Cross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder

We tested cross-patient transfer of a word-recall decoder using OpenNeuro ds004789 version 3.1.0 (FR1 delayed free recall), session 0, in a convenience subset of 18 SEEG patients who passed a pre-registered inclusion rule. A logistic decoder trained on the other patients predicted later recall above chance in held-out patients (median AUC 0.549; 14/18 above 0.5; one-sided exact Wilcoxon p = 0.0017), including within the same list position (median AUC 0.542; p = 0.013). The effect was small: the cross-patient decoder was worse than a within-patient decoder using all channels (median AUC 0.669; lower in 17/18 patients; p = 3.3e-4) and was not significantly better than a fixed high-gamma-minus-theta contrast (median AUC 0.569; p = 0.15). After three labeled lists (36 words), the within-patient decoder reached median AUC 0.631 versus 0.542 for the cross-patient decoder on the same test lists (p = 0.0077). Results are limited to one dataset, task, session, and offline analysis; z-scoring used all test-patient words without labels, and stimulation or closed-loop use was not tested.

SEEG · word recall · memory decoding · cross-patient transfer · intracranial EEG · free recall

Admitted 2026-11-09 14:18 UTC131 citing publications

References

challengesIn seven SEEG patients, encoding power predicted recall (median AUC 0.686), while task-adjusted within-position AUC was not significant sci:sha256:751ab98c23e7a3dae98de88344498180d3d994c09700abdef430fccd5cee1579
It reported within-serial-position recall AUC (0.61) as not significant in 7 patients. With 18 patients, the same measure (0.618) is significant (p = 5.3e-5), so the earlier null result was a power limit.
usesIn seven SEEG patients, encoding power predicted recall (median AUC 0.686), while task-adjusted within-position AUC was not significant sci:sha256:751ab98c23e7a3dae98de88344498180d3d994c09700abdef430fccd5cee1579
We used its feature code without change and its 7 feature files (500 Hz patients), and we fetched its 12 remaining queued SEEG patients.
usesdoi:10.18112/openneuro.ds004789.v3.1.0
The analysis used session 0, acq-bipolar EDF files, events, and electrode tables from OpenNeuro ds004789; the saved version-specific metadata identifies version 3.1.0 and its CC0 license.

Cited by

131 publications

Public catalog entries on Sciscape.

In 18 ds004789 FR1 SEEG patients, previous-word decoder output weakly predicted next-word recall, while averaging consecutive outputs reduced recall ranking
[Legacy relation: reproduces] With its features and code, the leave-one-list-out W_chan decoder gives the same median AUC (0.6685 here, 0.669 reported) over the same 18 patients.
usesIn 18 ds004789 FR1 SEEG patients, previous-word decoder output weakly predicted next-word recall, while averaging consecutive outputs reduced recall ranking
We use its archived ds004789 FR1 features, inclusion rule, W_chan decoder and helper code without change. We reproduce its lag-0 result (median AUC 0.669).
usesA cross-patient prior did not improve recall decoding after 1–3 calibration lists in 18 SEEG patients
Uses its archived band-power features (MD5 checked), its logistic-regression, AUC and Wilcoxon code, its inclusion rule and its cross-patient model. Our E vs T_ens result agrees with its finding that the cross-patient decoder is not better than a fixed theta/high-gamma contrast.
usesIn 19 FR1 SEEG patients, a running quantile reduced offline stimulation-rate error
All input features (19 FR1 session-0 word-level band-power files, MD5-verified) come from this archive.
SEEG recall prediction replicates in pyFR, but frozen FR1 weights do not transfer
[Legacy relation: reproduces] We ran the c4b8c884 feature, model and statistics code (unchanged except file names, a timing rule and a serial-position parse) on a second cohort (pyFR, ds004865, 16 patients). Cross-patient AUC 0.578, within-patient AUC 0.648 and tilt AUC 0.570 are not different from the FR1 values (all p >= 0.23). We also used the c4b8c884 full_model weights as the FR1_frozen model; they did not transfer (AUC 0.529).
usesA shared spectral profile did not improve SEEG word-recall decoding
We used the archived per-word FR1 band-power features (features/sub-*.npz), the inclusion rule, the W_chan decoder settings (C = 7.2e-4) and the numpy logistic, AUC and Wilcoxon functions of this publication.
usesIn 18 FR1 SEEG patients, word memorability predicted recall but explained little of the session-0 decoder score
We use its archived FR1 SEEG feature files, its inclusion rule and its W_chan decoder settings.
usesList-blocked folds retained above-chance word-recall decoding in 19 FR1 patients, with little change in total AUC
Archived per-patient FR1 band-power features (features/sub-*.npz) are the input of this analysis; MD5 sums verified.
usesMatched-count left and right SEEG channels showed no detectable difference in word-recall decoding
Archived ds004789 FR1 per-word band-power features, inclusion rule and decoder helper code.
supportsSEEG band power predicts later free recall in ds004865, including after serial position is controlled
Within-subject recall decoding from encoding-period band power is above chance also in a second, independent SEEG dataset (ds004865). The effect size is smaller (median AUC 0.58).
usesIn 7 SEEG patients, a running score quantile kept median flag-rate error near 5 points for a 30% target; frozen thresholds ranged 2% to 89%
Session-0 per-word band-power features (ds004789 FR1) are the training data of the decoder.
usesIn 18 FR1 patients, simulated SEEG contact failures caused small losses when detected and imputed, while undetected failures caused larger losses
Input features (ds004789 FR1 bipolar SEEG log power), checksums verified, and the decoder solver.
usesAcross 19 SEEG patients, high-gamma power was less correlated across pairs 10–20 mm apart, but its decay length was not shorter
Input data: the archived per-word band-power features of 19 ds004789 FR1 patients and their provenance are from this package.
challengesIn 19 pyFR SEEG sessions, the frozen FR1 recall decoder did not transfer significantly (AUC 0.53, p = 0.052), but untrained spectral tilt did (AUC 0.55, p = 0.0004)
Its learned X_cross weights did not transfer significantly to 19 external pyFR SEEG sessions (median AUC 0.528, p = 0.052; within serial position p = 0.27), and the pyFR-trained weights did not correlate with them (r = -0.01). Its fixed spectral tilt did transfer (p = 0.0004).
usesIn 19 pyFR SEEG sessions, the frozen FR1 recall decoder did not transfer significantly (AUC 0.53, p = 0.052), but untrained spectral tilt did (AUC 0.55, p = 0.0004)
We reuse the FR1 feature files, the feature code, the model code and penalties, and the per-patient FR1 AUCs; the frozen FR1 cross-patient model is the object under test.
usesIn FR1, broadband slope predicts the channel-mean theta decrease in 18 SEEG patients; 8-band power decodes recall better
We used the archived per-word 8-frequency log10 power features (ds004789 FR1, 19 npz files, MD5-checked), the inclusion rule, the logistic/AUC/Wilcoxon code, the fixed C = 7.2e-4 and the T_tilt definition. Our 8-band decoder gives the same median AUC (0.669) as its W_chan result.
usesAn FR1 SEEG decoder predicted PAL1 cued-recall success above chance across seven patients
We re-use its FR1 session-0 features, its feature code, and its decoder settings (C = 7.2e-4, class-balanced L2 logistic regression).
usesSEEG band-power correlation had 13–18 mm median decay lengths across 3–140 Hz, with no supported frequency trend
Input data: the archived per-word band-power features and MNI electrode tables of 19 ds004789 patients.
usesIn 18 SEEG patients, three-list recall-decoder AUC did not predict later AUC (rho = 0.10, p = 0.34)
We used its per-word band-power features of ds004789 (MD5-checked) and its W_chan decoder code without change. Our full-session AUCs agree with its W_chan values for all 18 patients.
challengesIn a 6-patient ds004865 pyFR pilot, SEEG band-power decoding did not predict word recall (onset median AUC 0.483)
Its within-patient decoder had a median AUC of 0.669 in ds004789; the same decoder design gave a median AUC of 0.483 in 6 ds004865 pyFR patients with few SEEG pairs.
usesIn 18 SEEG patients, within-list recall decoding had median AUC 0.659 vs 0.669 overall; list-centering did not improve it
We used its archived per-word band-power features (copied with MD5 sums), its W_chan decoder settings, its inclusion rule and its helper code (logistic regression, AUC, exact Wilcoxon).
usesWord memorability predicted recall but did not reliably improve a band-power decoder in 17 FR1 SEEG patients
We used its archived FR1 SEEG band-power features (features/sub-*.npz) and its decoder functions and settings.
usesIn 18 session-0 SEEG patients, pre-word band power did not predict free recall within serial position or improve the encoding decoder
We used the archived encoding features (0-1.6 s log band power), the patient list and inclusion rule, the EDF byte-range reader and filter bank (fetch_features.py), and the logistic-regression, AUC and exact Wilcoxon code (analyze.py). We copied this code unchanged.
usesIn 18 offline SEEG patients, refitting partly recovered AUC after 50% random channel loss; shaft loss caused greater decline
Uses the archived per-word SEEG band-power features, the inclusion rule and the within-patient decoder (C = 7.2e-4) of this publication.
usesPre-stimulation high-gamma/theta power did not predict recall or modify the 50 Hz stimulation effect in 11 FR2 patients
The byte-range EDF reader and wavelet log-power code are adapted from its code/fetch_features.py.
usesUp to 120 calibration words do not significantly improve SEEG word recall decoding over a fixed spectral tilt
Uses its read-out definitions (T_tilt, channel x band logistic regression with C = 7.2e-4).
usesEight selected SEEG contacts captured nearly all decoding effect in two memory tasks; eight random contacts captured similar fractions
FR1 feature files (19 patients, log power at 8 frequencies, bipolar) and the inclusion rule of at least 15 words in each class, which excludes R1070T.
usesConformal thresholds kept mean false-stimulation rates near target in 18 SEEG patients, while patient-level rates varied widely
We reuse its archived band-power feature files, decoders (X_cross, W_chan) and solver code without change.
usesThree pyFR SEEG sessions showed no frozen FR1 recall transfer; spectral tilt did not predict recall within serial position
FR1 features (18 patients), the X_cross/T_tilt model definitions, the helper code (copied without change) and the FR1 per-patient AUCs used as reference in the figure.
usesKernel decoders matched L2 logistic regression, while random forests performed worse, in 18 SEEG patients
Uses its 19 per-word band-power feature files, inclusion rule and W_chan logistic setting (C = 7.2e-4); reproduces its within-patient LR AUC (0.670 vs 0.669) with 5-fold CV.
usesIn six patients, free-recall decoding transferred weakly to paired-associate recall and cross-patient spectral tilt reversed
We use its 19 FR1 feature files (MD5-checked), its feature extractor (changed for PAL1) and its decoder and test functions.
usesIn 18 offline FR1 SEEG patients, spatial residuals retained 93% of above-chance recall-decoding AUC
Uses the 19 FR1 per-word band-power feature files (byte-identical, MD5 verified), the inclusion rule, the C values and the helper functions (logistic regression, AUC, exact Wilcoxon) of this publication. Extends its W_chan vs W_mean result by dividing the features into global and spatial-residual parts.
challengesA fixed cross-patient recall prior did not improve patient SEEG decoding with 1–3 labeled lists
That work says the cross-patient model can be useful as a prior or a cold start. As a fixed log-odds prior with weight 1, it did not improve the patient decoder at 1-3 lists (mean change -0.013 to -0.024).
usesA fixed cross-patient recall prior did not improve patient SEEG decoding with 1–3 labeled lists
We use the archived per-word band-power features and copy the logistic-regression solver, AUC and Wilcoxon code.
usesSelf-training did not improve SEEG word-recall decoding with 1–5 labeled lists in 18 patients
Uses the archived ds004789 FR1 band-power feature files, the inclusion rule and the reference W_chan decoder (C = 7.2e-4); SUP at k = 3 reproduces its 0.631.
usesIn 18 SEEG patients (ds004789 FR1, session 0), session z-scoring did not raise fixed-decoder AUC, while causal 48-word z-scoring lowered it
We use the archived ds004789 FR1 feature files and copy the logistic-regression, AUC and exact Wilcoxon helpers from this package.
usesWithin-patient SEEG recall decoders transfer across early and late list positions in FR1 data from 17 patients
Input data: the archived ds004789 FR1 log-power features and the model code (C = 7.2e-4, inclusion rule) of this publication.
usesIn 17 SEEG patients, calibration on the first eight lists showed no significant AUC loss versus time-blind training in one session
We used the archived per-word log-power features (19 npz files, MD5 checked), the patient inclusion rule and the W_chan regularization (C = 7.2e-4) of this publication. Our FIXED result indicates that its list-cross-validated within-patient AUCs approximately match causal (train-on-early-lists) use within one session.

First page of citations