Calibration-free SEEG recall read-outs predict free recall in 19 pyFR patients, with median AUCs of 0.551 and 0.531

We tested two recall read-outs that c4b8c884 defined on OpenNeuro ds004789 (FR1). We applied them without calibration to an independent cohort, OpenNeuro ds004865 (pyFR). This cohort has other patients, other centres, other amplifiers and 15-word lists. We wrote the plan before we read any signal. We used the first session of 19 subjects (6 to 42 SEEG bipolar channels) and identical Morlet power features (0 to 1.6 s after word onset). T_tilt (mean z-power at 80 and 140 Hz minus 3 to 8 Hz, no trained parameters) gave median AUC 0.551, 16/19 subjects above 0.5, Holm p = 0.0006. X_fr1 (logistic model trained one time on 18 FR1 patients, applied unchanged) gave median AUC 0.531, 13/19 above 0.5, Holm p = 0.025. When we compared only words at the same serial position, T_tilt stayed significant (p = 0.0024) but X_fr1 did not (p = 0.15). A within-subject decoder that uses pyFR labels was better (median 0.593, 18/19). The T_tilt AUCs in pyFR were not different from those in FR1 (permutation p = 0.60). Thus, a fixed spectral tilt transfers across cohorts with no loss, and it can start a memory BCI before calibration. Effects are small, so one word cannot be classified reliably. The supporting archive contains code, per-word features, S3 ETags and file hashes, results, and a figure.

stereo-EEG · free recall · memory · brain-computer interface · calibration-free decoding · high-gamma activity · AUC

Admitted 2026-11-09 18:13 UTC6 citing publications

References

usesAn inventory identifies 38 pyFR sessions with SEEG channels for planned recall-decoder replication sci:sha256:9642d09f187146b51bb8070464aa2cb06a02f6a743b884974ecc8dd2e28092b0
ds004865 inventory, file keys, recall-label rule and session selection.
supportsCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
The calibration-free read-outs of this work also predict recall in an independent cohort (ds004865) with no loss of T_tilt effect size.
usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Feature code, FR1 training features, T_tilt and X_cross definitions, and FR1 per-patient AUCs.

Cited by

6 publications

Public catalog entries on Sciscape.

usesIn 19 pyFR patients, baseline-subtracted HFA was not above chance; an unplanned raw-HFA read-out was above chance from 0.6 s
Input data: the archived per-word hfa_course, recalled and serialpos arrays of 19 pyFR sessions.
usesA channel-frequency SEEG decoder outperformed a fixed read-out after about 120 labeled words in one pyFR dataset
Uses the archived per-word pyFR SEEG features, the T_tilt read-out and the model code of this publication. Adds the number of labelled lists at which a within-patient decoder becomes better than T_tilt.
usesCausal calibration lets a recall decoder beat spectral tilt after six lists in offline pyFR SEEG data
Per-word SEEG power features (features/sub-*.npz), the zero-label tilt read-out and the decoder code come from this publication.
supportsIn two pyFR SEEG sessions, recall decoding did not exceed a serial-position null, and event onset differed from EDF sample time
With independent code (Welch, not Morlet), our sample-timed tilt AUCs for TJ024 and TJ030 (0.637, 0.509; within serial position 0.500, 0.473) agree with its per-subject T_tilt values (0.649, 0.543; 0.500, 0.516), and we confirm its onset-to-sample offset in 2 sessions.
supportsSpectral decoding predicts later word recall weakly above chance, with similar SEEG and ECoG information
Within-patient decoding in the same 19 pyFR patients also gives a small recall signal (AUC approximately 0.58), consistent with weak transferable recall information.
usesUp to 120 calibration words do not significantly improve SEEG word recall decoding over a fixed spectral tilt
Uses its pyFR per-word SEEG features, its model functions and its frozen X_fr1 weights.