Five pre-calibration patient properties did not predict recall-decoder AUC in 18 SEEG patients

In 18 patients with stereo-electroencephalography (SEEG) who completed session 0 of the FR1 free-recall task in OpenNeuro ds004789, five patient properties known before decoder calibration did not predict the offline AUC of an archived within-patient word-recall decoder. The properties were bipolar-channel count, medial temporal lobe (MTL) bipolar-pair count, shaft count, recall rate, and sampling rate. For the primary W_chan outcome, Spearman rho values ranged from -0.21 to +0.22; all two-sided permutation p values were at least 0.39 and all Holm-adjusted p values were 1.0. Results for AUC within serial position were also not significant (rho -0.30 to +0.28; p at least 0.22). Median W_chan AUC was 0.688 at 500 Hz (7 patients) and 0.660 at 1000 Hz (11 patients). With 18 patients, a two-sided test at alpha 0.05 detects only |rho| > 0.47, so small and medium effects remain possible. These findings do not establish that the properties have no effect. The outcome is from one fixed L2 logistic-regression decoder evaluated by leave-one-list-out; other decoders, tasks, sessions, or patient groups may differ. In this sample, these pre-calibration properties were not sufficient to identify patients with better decoder AUC, and labeled calibration lists remain necessary for each patient.

stereo-electroencephalography · free recall · memory decoding · brain-computer interface · AUC · patient properties · permutation test