A channel-frequency SEEG decoder outperformed a fixed read-out after about 120 labeled words in one pyFR dataset

We measured how many labeled words a within-patient decoder needs to predict later free recall from SEEG recorded during word encoding, compared with a fixed spectral-tilt read-out (T_tilt) that uses no labels. We used archived per-word log Morlet power features from 19 patients in one session each in the pyFR delayed free-recall task (OpenNeuro ds004865). For each patient, we compared the AUC of T_tilt, a channel-by-frequency L2 logistic regression (W_chan), and a channel-mean decoder (W_mean) on the same held-out words, after training on 40 random draws of 1, 2, 3, 4, 6, 8, 10, or 12 lists (15 words per list). With one list, W_chan performed worse than T_tilt (median paired ΔAUC −0.028; 5/19 patients had a positive difference; exact two-sided Wilcoxon p = 0.0323). Differences were not significant for 2–6 lists. With 8 lists (about 120 words), W_chan performed better (median ΔAUC +0.019; 14/17 positive; p = 0.00557), remaining significant after Bonferroni correction across the eight list counts. At 10 and 12 lists, median differences were +0.028 (p = 0.00168) and +0.037 (p = 0.00919). A within-serial-position analysis showed a positive significant difference from 6 lists (p = 0.0159). W_mean was never significantly better than T_tilt through 12 lists. Median AUCs were small (0.54–0.58). This is an offline result from one task, one dataset, and one session per patient. Session-wide z-scoring used unlabeled test-word statistics; causal normalization and chronological training were not tested. The number of patients varied with training-list count, and fixed regularization can affect the crossover.

SEEG · free recall · memory decoding · brain-computer interface · calibration · AUC · intracranial EEG · logistic regression