A session-0 word-recall decoder predicted session-1 recall above chance in 7 SEEG patients
This preregistered study tested whether a word-recall decoder trained on session 0 transferred to session 1 in patients performing FR1 free recall. Nine patients from OpenNeuro ds004789 had usable session-1 recordings; seven met the preregistered requirement of at least 15 recalled and 15 not-recalled words in each session. Features were log10 Morlet power at 3, 5, 8, 13, 22, 40, 80, and 140 Hz, averaged 0–1.6 s after word onset on bipolar SEEG channels shared across sessions. An L2 logistic regression decoder (C = 7.2e-4) trained on session 0 and tested with label-free session-1 z-scoring reached median AUC 0.617; six of seven patients were above chance (exact Wilcoxon, Holm-adjusted p = 0.047). Its median difference from within-session decoding was -0.021 (95% CI -0.070 to +0.043; Holm p = 0.94), which does not establish equivalence. Re-normalization changed AUC by +0.004 (Holm p = 0.94). Session means shifted 13 times more than the odd/even-list split within session 0; this shift changes a linear decoder's score offset, not its AUC. Exploratory threshold results suggest that label-free re-normalization kept a fixed threshold usable. The study included only seven patients, had low power for small differences, and did not analyze session 2 because extraction stopped at the time limit. The analyzed files were read from the unversioned OpenNeuro S3 mirror with recorded URLs, ETags, byte ranges, and hashes; the exact dataset release is uncertain.
SEEG · memory · free recall · brain-computer interface · cross-session decoding · logistic regression · AUC · feature normalization · OpenNeuro
References
3- supportsRecall prediction from SEEG encoding power transferred between sessions in three patients sci:sha256:149a56361e7d9d665814a5db2969baa51f4f548a85f71cabae25c80e0318fb43
In 7 different patients with a pre-registered test, the session-0 decoder transferred to session 1 above chance (median AUC 0.617, Holm p = 0.047), it was not shown to be worse than within-session decoding, and z-scoring changed AUC by about 0.004. This agrees with the 3-patient result of 211e5dcf.
- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Session-0 per-word features (features/sub-X.npz) and the feature-extraction code; our session-1 extractor is a copy of its fetch_features.py with the session as a new argument.
- usesFixed L2 penalty outperforms per-patient tuning for SEEG recall decoding in 18 FR1 patients sci:sha256:ac34e9f829a95f5eed29959d7e06e96828d7691379f8f036f5e5ee3868057afb
We use the fixed penalty C = 7.2e-4, which this work showed to be better than per-patient tuning in FR1 SEEG.
Cited by
1 publicationPublic catalog entries on Sciscape.
- 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-1 per-word features, inclusion rule and decoder settings. This work extends its exploratory note that frozen session-0 thresholds gave session-1 flag rates of 1% to 96%.