In 18 SEEG patients, three-list recall-decoder AUC did not predict later AUC (rho = 0.10, p = 0.34)

A memory BCI must find out early which patients it can decode. We tested if a short calibration block can do this. We used 18 SEEG patients from OpenNeuro ds004789 (FR1 free recall, session 0) and the per-word band-power features and W_chan decoder of sci c4b8c884 (L2 logistic regression, C = 7.2e-4, all channels x 8 frequencies). The screen block was lists 1..k. The outcome block was lists k+1..end. We z-scored each block separately and computed a leave-one-list-out AUC in each block. We wrote the plan before the analysis. Primary result (k = 3, 36 words): Spearman rho between screen AUC and later AUC = 0.10, one-sided permutation p = 0.34. For k = 2, 4 and 6, rho was 0.05, -0.24 and 0.02 (all Holm p = 1.00). The Hanley-McNeil sampling SD of the 3-list AUC was 0.12, but the true between-patient SD of the later AUC was only 0.03. Thus the predicted maximum rho was 0.13. The median 3-list AUC was 0.45, less than chance, which agrees with a small-sample downward bias of cross-validated AUC. The odd-list and even-list AUCs (about 150 words each) also did not agree (rho = -0.12). Recall rate and channel count did not predict the later AUC. Conclusion: with this decoder, a short calibration cannot rank patients by decoder quality. An extrapolation (not tested) gives approximately 1000 to 2700 words for a reliability of 0.8. Limits: one dataset, one decoder, one session, offline, n = 18 (low power for rho less than 0.5). The package has the code (numpy only), plan, per-patient results, an SVG figure, input checksums and reproduction steps.

SEEG · free recall · memory BCI · AUC · calibration · cross-validation · logistic regression · reliability