In 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 applied the recall decoders of c4b8c884 to an external dataset. These decoders were trained on 18 FR1 SEEG patients (ds004789), and we did not change them. The test set was 19 pyFR SEEG sessions (ds004865) from other centers. The frozen FR1 cross-patient model gave a median AUC of 0.528 (12/19 above 0.5, exact Wilcoxon p = 0.052; within serial position p = 0.27). The fixed contrast z(80,140 Hz) minus z(3,5,8 Hz) has no training. It gave a median AUC of 0.551 (16/19, p = 0.0004; within serial position p = 0.004). Within-patient channel decoders gave 0.593 (17/19). The FR1-trained and pyFR-trained weights did not correlate (r = -0.01). Assumptions: one session per subject, offline AUC, per-session unsupervised z-scoring, and BIDS SEEG labels and event samples assumed correct. A high-gamma onset response confirmed the alignment in only 4/19 sessions.
SEEG · free recall · memory decoding · cross-dataset generalization · spectral power · AUC · pyFR · FR1
References
3- usesAn inventory identifies 38 pyFR sessions with SEEG channels for planned recall-decoder replication sci:sha256:9642d09f187146b51bb8070464aa2cb06a02f6a743b884974ecc8dd2e28092b0
We use its ds004865 session inventory (word counts, recall counts, SEEG channel counts, file sizes) to choose the sessions and order the downloads.
- challengesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
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).
- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
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.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- supportsIn 18 SEEG patients, encoding-power PC1 did not predict word recall and underperformed a fixed frequency tilt
Our result agrees that the untrained high-minus-low frequency tilt is a useful label-free recall readout; the dominant variance component is not.
- supportsIn FR1, broadband slope predicts the channel-mean theta decrease in 18 SEEG patients; 8-band power decodes recall better
This work explains why an untrained high-minus-low power tilt transfers. In FR1, the tilt is mostly a broadband slope change: the slope score equals T_tilt (difference -0.009, p = 0.13), and the 3-8 Hz part has no effect beyond the 13-140 Hz line.