An FR1-trained SEEG classifier weakly predicts catFR1 recall in 22 patients, but HFA memory maps do not correlate

We tested whether SEEG features associated with later recall transfer between unrelated-word free recall (FR1) and category-word free recall (catFR1) in the same patients. In 22 patients with session 0 of both tasks, we used archived per-word log10 band-power features for matched bipolar pairs, computed from 0–1.6 s after word onset in five bands. An L2 logistic classifier trained on all FR1 words achieved mean catFR1 AUC 0.517 (14/22 above 0.5), against a within-list shuffle-null mean of 0.493 (1,000 shuffles; p = 0.002, Holm-adjusted p = 0.004). Transfer remained after serial-position removal (AUC 0.515, p = 0.009) and list centring (AUC 0.519, p = 0.027). The effect was small; only 2 of 22 patients were individually significant. Reverse transfer had AUC 0.520 (p = 0.091). An HFA-only classifier did not transfer significantly, and pair-level HFA subsequent-memory maps did not correlate between tasks (mean Fisher z = 0.020, p = 0.32) or show significant split-half reliability within either task. The analysis used one fixed regulariser, a periodogram rather than RAM Morlet features, and session 0 only; task order and time between tasks were unknown. These results indicate weak multivariate cross-task transfer in this sample, with no corresponding evidence for transferable HFA maps.

SEEG · free recall · memory · classifier transfer · band power · high-frequency activity · catFR1 · FR1