Across 18 SEEG patients at four sites, other-site training was non-inferior for word-recall decoding (median AUC difference +0.004; margin 0.02)

We tested whether a zero-calibration cross-patient SEEG word-recall decoder loses accuracy when trained only on patients from sites other than the test patient’s site. Using archived band-power features for 18 RAM FR1 patients from OpenNeuro ds004789 version 3.1.0 (session 0; four sites with 3, 7, 4, and 4 patients), we compared other-site training (OS) with random same-size training sets from all other patients (MIX). Median OS-minus-MIX AUC was +0.004 (bootstrap 95% CI −0.006 to +0.010; two-sided exact Wilcoxon p = 0.37); OS met the prespecified 0.02 AUC non-inferiority margin (one-sided p = 7.6×10⁻⁶). Same-site-only training was not better than same-size other-site training (median difference −0.007, p = 0.67). OS median AUC was 0.561, and decoder AUCs remained near 0.55. Per-site model weights were dissimilar, but random patient groups showed similarly low similarity (p = 0.40). These results support transfer across sites in this dataset, with limitations: only 18 patients and one task and session; site and sampling rate were confounded; features averaged across channels; and test-patient z-scoring used all words without labels. The 0.02 AUC non-inferiority margin was selected by the authors.

SEEG · memory BCI · word recall · cross-patient decoding · site transfer · logistic regression · AUC