In six patients, free-recall decoding transferred weakly to paired-associate recall and cross-patient spectral tilt reversed

We tested if a SEEG memory-encoding decoder that we train on free recall (FR1, OpenNeuro ds004789) also predicts cued recall of word pairs (PAL1, OpenNeuro ds005059) in the same patients, without PAL1 labels. We fixed the plan before we computed PAL1 features. Features: log10 power at 8 frequencies (3-140 Hz), 0-1.6 s after word or pair onset, on bipolar channels that are in both tasks, z-scored in each task. Primary result (n = 6, one patient excluded with 14 recalled pairs): the same-patient FR1 channel x frequency model gave median AUC 0.612 on PAL1 (4 of 6 > 0.5; exact one-sided Wilcoxon p = 0.078). This is not significant. It was significant by permutation in 2 patients. A decoder trained in PAL1 gave median AUC 0.665 (6 of 6; p = 0.016). The cross-patient FR1 channel-mean model gave AUC < 0.5 in all 6 patients (median 0.409; two-sided p = 0.031). The reason: in PAL1, more 3-22 Hz power and less 40-140 Hz power came before correct recall. This is the opposite of the FR1 spectral tilt. FR1 and PAL1 decoder weights were almost uncorrelated (median r = 0.10). Secondary: with a control for study position, FR1-to-PAL1 AUC was > 0.5 in 6 of 6 (p = 0.016). With 10 recalled pairs as the limit (n = 7), p = 0.039. For memory BCIs: a free-recall encoding classifier, and mainly a generic spectral-tilt rule, can have the wrong sign in a different memory task. The package has the code, the plan, all feature files with S3 provenance, results, a figure and the commands to reproduce the work.

SEEG · memory encoding · free recall · paired-associate learning · cued recall · brain-computer interface · decoder transfer · spectral tilt · AUC