A fixed cross-patient logit offset did not improve SEEG recall decoding over a patient-only model in 18 patients

We tested whether a cross-patient recall decoder can reduce calibration data for within-patient SEEG word-recall decoding. In 18 patients from FR1, we trained patient models on the first 1, 2, 3, 4, or 6 lists (12 words per list) and tested on lists 7 onward. The test method added the cross-patient logit as a fixed offset to an L2 logistic regression. For the prespecified primary endpoint, the mean change in AUC versus the patient-only model over k=1, 2, and 3 was -0.004 (median -0.007); 6 of 18 patients had a positive change (exact two-sided Wilcoxon p=0.47). The offset model was not better than the patient-only model at any k, including within-serial-position AUC and calibration-only z-scoring sensitivity analyses. It was better than the cross-patient model alone at every k (Holm-adjusted p<=0.010). In an exploratory analysis, gains over the patient-only model tracked the cross-patient model’s gap to that model (Pearson r=0.96), consistent with the offset moving predictions toward the cross-patient decoder. The result is limited to one task, one session per patient, 18 patients, band-power features, linear decoders, and later lists in the same session. Cross-session and closed-loop use were not tested; the prior weight and regularization were not tuned.

SEEG · free recall · word encoding · recall decoding · cross-patient model · logistic regression · calibration · AUC · band power · transfer learning