prestimulus                                                                                                                                                 d                  Occipital cortex
                                                       versus poststimulus ripples               Post < pre                                                                                                                                          prestimulus ripples
                                160                                                                   4                                           Pre                                                       Post                    160                                                                                 0.3




                                                                                                                                                                                                                                                                                                                               OCC power relative to baseline
                                                                                                           Peri-ripple OCC power (t value)
                                140                                                                                                                                                                                                 140
                                                                                                      2                                                                                                                                                                                                                 0.2
                                120                                                                                                                                                                                                 120
               Frequency (Hz)




                                                                                                                                                                                                                   Frequency (Hz)
                                100                                                                   0                                                                                                                             100                                                                                 0.1

                                80                                                                                                                                                                                                      80

                                                                                                      –2                                                                                                                                                                                                                0
                                60                                                                                                                                                                                                      60


                                40                                                                                                                                                                                                      40
                                                                                                      –4                                                                                                                                                                                                                –0.1
                                  –0.2                   –0.1       0           0.1       0.2     Pre < post                                                                                                                             –0.2       –0.1         0         0.1                          0.2

                                              Time relative to HPC ripple (s)                                                                                                                                                                   Time relative to HPC ripple (s)

Fig. 4 | Peri-hippocampal ripple cortical activity. a,b, Cortical TF analyses time                                                                                                                d, Further examination of this effect revealed that it was driven by a suppression
locked to the hippocampal ripple peak time. Across all ripple trials, there was an                                                                                                                of gamma power relative to baseline in trials with prestimulus hippocampal
overall increase in high gamma power in the fusiform (a) and occipital cortex (b)                                                                                                                 ripples and compared to trials with no ripples (Extended Data Fig. 6). See
time locked to hippocampal ripple. c, Comparing trials with prestimulus versus                                                                                                                    Extended Data Fig. 7 for replication using a different ripple detection method.
poststimulus hippocampal ripples, there was a reduction in occipital gamma                                                                                                                        OCC, occipital cortex; HPC, hippocampus.
power in trials with prestimulus ripples, compared to poststimulus ripples.



context. To test this hypothesis, we examined whether the occur-                                                                                                                                       This observation raised the question of what the impact of pres-
rence of prestimulus hippocampal ripples modulated poststimulus                                                                                                                                   timulus hippocampal ripples could be on cortical dynamics in the
(0–1 s) cortical responses to surprise (for example, the increase in                                                                                                                              fusiform that would later modulate prediction error gamma response.
gamma power in the fusiform). This hypothesis was confirmed in                                                                                                                                    We therefore examined aperiodic exponents in the fusiform cortex
the fusiform cortex. We found an early (250–550 ms poststimulus,                                                                                                                                  (30–70 Hz, from −750ms to +250 ms around stimulus onset) as a meas-
summed t = 936.9, P = 0.0078) increase in fusiform gamma power                                                                                                                                    ure of cortical excitation/inhibition (E/I) balance60,61. This analysis
(47.5–122.5 Hz) in response to higher surprise in trials that were                                                                                                                                revealed that the aperiodic exponent—and hence the underlying
preceded by a prestimulus hippocampal ripple. By contrast, a later                                                                                                                                cortical excitability state—differs significantly based on the occur-
(510–710 ms) increase in gamma power (50–72.5 Hz, summed t = 473.8,                                                                                                                               rence of a prestimulus hippocampal ripple (t(6) = 2.63, P = 0.039).
P = 0.0078) was observed to higher surprise when there was no pres-                                                                                                                               Specifically, trials without prestimulus ripples show smaller exponents
timulus ripple (Fig. 5d and Extended Data Fig. 10a,b). The presence of                                                                                                                            (mean = 2.485, s.d. = 0.405; that is, flatter spectral slopes), indicat-
a faster fusiform response to surprise in trials with prestimulus rip-                                                                                                                            ing a shift toward excitation (E>>I), whereas trials with prestimulus
ples could simply reflect a co-occurrence of the two. To rule out this                                                                                                                            ripples show steeper slopes (mean = 2.549, s.d. = 0.371), indicating a
possibility, we examined whether there was a double dissociation of                                                                                                                               more balanced E/I state (Extended Data Fig. 10h). Critically, the most
the surprise parameter estimate using a repeated measures analysis                                                                                                                                neurobiologically grounded accounts of precision encoding appeal
of variance with ripple status (no ripple, prestimulus ripple) and time                                                                                                                           to fast-spiking inhibitory interneurons see ref. 16 for a particular focus
window (early, later) as factors. This showed a significant interaction                                                                                                                           on the hippocampus). In this framework, the more balanced E/I state
(F(1,6) = 19.6, P = 0.004, η2P = 0.766), confirming the double dissocia-                                                                                                                          following ripples reflects increased inhibitory control, which directly
tion (Extended Data Fig. 10c) such that an early effect of surprise was                                                                                                                           implements higher precision weighting of prediction errors. Con-
observed for trials with prestimulus ripples, but not for trials without                                                                                                                          versely, the E>>I state in trials without ripples reflects reduced inhibi-
ripples; conversely, a later surprise effect was found in trials without                                                                                                                          tory control and correspondingly lower precision weighting.
ripples, but not in trials with prestimulus ripples. This result suggests
that the presence of a prestimulus hippocampal ripple modulates the                                                                                                                               Directed information flow
fusiform response to surprise, such that it is faster—and of greater                                                                                                                              Finally, we examined information flow between the hippocampus and
amplitude—compared to trials without a prestimulus ripple. The                                                                                                                                    fusiform and occipital cortex by testing for Granger causality (GC) as
presence of prestimulus hippocampal ripples did not modulate post­                                                                                                                                a function of surprise and entropy. To do so, a median split for each
stimulus occipital responses to surprise.                                                                                                                                                         information-theoretic measure was applied and a nonparametric


Nature Neuroscience
Article                                                                                                                                                                                                                                      https://doi.org/10.1038/s41593-026-02345-6


   a                                                    Hippocampus                                                           b                                                           Fusiform                                         c                                                     Fusiform
                 160                                                                                                4                                                                                                              4                                                                                         4
                                                                                                                                                     160

                                                                                                                                                                                                                                   3                           30                                                            3
                 140                                                                                                                                 140
                                                                                                                    2                                                                                                              2                                                                                         2
                                                                                                                                                                                                                                                                 25
Frequency (Hz)




                 120




                                                                                                                            Frequency (Hz)




                                                                                                                                                                                                                                        Frequency (Hz)
                                                                                                                                                     120
                                                                                                                                                                                                                                   1                                                                                         1
                                                                                                                                                                                                                                                                20




                                                                                                                                                                                                                                                                                                                                  t value
                 100                                                                                                0                               100                                                                            0                                                                                         0
                                                                                                                                                                                                                                                                   15
                      80                                                                                                                                 80                                                                        –1                                                                                        –1

                                                                                                                    –2                                                                                                             –2                             10                                                         –2
                     60                                                                                                                                  60
                                                                                                                                                                                                                                   –3                                      5                                                 –3
                     40                                                                                                                                  40
                                                                                                                    –4                                                                                                             –4                                                                                        –4
                                       0         0.2     0.4    0.6      0.8                                                                                    0             0.2     0.4       0.6          0.8                                                               0       0.2       0.4       0.6        0.8
                                                           Time (s)                                                                                                                       Time (s)                                                                                                Time (s)



                                           0.2
                                                                                                                                                                0.1                                                                                                             0.4




                                                                                                                                                                                                                                                     Parameter estimates
                                                                                                                                         Parameter estimates
                 Parameter estimates




                                           0.1                                                                                                                      0                                                                                                           0.2


                                            0                                                                                                                  –0.1                                                                                                                0

                                                                                                                                                               –0.2                                                                                                            –0.2
                                       –0.1
                                                                                                                                                               –0.3                                                                                                            –0.4
                                       –0.2
                                                                                                                                                               –0.4                                                                                                            –0.6
                                                     Entropy          Surprise                                                                                                 Entropy                Surprise                                                                              Entropy              Surprise

                                                 d      Hippocampus                                                                                                           Fusiform                                       e               Fusiform to hippocampus
                                                                                                                  0.1
                                                                                                                                                                                                                                                                                                             High surprise
                                                                                                                                                                                                                             –5
                                                                                                                                                                                                                                                                                                             Low surprise
                                                                                 Surprise parameter estimates




                                                                                                                                                                                                                             –6
                                                                                                                0.05
                                                                                                                                                                                                                             –7
                                                                                                                                                                                                                   log(GC)




                                                                                                                   0                                                                                                         –8


                                                                                                                                                                                                                             –9
                                                                                                                –0.05
                                                                                                                                                                                                                             –10
                                                                                                                                     Prestimulus ripple
                                                                                                                                     No ripple
                                                                                                                 –0.1                                                                                                        –11
                                                                                                                        0                          0.2                  0.4         0.6       0.8        1                                                20                           40             60         80
                                                        Prestimulus                                                                                                      Time (s)                                                                                              Frequency (Hz)

Fig. 5 | Surprise-driven responses. a, Hippocampal increase in gamma power as                                                                                                                  low-frequency responses to surprise. d, The increase in fusiform gamma power
surprise increases. Bottom, parameter estimates for each predictor within the                                                                                                                  of 40–95 Hz (extracted from significant cluster) occurs earlier in trials with
significant cluster (n = 17). Each point pertains to one patient, horizontal bars                                                                                                              prestimulus hippocampal ripples (red) compared to trials without ripples (blue).
represent the group mean and vertical error bars represent 95% CI. b, Fusiform                                                                                                                 Lines represent the mean, and shaded areas represent the s.e.m. See Extended
gamma power increases as surprise increases (n = 7). Each point pertains to                                                                                                                    Data Fig. 8 for replication using a different ripple detection method. e, Directed
one patient, horizontal bars represent the group mean and vertical error bars                                                                                                                  information flow from fusiform to hippocampus for high (orange) versus low
represent 95% CI. c, Fusiform beta oscillations decrease as surprise increases                                                                                                                 (green) surprise trials was observed in the gamma band (shaded rectangle). Lines
(n = 7). Each point pertains to one patient, horizontal bars represent the group                                                                                                               represent the mean, and shaded areas represent the s.e.m.
mean and vertical error bars represent 95% CI. See Extended Data Fig. 9 for




spectral GC analysis was performed. Guided by the TF effects above,                                                                                                                                 Subsequent GC analyses showed no significant effects in the
we focused on hippocampus–occipital cortex for entropy, and hip-                                                                                                                               reverse direction (hippocampus → fusiform; Extended Data Fig. 10f).
pocampus–fusiform cortex for surprise, each centered in time around                                                                                                                            Thus, top-down effects are expressed as hippocampal ripple-dependent
the observed TF effect (−1000 to −500 ms prestimulus for entropy,                                                                                                                              modulation of fusiform cortical dynamics ‘before’ surprise, whereas
and 150 to 650 ms poststimulus for surprise). For surprise, we found                                                                                                                           bottom-up effects manifest as fusiform to hippocampal-directed con-
increased bottom-up information flow (fusiform → hippocampus) in                                                                                                                               nectivity in the gamma range ‘after’ surprise. No GC effects between
the gamma range (52–60 Hz; Fig. 5e and Extended Data Fig. 10a–c) for                                                                                                                           the hippocampus and occipital cortex as a function of entropy were
high versus low surprise (summed t = 13.2, P = 0.0078). To ensure that                                                                                                                         significant in either direction (Extended Data Fig. 10d,e). The absence
this effect was valid, we ran two further analyses—first, a GC analysis on                                                                                                                     of directed information flow alongside clear functional modulation
the time-reversed data, showing an effect in the same frequency ranges                                                                                                                         (prestimulus ripple-associated reduction in occipital gamma activity)
in the opposite direction, as expected, and, second, a partial directed                                                                                                                        suggests that hippocampal influence on uncertainty processing either
coherence analysis that revealed a significant information flow effect                                                                                                                         operates through state preparation rather than continuous informa-
from the fusiform to the hippocampus in high versus low surprise trials                                                                                                                        tion transfer or is mediated along the processing hierarchy to enable
for the same frequency range of 52–59 Hz (cluster t(6) = 4, P < 0.001;                                                                                                                         such state preparation. As there were few participants with contacts in
Extended Data Fig. 10g).                                                                                                                                                                       both cortical regions (n = 4), GC between occipital and fusiform cortex


Nature Neuroscience
Article                                                                                                                         https://doi.org/10.1038/s41593-026-02345-6




                 Hippocampalripple




                 Gamma activity



                    GC
                                                      High entropy (~1.6–1.9 bits)                                                 High surprise (>2 bits)

                 Temporal association


                                        –1,000    –800                 –600                 –400   –200       0         200       400        600         800      1,000



                    Hippocampus




                   Occipital cortex
                                                            log10(power)




                   Fusiform gyrus
                                                                           Frequency [Hz]




Fig. 6 | Human hippocampal–cortical dialog in precision weighting of                                 a more inhibitory state. We interpret these prestimulus effects to reflect the
prediction errors. Peri-stimulus time is indicated in ms, and involvement                            setting of cortical dynamics to imbue precision to subsequently surprising
of relevant brain areas (hippocampus in blue, occipital cortex in green and                          stimuli in the high entropy context. In line with these predictions, different
fusiform cortex in red) at different time points is indicated by the inclusion of                    levels of entropy are not associated with any poststimulus changes in spectral
that structure in the top row. Before stimulus presentation (left), unpredictable                    power. Surprising stimuli, however, lead to poststimulus increases in gamma
visual stimulus streams modulate hippocampal gamma activity and ripple                               power in fusiform and hippocampus (right), with directed functional
occurrence, with a change in dynamics in visual cortical areas temporally                            connectivity following fusiform to hippocampus in this gamma range. The
associated with hippocampal ripples. These effects are observed before                               critical observation is that a prestimulus hippocampal ripple is associated
stimulus presentation. From ~1,000 ms before stimulus presentation, there is                         with a change in excitatory–inhibitory balance in the fusiform cortex (left),
a reduction in hippocampal gamma activity, coinciding with a peak in ripple                          and a subsequent poststimulus gamma response in the fusiform to surprising
rate at ~800 ms prestimulus. By contrast, occipital cortex shows an increase                         stimuli (right) that is greater in amplitude and of shorter latency (~200 ms). No
in gamma power ~500 ms prestimulus, but this increase is suppressed if the                           poststimulus responses to surprising stimuli are observed in occipital cortex,
hippocampus generates a ripple. Spectral power in the fusiform cortex is not                         in keeping with an established role of higher-order visual cortices (that is,
modulated by entropy prestimulus. However, the occurrence of a prestimulus                           fusiform) in generating prediction error neuronal responses. For illustrative
hippocampal ripple (which is most frequent under high entropy) shifts the                            purposes, gamma activity is shown as a continuous oscillation.
excitatory–inhibitory balance of the prestimulus fusiform field potential to



was not examined. All hippocampal–cortical interactions reported                                          The properties and distribution of ripples as a function of
above are summarized schematically for high entropy (prestimulus                                     peri-stimulus time and entropy strongly suggest that they have an
time periods) and high surprise (poststimulus time periods) in Fig. 6.                               important role in generating a certain kind of prediction; namely, a
                                                                                                     prediction of predictability. Previous work in rodents has demon-
Discussion                                                                                           strated that ripples represent future trajectories of experienced and
We measured hippocampal and cortical activity using depth electrodes                                 new paths41–43, indicative of a generative process. Our results provide
in patients with epilepsy as they observed sequences of simple visual                                support for, and extend, these previous studies by showing increased
stimuli drawn from a block-specific distribution, yielding varying lev-                              ripple probability in uncertain contexts before stimulus onset, in awake
els of stimulus-bound uncertainty and surprise. Electrophysiological                                 human patients, as they anticipate upcoming sensory input. Specifi-
intracranial activity was examined prestimulus and poststimulus to                                   cally, predictive processing under uncertainty may represent an inte-
assess prediction under different levels of uncertainty and subsequent                               gral part of the proposed role for ripples in planning future actions40,41.
validation against observed inputs, respectively. Our results highlight                              Our findings are in keeping with a recent synthesis of rodent studies
an important hippocampal function in generating putative predictions                                 suggesting that ripple trajectories comprise not only the path to be cho-
and communicating them downstream to the cortex. Specifically, our                                   sen or the path just completed, but also many of the potential options
findings suggest that hippocampal ripples are key to this generative                                 available62. With an increasing number of options—and implicit unpre-
process—ripple probability was highest in the prestimulus window                                     dictability of ensuing outcomes—the ripple rate would be expected to
when uncertainty was high but not at its maximum, and prestimulus                                    increase, as observed here. This is also in line with recent fMRI work in
ripple duration increased with entropy. These prestimulus ripples                                    humans showing successor-like representations in the hippocampus
were associated with subsequent faster responses to poststimulus                                     and visual cortex29. Notably, our visuomotor mapping task did not
surprise in the fusiform cortex. This is exactly consistent with a role                              impose memory demands, as both stimulus and response mapping
of hippocampal ripples reporting the predictability of an upcoming                                   appeared on the screen simultaneously. The key manipulation was
stimulus and thereby affording the ensuing prediction errors greater                                 the underlying probability of the stimulus distribution within block,
precision, in the context of greater expected information gain (that                                 which we suggest elicited an increased need for precise predictive
is, higher entropy).                                                                                 processing, and thus more ripple events, with increased uncertainty


Nature Neuroscience
Article                                                                                              https://doi.org/10.1038/s41593-026-02345-6

regarding the upcoming stimulus. This allowed us to elucidate an             processing35,36, so that the faster response to surprise in the presence
important functional role that ripples have in anticipation of sensory       of prestimulus ripples indicates that a prior prediction facilitates
inputs and associated action planning48.                                     signaling of prediction errors31, and their propagation of the visual
      The increased frequency of prestimulus ripples occurred around         processing hierarchy. Further support for this claim arises from the
the same time as the negative association between entropy and gamma          E/I analyses, which provided direct neurobiological evidence for our
power (overlapping in 80–97.5 Hz), namely, around 800 ms prestimu-           precision-weighting hypothesis—prestimulus hippocampal ripples
lus. The early hippocampal gamma effect (that is, negative association       enhance the precision afforded to subsequent prediction errors
with entropy) is—on a predictive coding narrative—consistent with an         through enhanced inhibitory control mechanisms, resulting in faster,
attenuation of neuronal fluctuations in populations reporting predic-        stronger and more controlled cortical responses to surprising stimuli.
tion errors (for example, superficial pyramidal cells) in proportion to           Furthermore, we also observed increased fusiform to hippo­
the predicted precision of subsequent (stimulus-bound) prediction            campus information flow in high surprise as reflected in GC for
errors11. Such an early effect is consistent with (Bayesian belief) an       gamma activity from 150 to 650 poststimulus, which we associate
updating of information from the previous trial and broadcasting the         with bottom-up prediction error signaling. This is compatible with a
ensuing predictions down the cortical hierarchy. This finding also           view that prestimulus hippocampal ripples modulate (or prime) the
suggests that detected ripples were likely discrete events, dissociable      fusiform to respond at earlier latencies to surprising stimuli (top-down
from gamma activity37,63. The observed reduction in gamma activity           effect), and the fusiform poststimulus gamma response to surprise
(spanning frequencies lower than that of the ripple activity) occurred       feeds back to the hippocampus (bottom-up effect). The precise nature
in the prestimulus period on trials in which prestimulus ripples were        of hippocampal ripple-triggered changes in cortical state remains to be
evident (Extended Data Fig. 4). This co-occurrence might be driven by        determined (for example, cortical layer-specific engagement of inhibi-
the opposing effects of acetylcholine on ripple generation and gamma         tory interneurons), although the simplicity of the current behavioral
power (increase in ripple rate accompanied by decreased gamma-band           task lends itself to further mechanistic interrogation using nonhuman
activity, and vice versa37). This is interesting because acetylcholine has   animal models, where homologous predictive coding mechanisms
been implicated in the encoding of precision in predictive processing        are evident66.
in several computational studies9,58,64.                                          Finally, we also observed poststimulus hippocampal ripples, but
      To demonstrate that ripples carry predictive information that is       did not, however, find a relationship between these and behavioral
subsequently propagated to the cortex, it is important to establish          measures or cortical activity. Different cortical effects linked to pres-
peri-ripple modulation of behavior or cortical activity, as suggested        timulus versus poststimulus hippocampal ripples could imply that
by predictive processing5,16. There are mixed findings in the literature,    these ripples arise in different neuronal populations whose outputs are
with some studies showing no ripple occurrence-related modulation            routed differently. Therefore, it is possible that poststimulus ripples
of subsequent behavior in rodents48–50. Nevertheless, hippocampal            serve a different functional role38,67, or facilitate communication with
ripples modulated cortical activity in several ways; in line with previ-     other brain regions, such as the prefrontal cortex45,46.
ous findings from rodents and nonhuman primates45–47, we found an                 In conclusion, our findings speak to an important function of hip-
overall increase in high gamma power in fusiform and occipital cortex        pocampal ripples in predictive processing, reporting the predictability
around the ripple time, supporting the notion that the occurrence of         or expected information gain of stimuli (that is, prediction errors)
ripples facilitates enhanced interaction between the hippocampus and         before they are encountered, thereby facilitating their propagation
cortex. These findings demonstrate that ripples implement general            through the visual cortex (in the absence of any memory demands).
precision-weighting mechanisms that modulate cortical readiness              More specifically, our results reveal an increase in ripple events in
across different computational contexts, rather than being tied to           uncertain trials, before stimulus onset, and subsequent modulation
any single variable.                                                         of cortical activity, pointing to enhanced hippocampal–cortical com-
      In addition to the cortical modulation across all ripple trials, a     munication facilitated by ripples that serve the propagation of precise
unique modulation of prestimulus ripples on cortical processing was          prediction errors. These findings reveal complementary mechanisms
observed, in keeping with their predictive role. First, we found a sup-      through which hippocampal ripples achieve their modulatory effects—
pression of occipital gamma activity that was time locked to the hip-        preparing cortical states for upcoming uncertain stimuli (entropy
pocampal ripple event. Together with the overall positive association        pathway) and enhancing the precision of prediction–error responses
between entropy and occipital gamma power prestimulus, this sug-             when surprising events occur (surprise pathway).
gests that the ripple-triggered suppression may reflect an overriding
hippocampal signal over lower-level cortical prediction of the upcom-        Online content
ing stimulus, or facilitate a sharper representation of the possible         Any methods, additional references, Nature Portfolio reporting sum-
outcomes in occipital cortex14,65. We suggest that this suppression rep-     maries, source data, extended data, supplementary information,
resents state optimization that prepares the cortex for more effective       acknowledgements, peer review information; details of author contri-
prediction error processing when stimuli arrive. Therefore, although         butions and competing interests; and statements of data and code avail-
there is a general increase (positive association) between occipital         ability are available at https://doi.org/10.1038/s41593-026-02345-6.
gamma as entropy increases, this is modulated (suppressed) in the
presence of prestimulus hippocampal ripples (Extended Data Fig. 6a).         References
This anticipatory prediction generation is exactly what predictive cod-      1.   Gregory, R. L. Perceptions as hypotheses. Philos. Trans. R. Soc.
ing theory predicts; when upcoming stimuli are unpredictable (high                Lond. B Biol. Sci. 290, 181–197 (1980).
entropy), the visual system must prepare more robust predictions to          2.   Friston, K. J. & Kiebel, S. J. Predictive coding under the free-energy
effectively process the expected information.                                     principle. Philos. Trans. R. Soc. B Biol. Sci. 364, 1211–1221 (2009).
      Second, the presence of prestimulus ripples modulated poststimu-       3.   Clark, A. Whatever next? Predictive brains, situated agents, and the
lus fusiform activity; while there was an overall positive association            future of cognitive science. Behav. Brain Sci. 36, 181–204 (2013).
between trial-wise surprise and gamma power in fusiform around               4.   Rao, R. P. N. & Ballard, D. H. Predictive coding in the visual cortex:
300 ms poststimulus, when splitting trials based on the presence of               a functional interpretation of some extra-classical receptive-field
prestimulus hippocampal ripples, we found a faster fusiform gamma                 effects. Nat. Neurosci. 2, 79–87 (1999).
response to surprise, compared to trials without prestimulus rip-            5.   Friston, K. J. A theory of cortical responses. Philos. Trans. R. Soc. B
ples. Gamma-band activity has been shown to support bottom-up                     Biol. Sci. 360, 815–836 (2005).


Nature Neuroscience
Article                                                                                             https://doi.org/10.1038/s41593-026-02345-6

6.  Mumford, D. On the computational architecture of the neocortex.         28. Dayan, P. Improving generalization for temporal difference learning:
    I. The role of the thalamo-cortical loop. Biol. Cybern. 65, 135–145         the successor representation. Neural Comput. 5, 613–624 (1993).
    (1991).                                                                 29. Ekman, M., Kusch, S. & de Lange, F. P. Successor-like
7. Yon, D. & Frith, C. D. Precision and the Bayesian brain. Curr. Biol.         representation guides the prediction of future events in human
    31, R1026–R1032 (2021).                                                     visual cortex and hippocampus. eLife 12, e78904 (2023).
8. Clark, A. The many faces of precision (replies to commentaries on        30. Moser, E. I., Kropff, E. & Moser, M. B. Place cells, grid cells, and
    ‘whatever next? Neural prediction, situated agents, and the future          the brain’s spatial representation system. Annu. Rev. Neurosci. 31,
    of cognitive science’). Front. Psychol. 4, 270 (2013).                      69–89 (2008).
9. Moran, R. J. et al. Free energy, precision and learning: the role        31. Bell, A. H., Summerfield, C., Morin, E. L., Malecek, N. J. &
    of cholinergic neuromodulation. J. Neurosci. 33, 8227–8236                  Ungerleider, L. G. Encoding of stimulus probability in macaque
    (2013).                                                                     inferior temporal cortex. Curr. Biol. 26, 2280–2290 (2016).
10. FitzGerald, T. H. B., Moran, R. J., Friston, K. J. & Dolan, R. J.       32. Arnal, L. H. & Giraud, A. L. Cortical oscillations and sensory
    Precision and neuronal dynamics in the human posterior parietal             predictions. Trends Cogn. Sci. 16, 390–398 (2012).
    cortex during evidence accumulation. Neuroimage 107, 219–228            33. Bastos, A. M. et al. Canonical microcircuits for predictive coding.
    (2015).                                                                     Neuron 76, 695–711 (2012).
11. Hesselmann, G., Sadaghiani, S., Friston, K. J. & Kleinschmidt, A.       34. Chao, Z. C., Takaura, K., Wang, L., Fujii, N. & Dehaene, S.
    Predictive coding or evidence accumulation? False inference and             Large-scale cortical networks for hierarchical prediction and
    neuronal fluctuations. PLoS ONE 5, e9926 (2010).                            prediction error in the primate brain. Neuron 100, 1252–1266
12. Haarsma, J. et al. Precision weighting of cortical unsigned                 (2018).
    prediction error signals benefits learning, is mediated by              35. Bastos, A. M. et al. Visual areas exert feedforward and feedback
    dopamine, and is impaired in psychosis. Mol. Psychiatry 26,                 influences through distinct frequency channels. Neuron 85,
    5320–5333 (2021).                                                           390–401 (2015).
13. Limanowski, J. Precision control for a flexible body                    36. Bastos, A. M., Lundqvist, M., Waite, A. S., Kopell, N. & Miller, E. K.
    representation. Neurosci. Biobehav. Rev. 134, 104401 (2022).                Layer and rhythm specificity for predictive routing. Proc. Natl
14. Hindy, N. C., Ng, F. Y. & Turk-Browne, N. B. Linking pattern                Acad. Sci. USA 117, 31459–31469 (2020).
    completion in the hippocampus to predictive coding in visual            37. Liu, A. A. et al. A consensus statement on detection of
    cortex. Nat. Neurosci. 19, 665–667 (2016).                                  hippocampal sharp wave ripples and differentiation from other
15. Dimakopoulos, V., Mégevand, P., Stieglitz, L. H., Imbach, L. &              fast oscillations. Nat. Commun. 13, 6000 (2022).
    Sarnthein, J. Information flows from hippocampus to auditory            38. Joo, H. R. & Frank, L. M. The hippocampal sharp wave–ripple in
    cortex during replay of verbal working memory items. eLife 11,              memory retrieval for immediate use and consolidation. Nat. Rev.
    e78677 (2022).                                                              Neurosci. 19, 744–757 (2018).
16. Barron, H. C., Auksztulewicz, R. & Friston, K. J. Prediction and        39. Swanson, R. A., Levenstein, D., McClain, K., Tingley, D. &
    memory: a predictive coding account. Prog. Neurobiol. 192,                  Buzsáki, G. Variable specificity of memory trace reactivation
    101821 (2020).                                                              during hippocampal sharp wave ripples. Curr. Opin. Behav. Sci.
17. Bornstein, A. M. & Daw, N. D. Cortical and hippocampal correlates           32, 126–135 (2020).
    of deliberation during model-based decisions for rewards in             40. Buzsáki, G. Hippocampal sharp wave–ripple: a cognitive
    humans. PLoS Comput. Biol. 9, e1003387 (2013).                              biomarker for episodic memory and planning. Hippocampus 25,
18. Fuhrer, J., Glette, K. & Ivanovic, J. Direct brain recordings reveal        1073–1188 (2015).
    implicit encoding of structure in random auditory streams. Sci.         41. Diba, K. & Buzsáki, G. Forward and reverse hippocampal place-
    Rep. 15, 14725 (2025).                                                      cell sequences during ripples. Nat. Neurosci. 10, 1241–1242 (2007).
19. Schapiro, A. C., Turk-Browne, N. B., Norman, K. A. & Botvinick, M. M.   42. Pfeiffer, B. E. & Foster, D. J. Hippocampal place-cell sequences
    Statistical learning of temporal community structure in the                 depict future paths to remembered goals. Nature 497, 74–79 (2013).
    hippocampus. Hippocampus 26, 3–8 (2016).                                43. Gupta, A. S., van der Meer, M. A. A., Touretzky, D. S. & Redish, A. D.
20. Strange, B. A., Duggins, A., Penny, W., Dolan, R. J. & Friston, K. J.       Hippocampal replay is not a simple function of experience.
    Information theory, novelty and hippocampal responses:                      Neuron 65, 695–705 (2010).
    unpredicted or unpredictable? Neural Netw. 18, 225–230                  44. Dragoi, G. & Tonegawa, S. Preplay of future place cell sequences
    (2005).                                                                     by hippocampal cellular assemblies. Nature 469, 397–401 (2011).
21. Barron, H. C., Dolan, R. J. & Behrens, T. E. J. Online evaluation       45. Logothetis, N. K. et al. Hippocampal-cortical interaction during
    of novel choices by simultaneous representation of multiple                 periods of subcortical silence. Nature 491, 547–553 (2012).
    memories. Nat. Neurosci. 16, 1492–1498 (2013).                          46. Nitzan, N., Swanson, R., Schmitz, D. & Buzsáki, G. Brain-wide
22. Kaplan, R. et al. The neural representation of prospective choice           interactions during hippocampal sharp wave ripples. Proc. Natl
    during spatial planning and decisions. PLoS Biol. 15, e1002588              Acad. Sci. USA 119, e2200931119 (2022).
    (2017).                                                                 47. Kaplan, R. et al. Hippocampal sharp-wave ripples influence
23. Axmacher, N. et al. Intracranial EEG correlates of expectancy               selective activation of the default mode network. Curr. Biol. 26,
    and memory formation in the human hippocampus and nucleus                   686–691 (2016).
    accumbens. Neuron 65, 541–549 (2010).                                   48. Silva, D., Feng, T. & Foster, D. J. Trajectory events across
24. Lisman, J. E. & Redish, A. D. Prediction, sequences and the                 hippocampal place cells require previous experience. Nat.
    hippocampus. Philos. Trans. R. Soc. Lond. B Biol. Sci. 364,                 Neurosci. 18, 1772–1779 (2015).
    1193–1201 (2009).                                                       49. Gillespie, A. K. et al. Hippocampal replay reflects specific past
25. Hassabis, D. & Maguire, E. A. Deconstructing episodic memory                experiences rather than a plan for subsequent choice. Neuron
    with construction. Trends Cogn. Sci. 11, 299–306 (2007).                    109, 3149–3163 (2021).
26. Frank, D. & Kafkas, A. Expectation-driven novelty effects in            50. Findlay, G., Tononi, G. & Cirelli, C. The evolving view of replay and
    episodic memory. Neurobiol. Learn. Mem. 183, 107466 (2021).                 its functions in wake and sleep. Sleep Adv. 1, zpab002 (2020).
27. Kay, K. & Frank, L. M. Three brain states in the hippocampus and        51. Norman, Y. et al. Hippocampal sharp-wave ripples linked to visual
    cortex. Hippocampus 29, 184–238 (2019).                                     episodic recollection in humans. Science 365, eaax1030 (2019).


Nature Neuroscience
Article                                                                                                 https://doi.org/10.1038/s41593-026-02345-6

52. Vaz, A. P., Inati, S. K., Brunel, N. & Zaghloul, K. A. Coupled ripple       64. Iglesias, S. et al. Hierarchical prediction errors in midbrain and
    oscillations between the medial temporal lobe and neocortex                     basal forebrain during sensory learning. Neuron 80, 519–530
    retrieve human memory. Science 363, 975–978 (2019).                             (2013).
53. Axmacher, N., Elger, C. E. & Fell, J. Ripples in the medial temporal        65. De Lange, F. P., Heilbron, M. & Kok, P. How do expectations shape
    lobe are relevant for human memory consolidation. Brain 131,                    perception? Trends Cogn. Sci. 22, 764–779 (2018).
    1806–1817 (2008).                                                           66. Pezzulo, G., Parr, T. & Friston, K. The evolution of brain
54. Hick, W. E. On the rate of gain of information. Q. J. Exp. Psychol. 4,          architectures for predictive coding and active inference. Philos.
    11–26 (1952).                                                                   Trans. R. Soc. B Biol. Sci. 377, 20200531 (2022).
55. Van der Meer, M. A. A. & Bendor, D. Awake replay: off the clock but         67. Ramirez-Villegas, J. F., Logothetis, N. K. & Besserve, M. Diversity
    on the job. Trends Neurosci. 48, 257–267 (2025).                                of sharp-wave-ripple LFP signatures reveals differentiated
56. Hellerstedt, R. et al. Emotional amnesia in humans with focal                   brain-wide dynamical events. Proc. Nat. Acad. Sci. USA 112,
    temporal pole lesions. Preprint at bioRxiv https://doi.org/10.1101/             E6379–E6387 (2015).
    2025.10.22.683695 (2025).
57. Fernández-Ruiz, A. et al. Long-duration hippocampal sharp wave              Publisher’s note Springer Nature remains neutral with regard to
    ripples improve memory. Science 364, 1082–1086 (2019).                      jurisdictional claims in published maps and institutional affiliations.
58. Friston, K. Hierarchical models in the brain. PLoS Comput. Biol. 4,
    e1000211 (2008).                                                            Open Access This article is licensed under a Creative Commons
59. Feldman, H. & Friston, K. J. Attention, uncertainty, and                    Attribution 4.0 International License, which permits use, sharing,
    free-energy. Front. Hum. Neurosci. 4, 215–238 (2010).                       adaptation, distribution and reproduction in any medium or format,
60. Gao, R., Peterson, E. J. & Voytek, B. Inferring synaptic excitation/        as long as you give appropriate credit to the original author(s) and the
    inhibition balance from field potentials. Neuroimage 158, 70–78             source, provide a link to the Creative Commons licence, and indicate
    (2017).                                                                     if changes were made. The images or other third party material in this
61. Donoghue, T. et al. Parameterizing neural power spectra                     article are included in the article’s Creative Commons licence, unless
    into periodic and aperiodic components. Nat. Neurosci. 23,                  indicated otherwise in a credit line to the material. If material is not
    1655–1665 (2020).                                                           included in the article’s Creative Commons licence and your intended
62. Diba, K. Hippocampal sharp-wave ripples in cognitive map                    use is not permitted by statutory regulation or exceeds the permitted
    maintenance versus episodic simulation. Neuron 109, 3071–3074               use, you will need to obtain permission directly from the copyright
    (2021).                                                                     holder. To view a copy of this licence, visit http://creativecommons.
63. Tong, A. P. S., Vaz, A. P., Wittig, J. H., Inati, S. K. & Zaghloul, K. A.   org/licenses/by/4.0/.
    Ripples reflect a spectrum of synchronous spiking activity in
    human anterior temporal lobe. eLife 10, e68401 (2021).                      © The Author(s) 2026




Nature Neuroscience
Article                                                                                                https://doi.org/10.1038/s41593-026-02345-6

Methods                                                                        were then applied to the brain mask from SPM8 to transform the brain
Participants                                                                   mask into the native space of the pre-MRI. All voxels in the pre-MRI
Eighteen patients with medication-resistant presurgical epilepsy (mean         that were outside the brain mask and with a signal value in the top
age = 37.3 years, s.d. = 11.7; eight males) with depth electrodes surgically   15% were filtered out. The skull-stripped pre-MRI was then coregis-
implanted to aid seizure focus localization took part in the experiment.       tered and resliced to the skull-stripped post-CT. Next, the pre-MRI
Data were acquired at two epilepsy centers, one in Madrid and one in           was affine registered to the post-CT, thus transforming the pre-MRI
Zurich. All patients signed informed consent before participating in the       image into a native post-CT space. The two images were then overlaid,
experiment. Implantation sites were chosen solely on the basis of clini-       with the post-CT thresholded such that only electrode contacts were
cal criteria. Patients had normal or corrected-to-normal vision and had        visible. For all patients, only contacts in the hippocampus head and
no history of head trauma or encephalitis. All patients had electrodes         body were selected. For patients with multiple electrodes localized
implanted in the hippocampus. For patients showing unilateral hip-             in the hippocampus, only the most anterior contacts were used in the
pocampal sclerosis, only the nonpathological side was included in the          analyses (Fig. 1c). Electrode contacts for each patient are shown in
analysis; in all other patients, hippocampi were radiologically normal         Supplementary Fig. 1.
on preoperative MRI.
     Of the 18 patients who completed the task, 1 patient was excluded         Electrode localization: cortical regions
from any analysis due to poor performance on the task (less than 25%           To localize electrodes with contacts in the occipital and fusiform cor-
accuracy on all trials). We therefore analyzed electrophysiological            tices, we used a semiautomatic procedure with Lead-DBS69 (https://
responses from 17 patients. In total, we analyzed contacts in the              www.lead-dbs.org/). First, the postoperative CT was registered to the
right hippocampus from seven patients, and eight patients with                 preoperative MRI using a two-stage linear registration (rigid followed
contacts on the left side. Of these 17 patients, 8 patients also had           by affine) as implemented in Advanced Normalization Tools70. The
electrodes in the occipital cortex, and 7 patients also had contacts           images were then normalized to the MNI template based on the pre-MRI
in the fusiform. Our sample size was not determined before the                 using the SyN registration approach as implemented in Advanced
experiment, but it is equal or larger than that reported in previ-             Normalization Tools. To reduce bias introduced by brain shift, the
ous publications using cognitive tasks with intracranial electroen-            brain shift correction implemented in Lead-DBS was performed on
cephalography (iEEG)51,68. This study has been approved by the local           the postoperative CT. Manual prereconstruction was then performed,
ethics committees of Hospital Ruber Internacional and Kantonale                in which the tip of each electrode and another point along its trajec-
Ethikkommission (PB-2016-02055).                                               tory were manually marked, followed by automatic reconstruction
                                                                               guided by the electrode specification (that is, the number of contacts
Stereotactic electrode implantation                                            and their spacing). The reconstructed electrodes were then visually
A contrast-enhanced MRI was performed pre-operatively under stereo-            inspected and refined, and the processes were iterated in case of any
tactic conditions to map vascular structures before electrode implanta-        misalignments. Once all electrodes (for any given patient) have been
tion, and to calculate stereotactic coordinates for trajectories using the     reconstructed, they were visualized using Lead-Group71. After this pro-
Neuroplan system (Integra Radionics). For patients whose data were             cess, each electrode contact was associated with MNI coordinates. To
collected in Madrid (n = 12), DIXI Medical Microdeep depth electrodes          identify contacts in our cortical regions of interest, a custom MATLAB
(multicontact, semirigid, diameter of 0.8 mm, contact length of 2 mm,          code, together with the FindStructure function (https://alivelearn.
intercontact isolator length of 1.5 mm) were implanted based on the            net/?p=1456), was used. Each MNI coordinate was associated with a
stereotactic Leksell method. For patients whose data were collected            label from the Automated Anatomical Labeling atlas. For occipital cor-
in Zurich (n = 5), the depth electrodes (diameter of 1.3 mm, eight con-        tex, MNI coordinates labeled as ‘inferior occipital’ or ‘middle occipital’
tacts of 1.6 mm length and spacing between contact centers of 5 mm;            were used; for fusiform, the ‘fusiform’ label was selected. The selected
Ad-Tech, www.adtechmedical.com) were stereotactically implanted                contacts were then visually inspected with the Automated Anatomi-
in the medial temporal lobes.                                                  cal Labeling overlay in MRIcron, as well as in native space (Fig. 1d and
                                                                               Supplementary Fig. 1).
Data acquisition
In Madrid, iEEG activity was acquired using an XLTEK EMU128FS ampli-           Behavioral task
fier (XLTEK). iEEG data were recorded at each electrode contact site at        Patients performed a visuomotor mapping task, consisting of 12 blocks
a 500 Hz sampling rate (online bandpass filter of 0.1–150 Hz) and ref-         with 40 trials per block, using the same design as discussed in ref. 20.
erenced to linked mastoid electrodes. For four patients, the data were         Each trial included a brief presentation of a colored shape, for 500 ms,
recorded with a higher sampling rate, but were later downsampled to            with an interstimulus interval of 2200 ms. In all trials within a block, two
500 Hz. In Zurich, data were acquired using a Neuralynx ATLAS system           colors and two shapes were combined to form four possible outcomes,
with a sampling rate of 4000 Hz (online bandpass filter of 0.5–1000 Hz)        with different stimuli presented in the different blocks. Patients were
against a common intracranial reference, and then downsampled to               asked to respond to the sampled item by pressing a key to identify the
500 Hz. Data from the two centers were comparable, and we did not              target’s position in the row (Fig. 1a). Each trial used an independent
observe differences across centers in any of the analyses reported in          sample from a distribution that remained constant within a block,
the Results section.                                                           but that varied over blocks. There was no underlying sequence gov-
                                                                               erning stimulus presentation; only the relative proportions of stimuli
Electrode localization: hippocampus                                            were varied from block to block. Two information theoretic measures,
To localize electrodes with contacts in the hippocampus, we used the           entropy and surprise, were then calculated (Fig. 1b). Surprise quantifies
manual procedure described previously68. For each patient, the post-           the improbability of a given event:
electrode placement CT (post-CT) was coregistered to the preoperative
T1-weighted MRI (pre-MRI). To optimize coregistration, both brain                                            I (xi ) = −ln P (xi )
images were first skull stripped. For CTs, this was done by filtering out
all voxels with signal intensities between 100 and 1,300 HU. Skull strip-          Entropy quantifies the expected (running average of) surprise
ping of the pre-MRI proceeded by first spatially normalizing the image         overall the trials:
to MNI space using the New Segment algorithm in SPM8 (http://www.
fil.ion.ucl.ac.uk/spm). The resultant inverse normalization parameters                                   H(X) = ∑ −P (xi ) ln P (xi )


Nature Neuroscience
Article                                                                                             https://doi.org/10.1038/s41593-026-02345-6

     Patients were told the proportion of colored shapes presented          a Kolmogorov–Smirnov test. To identify the bins with the highest
within block was random, and independent of the other blocks in the         ripple counts, we also performed a mixed-effects gamma regression
task. To account for nonspecific time effects within block, the mean        on the normalized ripple counts (as they follow a right-skewed dis-
entropy overall blocks was calculated for each patient. Trial number        tribution) with entropy and ripple peak time bin as predictors. This
is therefore accounted for by the inclusion of mean entropy values          was done using the lme4 package77 in R (https://www.r-project.org/),
that are the same across blocks. Therefore, for each trial, three val-      and the following model setup: normalised_count ~ entropy × peak_
ues of interest were computed—entropy, surprise and mean entropy.           time + (1|subI), family = Gamma. After the significant interaction, the
These values were modeled from the perspective of an ideal Bayesian         top five bins with the highest estimated marginal means were identi-
observer, using the Dirichlet distribution20. Only correct responses        fied using the pairwise function from the emmeans package (https://
given within 1 s of stimulus onset were used for subsequent analyses.       rvlenth.github.io/emmeans/). Finally, TFCE was calculated on the
                                                                            normalized 2D histogram, with a t test against the average ripple rate
Electrophysiological data analysis                                          per patient. Therefore, a positive value indicates a higher-than-average
Preprocessing. iEEG data analysis was carried out using the FieldTrip       ripple count observed.
toolbox72 (https://www.fieldtriptoolbox.org; v20200310) running                  Finally, to test whether the increased ripple frequency was related
on MATLAB R2019b (MathWorks). For all patients and regions of               to entropy, over the effect of surprise, we used a standard summary
interest, recordings were transformed to a bipolar derivation by sub-       statistic approach to mixed-effects modeling, using a separate Poisson
tracting signal from adjacent electrode contacts within the region          GLM for each participant, for each time bin, that included a constant
of interest (hippocampus, occipital cortex and fusiform). Previous          term, modeling the ripple rate for that participant and time bin. The
studies demonstrate that bipolar referencing optimizes estimates of         predictors in this model were entropy and surprise, as well as mean
local activity73,74 and connectivity patterns across brain regions75, as    entropy as covariate of no interest (as was done in all other TF analyses
well as for analysis of SWRs52. Data were epoched from −1 to 1 s with       described below). The ensuing parameter estimates were then subject
respect to stimulus onset (a time window selected a priori), demeaned       to a second-level analysis using cluster-based permutation tests as
and detrended. For each region of interest, every epoch was visually        implemented in FieldTrip (using the TFCE method did not change
inspected for artifacts caused by epileptic spikes or electrical noise,     the results).
first in the time domain and then in the TF domain. Trials with artifacts        Next, we examined ripple duration, calculated as ripple end point
were excluded from all subsequent analyses. Of note, for analyses           minus ripple start point (both determined in the detection algorithm
involving more than one region, only trials that were artifact free in      as two z scores above average) as a function of peri-stimulus time
all regions were used.                                                      and entropy and surprise. This was done using generalized linear
                                                                            mixed-effects models implemented by the lme4 package77 in R (https://
SWR analyses                                                                www.r-project.org/). Because ripple duration follows a right-skewed
Ripples were detected using a previously established method52               distribution, we used a model from the gamma family with a random
(Supplementary Fig. 2a). After trial-wise artifact rejection and basic      intercept for patient using the following syntax:
preprocessing (described above), the iEEG signal was bandpass fil-               model ← glmer(rip_duration ~ entropy × peri_stim_time + sur-
tered between 80 and 120 Hz using a second-order Butterworth filter.        prise × peri_stim_time + mean_entropy + (1 | patient), family = Gamma).
A Hilbert transform was then applied to the filtered signal to extract           A similar approach using mixed-effects linear regression was used
its instantaneous amplitude. Ripples were identified as events with a       to examine RT, which was normalized using log transformation. This
maximum amplitude three s.d. above the mean, and with a duration            approach accommodates any individual differences among partici-
of at least 25 ms. Ripples that were detected within 15 ms of each other    pants (for example, the overall ripple rate for each participant).
were merged. The start, peak and end time of each ripple were noted.
To ensure that artifacts or interictal epileptiform discharges (IEDs;       TF analyses
Extended Data Fig. 3) were not mistakenly classified as ripples, we used    Time-resolved spectral decomposition was computed for each trial
an automated procedure as described in ref. 52 to identify and reject       using seven Slepian multitapers for high frequencies ( ≥35 Hz) and a
IEDs. The iEEG signal was high-pass filtered at 200 Hz and a z score was    single Hann taper for low frequencies (<3 5Hz). The selected Slepian
calculated based on the gradient and amplitude of the filtered signal.      tapers for the analysis of high frequencies were based on windows with
Any time-point exceeding a z score of 5 was marked as an IED, together      width of 0.4 s and a 10 Hz frequency smoothing. The time-resolved
with the 100 ms before and after the event. All IEDs were excluded          spectral estimation was done in steps of 2.5 Hz. No baseline correction
from the analysis, increasing the likelihood of the detected ripples        was performed, given our interest in both prestimulus and poststimu-
being physiological52. We have also applied another ripple detection        lus activities. Trial-wise TF estimates were then entered into a GLM;
method51 and replicated our findings (Extended Data Figs. 1, 7 and 8).      two predictors of interest (entropy and surprise) and a covariate (mean
For patients who had more than one anterior hippocampal electrode           entropy) were used to predict power at each TF point. The parameter
(for example, head and body), we examined whether the same ripple           estimates for entropy, mean entropy and surprise are proportional to
was picked up by the different electrodes (within 20 ms). In line with      the range these values can assume. This resulted in a ‘first-level’ beta
previous findings in rodents76, we found the vast majority of ripples       map (with size time × frequency) per predictor. The spectral activity
(96% on average) were not detected in the adjacent electrode along          was then averaged over bipolar channels, within each region, for each
the hippocampal long axis.                                                  patient (in some cases, there was only one bipolar channel in each brain
      We then examined whether ripple occurrence was influenced by          region). These beta maps were then used for statistical inference with
peri-stimulus time and information-theoretic measures. To do so, a 2D       a cluster-based permutation test with the maximum number of per-
histogram of ripple peak count using peri-stimulus time and entropy or      mutations allowed by our sample size, up to a maximum of 5,000
surprise was derived for each patient. We normalized each histogram by      permutations. In each permutation step, clusters were formed by
the total number of trials per entropy/surprise bin to account for poten-   temporal and frequency adjacency using a cluster-forming threshold
tial imbalances. This was done by calculating the 2D count histogram        of P = 0.05. In each permutation step, a one-sample two-tailed t test
(time and entropy or surprise) and by multiplying by the frequency of       against a value of 0 (equivalent to H0— β = 0), using a threshold of
observed ripples given available trials in each entropy or surprise bin     P = 0.025, was calculated for each TF estimate, separately for low (2.5–
(that is, postcleaning). The normalized histograms were then averaged       32.5 Hz) and high (35–160 Hz) frequencies, as well as prestimulus (−1
across participants and tested against a uniform distribution using         to 0 s) and poststimulus (0 to 1 s) time windows. Gamma frequency


Nature Neuroscience
Article                                                                                              https://doi.org/10.1038/s41593-026-02345-6

band is used to refer to frequencies of 30–100 Hz, whereas high gamma        Reporting summary
refers to effects above 100 Hz. The precise frequency ranges (and            Further information on research design is available in the Nature
associated time windows) that we report for each significant effect          Portfolio Reporting Summary linked to this article.
are derived from a cluster-based permutation correction for
multiple comparisons.                                                        Data availability
                                                                             Preprocessed data needed to generate the figures, and over which
Ripple-modulated cortical activity                                           statistics were computed, are available in the following GitHub reposi-
To examine the temporal relationship between hippocampal ripples             tory: https://github.com/frdarya/GenerativeRipples.
and cortical responses, we extracted TF estimates in occipital cor-
tex and fusiform with respect to peri-ripple peak time (−0.2 to 0.2 s,       Code availability
selected a priori), as described above for peri-stimulus responses,          Analysis codes are available in the following GitHub repository: https://
for high and low frequencies. Due to the shorter time window, the            github.com/frdarya/GenerativeRipples.
selected Slepian tapers for the analysis of high frequencies were based
on windows with width of 0.2 s and a 10 Hz frequency smoothing. The          References
time-resolved spectral estimation was done in steps of 5 Hz. These TF        68. Méndez-Bértolo, C. et al. A fast pathway for fear in human
estimates were baseline corrected by calculating the relative change             amygdala. Nat. Neurosci. 19, 1041–1049 (2016).
with respect to a baseline period of −1.4 to −1 s peri-stimulus. Averaged    69. Horn, A. & Kühn, A. Lead-DBS: A toolbox for deep brain stimulation
TF estimates across patients were used for statistical inference using           electrode localizations and visualizations. NeuroImage 107, 127–135
cluster-based permutations, as described above. In each permutation              (2014).
step, a one-sample two-tailed paired t test was performed between            70. Avants, B., Tustison, N. J. & Song, G. Advanced normalization tools
prestimulus and poststimulus periods (threshold of P = 0.025). Next,             (ANTS). Insight J. 2, 1–35 (2009).
to examine whether prestimulus hippocampal ripples modulated                 71. Treu, S. et al. Deep Brain Stimulation: Imaging on a group level.
poststimulus cortical responses to surprise, we used the same GLM                NeuroImage 219, 3086 (2020).
approach as described above, fitting a trial-wise regression at each         72. Oostenveld, R., Fries, P., Maris, E. & Schoffelen, J. M. FieldTrip:
TF point with entropy, surprise and mean entropy as predictors                   open source software for advanced analysis of MEG, EEG, and
(TF analyses).                                                                   invasive electrophysiological data. Comput. Intell. Neurosci.
                                                                                 2011, 156869 (2011).
E/I balance analysis                                                         73. Hamamé, C. M. et al. Functional selectivity in the human
E/I balance analysis of activity in fusiform cortex was performed using          occipitotemporal cortex during natural vision: evidence from
the ‘fitting oscillations and one over f’ (FOOOF)61 method as imple-             combined intracranial EEG and eye-tracking. Neuroimage 95,
mented in FieldTrip. To achieve this, we extracted the aperiodic expo-           276–286 (2014).
nent value for trials with or without prestimulus hippocampal ripples.       74. Shirhatti, V., Borthakur, A. & Ray, S. Effect of reference scheme on
The time series was first epoched from −750 to +250 ms around stimulus           power and phase of the local field potential. Neural Comput. 28,
onset, as this reflects the majority of hippocampal ripples detected             882–913 (2016).
(Fig. 2c), and around the onset of the fusiform gamma response to            75. Trongnetrpunya, A. et al. Assessing Granger causality in electro­
surprise (Fig. 5b). The data were then bandpass filtered between 30 and          physiological data: removing the adverse effects of common
70 Hz, filtering out the 50 Hz line noise. Fractal spectral estimation was       signals via bipolar derivations. Front. Syst. Neurosci. 9, 189 (2016).
performed with a spectral smoothing of 2 Hz. This resulted in a single       76. Patel, J., Schomburg, E. W., Berényi, A., Fujisawa, S. & Buzsáki, G.
exponent value for trials with or without prestimulus ripples, for each          Local generation and propagation of ripples along the septotemporal
patient. At the group level, a paired two-tailed t test was performed to         axis of the hippocampus. J. Neurosci. 33, 17029–17041 (2013).
compare the exponent values, with smaller exponents reflecting flatter       77. Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear
spectral slopes, indicating a shift toward excitation (E>>I).                    mixed-effects models using lme4. J. Stat. Softw. 67, 1–48 (2015).

GC analyses                                                                  Acknowledgements
Finally, to evaluate the direction of information flow between the hip-      The authors thank the patients who participated in the study, the
pocampus and our cortical regions of interest, we calculated spectral        electroencephalography technicians at Hospital Ruber Internacional
nonparametric GC as a measure of directed functional connectivity            and the Swiss Epilepsy Center in Zurich, and members of the
using the FieldTrip toolbox, on 500 ms time windows centered on the          Laboratory for Clinical Neuroscience for helpful discussions.
induced responses observed to surprise in the fusiform gyrus and
hippocampus, and the induced responses to entropy in the occipital           Author contributions
cortex and hippocampus. For patients with multiple bipolar chan-             B.A.S. and K.J.F. designed the experiment. B.A.S., J.S. and R.T. collected
nels in any of the regions, the most medial and anterior channel was         the data. R.T., L.I., L.S. and A.G.-N. monitored patients and performed
used to compute GC. Briefly, time-resolved spectral decomposition            clinical evaluation. N.L. and A.H. developed software to perform
was computed using nine Slepian multitapers (frequency range of              electrode localization. D.F., S.M., R.H. and B.A.S. performed analyses.
2–80 Hz and 10 Hz smoothing). The spectral transfer matrix was then          D.F., K.J.F. and B.A.S. wrote the paper with input from all other authors.
obtained from the Fourier transformation of the data and together
with the noise covariance matrix, they were used to calculate the            Funding
total and intrinsic power through which GC is computed. The result-          This project was supported by the European Research Council (ERC)
ing GC values were then log-transformed to normalize the data for            under the European Union’s Horizon 2020 research and innovation
statistical inference. We then compared information in high versus           program (ERC-2018-COG 819814 to B.A.S.). D.F. is supported by
low entropy and surprise (using a median split) in each direction (hip-      a Royal Society University Research Fellowship (URF/R1/241499).
pocampus → cortex and cortex → hippocampus) with nonparametric               K.J.F. is supported by funding for the Wellcome Centre for Human
cluster-based permutations (using a dependent samples of two-tailed          Neuroimaging (205103/Z/16/Z), a Canada-UK Artificial Intelligence
t test), using the maximum number of permutations available for              Initiative (ES/T01279X/1) and the European Union’s Horizon 2020
each pair of regions.                                                        Framework Programme for Research and Innovation under the


Nature Neuroscience
Article                                                                                       https://doi.org/10.1038/s41593-026-02345-6

Specific Grant Agreement 945539 (Human Brain Project SGA3). This       Supplementary information The online version
research was supported in part by the Wellcome Trust (205103/Z/16/Z)   contains supplementary material available at
and the Swiss National Science Foundation (supported by SNSF           https://doi.org/10.1038/s41593-026-02345-6.
204651). For the purpose of Open Access, the authors have applied
a CC BY public copyright license to any Author Accepted Manuscript     Correspondence and requests for materials should be addressed to
version arising from this submission.                                  Darya Frank or Bryan A. Strange.

Competing interests                                                    Peer review information Nature Neuroscience thanks the anonymous
All authors declare no competing interests.                            reviewer(s) for their contribution to the peer review of this work.
                                                                       Peer reviewer reports are available.
Additional information
Extended data is available for this paper at                           Reprints and permissions information is available at
https://doi.org/10.1038/s41593-026-02345-6.                            www.nature.com/reprints.




Nature Neuroscience
Article                                                                                                         https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 1 | Hippocampal ripples detected using an alternative             counts for each time and entropy bin are plotted as bars above and to the right of
algorithm. a, An example ripple detected using the method discussed in               the 2D distribution, respectively. The majority of ripples occurred prestimulus
ref. 51 showing the raw LFP trace (top), the ripple band signal (middle) and the     and in high entropy levels. d, Ripple duration as a function of peri-stimulus time
standardized envelope (bottom). b, Grand average raw field potential centered        and information-theoretic measures. Ripple duration increased with entropy
on ripple peak (left) and peri-ripple wavelet spectrogram (n = 4653 ripple events    (that is, expected information gain) for prestimulus ripples, but decreased for
from 17 participants). c, Normalized ripple distribution across peri-stimulus time   poststimulus ripples (χ2(1) = 4.54, p = 0.032). Lines represent the mean, and
and entropy levels, accounting for the total number of trials in each entropy bin.   shaded areas represent the 95% confidence intervals. *p < 0.05.
Normalized ripple count for each time and entropy bin is color-coded; average




Nature Neuroscience
Article                                                                                                    https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 2 | Estimated marginal means of the 2D ripple count as a function of entropy and time bins. The top 5 estimated marginal means (emmean) are
marked in red, error bars represent upper and lower 95% confidence interval of the estimated marginal mean of the model (N = 17 patients).




Nature Neuroscience
Article                                             https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 3 | See next page for caption.




Nature Neuroscience
Article                                                                                                           https://doi.org/10.1038/s41593-026-02345-6

Extended Data Fig. 3 | Hippocampal ripples. a, An example of ripple detection,        low surprise (1–1.5 bits) at later time bins. This designation of 1–1.5 bits as ‘low
showing the filtered ripple band and Hilbert amplitude thresholds. Gray shading       surprise’ is mathematically determined by our task design (maximum surprise
indicates ripple duration. b, An example interictal epileptiform discharge            ~5 bits, minimum approaching 0). This range of surprise values corresponds, on
(IED). c, Ripple frequency as a function of time bin and peri-stimulus time; a        average, to a value of 1.67 bits of entropy. This plot is, therefore, consistent with
main effect of peri-stimulus time (F(1,16) = 6.02, p = 0.026), with more ripples      the increase in ripple count in the prestimulus period for upcoming entropy
observed prestimulus. Error bars represent standard error of the mean (N = 17         values around 1.6 to 1.9 bits. Thus, when we plot ripple occurrence as a function
patients). d, Normalized distribution of ripples as a function of peri-stimulus       of surprise, we observe no increase in ripple count for high surprise trials in the
time and level of stimulus-bound surprise. e, Cluster-based statistical test on the   poststimulus period. Instead, ripples cluster precisely following low-surprise
normalized ripple count in the 2D entropy-time histogram using TFCE revealed          trials, very late in the trial (which would correspond to prestimulus bins of the
an extended cluster (−800 ms to +800 ms) of increased ripple count for high           next trial). g, Distribution of surprise values in the task (error bars depict s.e.m.;
levels of entropy, with peaks around −800 to −600 ms prestimulus, −200 to 0           N = 17 patients). h, Speed (1/RT) as a function of prestimulus ripple status and
and 0 to 200 ms poststimulus. Color bar indicates t-value. f, Poststimulus ripple     entropy. For trials with a prestimulus ripple there was a faster response, more
rate as a function of surprise (normalized by the number of trials within each        pronounced (at a trend level) for entropy levels around 1.6 to 1.8 bits. Error bars
surprise bin). If entropy-associated prestimulus ripples were instead reflecting      represent the standard error of the mean. i, Ripple rate as a function of trial in
surprise-associated postprocessing of the previous stimulus, an increase in           block (N = 17 patients), error bars represented standard error of the mean.
ripple count should be observed for high surprise later in the trial. When locked     j, Ripple duration as a function peri-stimulus time and information-theoretic
to stimulus onset and going forward in time (up to 2.2 s poststimulus onset,          measures. Entropy by ripple peak time interaction (χ2(1) = 0.66 p = 0.41) and
which corresponds to the time between presentation of successive stimuli), such       surprise by ripple peak time interaction (χ2(1) = 0.17, p = 0.678). Lines represent
an increase was not observed, that is, the ripple rate in the top right quadrant of   the mean and shaded area the 95% CI.
the plot shows a low ripple rate. Instead, there is an increase in ripple count for




Nature Neuroscience
Article                                                                                                       https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 4 | Prestimulus hippocampal response to entropy split           between hippocampal gamma and entropy in trials with a prestimulus ripple
by whether a prestimulus ripple was present. Examining the hippocampal             (around the same time and frequencies as reported in Fig. 3a). Cluster t = −488.1,
response to entropy on trials in which a prestimulus ripple was present,           p = 0.0248 95% CI = 0.0043.
compared to trials without a prestimulus ripple, we found a negative association




Nature Neuroscience
Article                                                                                                    https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 5 | Entropy-related responses. a, Log-transformed            poststimulus. c, Log-transformed gamma power in occipital cortex prestimulus
gamma power in hippocampal prestimulus entropy cluster (shown in Fig. 3a),      entropy cluster (shown in Fig. 3b), as a function of trial in block. Shaded areas
as a function of trial in block. Shaded areas represent standard error of the   represent standard error of the mean. d,e, Occipital (d) and fusiform (e) cortex
mean. b, Hippocampal lower frequencies entropy responses, prestimulus and       lower frequencies entropy responses, prestimulus and poststimulus.




Nature Neuroscience
Article                                                                                                                https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 6 | Peri-ripple cortical modulation. a, Mean of log power                bars the group mean, and vertical error bars represent 95% confidence interval.
in occipital cortex cluster (N = 8) positively associated with entropy in the               Occipital gamma power was suppressed when there was a hippocampal ripple
prestimulus time window, split by whether there was a hippocampal ripple just               compared to when there was not. b, T-statistics of peri-ripple activity in occipital
before the significant effect or if there was not, two-tailed paired t-test (t(7) = 2,51,   cortex for poststimulus ripples. c,d, Peri-ripple activity in the fusiform for (c)
p = 0.0401, CI = (0.011, 0.377)); each point pertains to one patient, horizontal            prestimulus and (d) poststimulus ripples. *p < 0.05.




Nature Neuroscience
Article                                                                                                         https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 7 | Peri-ripple cortical modulation using an alternative          cluster positively associated with entropy in the prestimulus time window, split
detection algorithm. a,b, Cortical time-frequency analyses time-locked to the        according to whether there was a hippocampal ripple just before the significant
hippocampal ripple peak time, detected using the algorithm discussed in ref. 51.     effect or if there was not (N = 8), each point pertains to one patient, horizontal
Across all ripple trials there was an overall increase in high gamma power in        bars the group mean, and vertical error bars represent 95% confidence interval.
the fusiform (a) and occipital cortex (b) time-locked to hippocampal ripple.         Occipital gamma power was suppressed when there was a hippocampal ripple
c, Comparing trials with prestimulus versus poststimulus hippocampal ripples,        compared to when there was not; dependent t-test t(7) = 2.54, p = 0.038, Cohen’s
there was a reduction in occipital gamma power in trials with prestimulus ripples,   d = 0.9. *p < 0.05 (two-tailed).
compared to poststimulus ripples. d, Mean of log power in occipital cortex




Nature Neuroscience
Article                                                                                                          https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 8 | Hippocampal prestimulus ripples modulate fusiform              (extracted from significant cluster in trials with prestimulus hippocampal ripples
response to surprise using ripples detected with an alternative algorithm.            (red) compared to trials without ripples (blue). Ripples detected using the
a,b, Fusiform poststimulus association with surprise in trials with (a) prestimulus   algorithm described in ref. 51. Lines represent the mean, and shaded areas the
ripple compared to (b) without prestimulus ripples (no significant response           standard error of the mean.
to surprise observed). c, The increase in fusiform gamma power 40–95 Hz




Nature Neuroscience
Article                                                                                                       https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 9 | Surprise responses. a, Hippocampal low frequency response to surprise. b,c, Occipital cortex (b) low and (c) high frequency responses
to surprise.




Nature Neuroscience
Article                                                                                                      https://doi.org/10.1038/s41593-026-02345-6




Extended Data Fig. 10 | Fusiform poststimulus response to surprise is             trials were significantly higher in the gamma band (shaded rectangle, p < 0.001).
modulated by prestimulus hippocampal ripples and directed cortical-               This is analogous to the results presented in Fig. 5e, using a Granger causal
hippocampal information flow. a,b, Fusiform poststimulus association (a) with     analysis. PDC was computed using the FieldTrip toolbox in Matlab. The median-
surprise in trials with prestimulus ripple earlier than (b) without prestimulus   split data were downsampled to 250 Hz, and PDC was calculated for frequencies
ripples. c, Double-dissociation of fusiform gamma response to surprise (early     ranging between 0 and 80 Hz with a smoothing of 10 Hz. Shaded areas represent
and late clusters) in the presence and absence of a prestimulus hippocampal       the standard error of the mean. param. est.: parameter estimate of fusiform
ripple (n = 7). Error bars represent the standard error of the mean, p = 0.004.   response to surprise. h, Trials without prestimulus ripples show smaller
d,e, Granger causality (GC) analysis between the hippocampus and occipital        exponents (that is, flatter spectral slopes), indicating a shift toward excitation
cortex for entropy. f, Granger causality from the hippocampus to fusiform         (E>>I), whereas trials with prestimulus ripples show steeper slopes, indicating
for surprise. Shaded areas represent the standard error of the mean. g, Partial   a more balanced E/I state. Error bars represent the standard error of the mean
directed coherence (PDC) values for high (orange) versus low (green) surprise     (N = 7 patients).




Nature Neuroscience
                                                                                                                                                                                         nature portfolio | reporting summary
                                                                                                                                         Darya Frank
                                                                                                     Corresponding author(s): Bryan Strange
                                                                                                     Last updated by author(s): 30/04/2026




Reporting Summary
Nature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.




Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
           The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
           A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
           The statistical test(s) used AND whether they are one- or two-sided
           Only common tests should be described solely by name; describe more complex techniques in the Methods section.

           A description of all covariates tested
           A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
           A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
           AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

           For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
           Give P values as exact values whenever suitable.

           For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
           For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
           Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
                                                Our web collection on statistics for biologists contains articles on many of the points above.



Software and code
Policy information about availability of computer code
  Data collection        Data was collected in two locations: in Madrid ongoing intracranial EEG activity was acquired using XLTEK EMU128FS amplifier (XLTEK,
                         Ontario, Canada). In Zurich, intracranial recordings were acquired using a Neuralynx ATLAS system.

  Data analysis          Data was analysed in Matlab 2019b using the Fieldtrip toolbox (version 20200310) and custom-code available in the following Github
                         repository: https://github.com/frdarya/GenerativeRipples
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.



Data
                                                                                                                                                                                                    March 2021




Policy information about availability of data
 All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
    - Accession codes, unique identifiers, or web links for publicly available datasets
    - A description of any restrictions on data availability
    - For clinical datasets or third party data, please ensure that the statement adheres to our policy


All data needed to generate the figures, and over which statistics were computed, are available in the following Github repository: https://github.com/frdarya/
GenerativeRipples.


                                                                                                                                                                                                   1
                                                                                                                                                                              nature portfolio | reporting summary
Human research participants
Policy information about studies involving human research participants and Sex and Gender in Research.

  Reporting on sex and gender               Of the 18 patients who participated (17 who fulfill the inclusion criteria) 8 were males.

  Population characteristics                Eighteen medication-resistant presurgical epilepsy patients (8 males; age range 24-59, mean age = 37.3 SD = 11.7), see Table
                                            S1 for full details

  Recruitment                               We included in the study all patients that agreed to participate and that had electrodes implantation in the
                                            hippocampus. Implantation sites were chosen solely on the basis of clinical criteria.

  Ethics oversight                          All patients signed informed consent. The study had full approval from the local ethics committees of the Hospital Ruber
                                            Internacional, Madrid, Spain and Kantonale Ethikkommission, Zurich, Switzerland (PB-2016-02055).

Note that full information on the approval of the study protocol must also be provided in the manuscript.




Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

    Life sciences                        Behavioural & social sciences                  Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf




Life sciences study design
All studies must disclose on these points even when the disclosure is negative.
  Sample size              No statistical methods were used to pre-determine sample sizes but our sample sizes are larger or equal to those reported in previous
                           publications (Zheng et al, 2017; Mendez-Bertolo et al 2016; Costa et al., 2022).

  Data exclusions          We included in the analyses all patients that agreed to participate and completed the task (at least 25% of trials after exclusion due to
                           performance or electrophysiological artefacts) and that had electrodes implantation in the
                           hippocampus. Implantation sites were chosen solely on the basis of clinical criteria.

  Replication              No out-of-sample replications were attempted, single-subject datapoints are presented in main and supplementary figures to demonstrate
                           within-study between-subjects effects are consistent.

  Randomization            The design is a within-subject with trial-by-trial variation (no categorical conditions), the underlying probability distribution of stimuli was
                           randomly chosen per patient prior to the task from all possible combinations.

  Blinding                 Patients were not allocated into different experimental groups, therefore blinding was not relevant for data collection.
                           Participants were naive to the task goals.




Reporting for specific materials, systems and methods
We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.


Materials & experimental systems                                   Methods
n/a Involved in the study                                         n/a Involved in the study
           Antibodies                                                        ChIP-seq
           Eukaryotic cell lines                                             Flow cytometry
                                                                                                                                                                                         March 2021




           Palaeontology and archaeology                                     MRI-based neuroimaging
           Animals and other organisms
           Clinical data
           Dual use research of concern




                                                                                                                                                                                        2
Magnetic resonance imaging




                                                                                                        nature portfolio | reporting summary
Experimental design
 Design type                                Clinical protocol for structural scanning

 Design specifications                      Structural scans used for electrode contact localization.

 Behavioral performance measures            N/A


Acquisition
 Imaging type(s)                            Structural

 Field strength                             3

 Sequence & imaging parameters              T1

 Area of acquisition                        Whole-brain

 Diffusion MRI               Used                 Not used

Preprocessing
 Preprocessing software               SPM8 (http://www.fil.ion.ucl.ac.uk/spm).

 Normalization                        Data were normalized to MNI space

 Normalization template               MNI

 Noise and artifact removal           N/A

 Volume censoring                     N/A


Statistical modeling & inference
 Model type and settings              N/A

 Effect(s) tested                     N/A

 Specify type of analysis:         Whole brain            ROI-based              Both
 Statistic type for inference         N/A
 (See Eklund et al. 2016)

 Correction                           N/A


Models & analysis
 n/a Involved in the study
           Functional and/or effective connectivity
           Graph analysis
           Multivariate modeling or predictive analysis
                                                                                                                   March 2021




                                                                                                                  3
