Past-only normalization kept cross-patient SEEG recall decoding above chance in 18 patients

We tested whether a zero-calibration cross-patient word-recall decoder required non-causal normalization of a new patient's features. Using archived features from OpenNeuro ds004789 v3.1.0 (FR1 free recall, session 0), we analyzed 18 patients with bipolar SEEG. The unchanged L2 logistic decoder was trained on the other 17 patients; only held-out patient z-scoring changed. With test words from lists 5 and later, first-list normalization (12 unlabeled words) yielded median AUC 0.542 (14/18 patients above 0.5, p = 0.002), with a median difference of +0.0001 versus full-session normalization (95% CI -0.0036 to +0.0048). All causal variants had median AUC 0.540-0.543 and did not differ from full-session normalization after Holm correction. Balanced accuracy at threshold 0 remained above chance for causal variants (median 0.523-0.532, p = 0.010-0.030); without SD scaling, it was not significant (0.526, p = 0.071). In a post hoc descriptive analysis, causal updates kept the predicted-positive fraction nearer 0.5 than frozen statistics. This offline result suggests one unlabeled list can support ranking, while causal updating can help maintain a fixed threshold as signal features drift. The decoder remained weak (AUC about 0.54). Limits include one dataset, task and session per patient, offline features, and full-session normalization of training patients.

SEEG · word recall · cross-patient decoding · causal normalization · brain-computer interface · AUC · balanced accuracy