A recall classifier showed little word-specific signal in 81 pure-SEEG patients

We evaluated a RAM-style L2 logistic classifier of later recall from encoding band power in 81 pure-SEEG patients performing FR1, using session 0 from OpenNeuro ds004789 version 3.1.0. The full model used five bands from a Hann periodogram over 0–1.6 s and leave-one-list-out cross-validation. Its mean AUC was 0.516 (AUC − 0.5 = +0.016; 95% bootstrap CI −0.001 to +0.035; confirmatory sign-flip p = 0.078, Holm p = 0.157). A two-feature global HFA/theta model was at chance (AUC − 0.5 = −0.009, p = 0.33); the full model exceeded it by 0.025 (95% CI +0.005 to +0.046; Holm p = 0.043). However, the full model's serial-position-stratified AUC was 0.485 (AUC − 0.5 = −0.015; 95% CI −0.034 to +0.004; Holm p = 0.157). Ten within-list label shuffles per subject gave a mean null AUC of 0.489; the full model exceeded this estimated null by 0.027 (95% CI +0.007 to +0.047, secondary p = 0.009, not Holm-corrected). The full AUC exceeded the stratified AUC by 0.031 (95% CI +0.021 to +0.042, exploratory p < 0.0001). These results indicate at most a very small prediction signal, which does not persist when comparing words at the same serial position. Most of the observed group-level signal is consistent with list position rather than word-specific encoding state. The result applies to this feature set and session 0; the shuffle estimate is based on only 10 shuffles per subject.

free recall · SEEG · memory · logistic regression · cross-validation · serial position · AUC · encoding power