Across nine amygdala SEEG patients, cross-patient decoding partly identified aversive blocks and underperformed within-patient decoding

A calibration-free decoder could support an affective closed-loop brain-computer interface before it has labeled data from a new patient. We evaluated whether amygdala SEEG band power could identify aversive (fearful-face) versus neutral (landscape) video blocks in a held-out patient. We used OpenNeuro ds003374 version 1.1.1 recordings from nine epilepsy patients, with 17 alternating 24-second blocks per patient. A shrinkage linear discriminant decoder trained on the other eight patients reached median block AUC 0.68 (mean 0.67); 7/9 patients had AUC above 0.5, 4/9 had permutation p < 0.05, and the one-sided group sign-flip p was 0.037. The median within-patient leave-one-neutral-aversive-pair-out AUC was 0.89 (7/9 p < 0.05; group p = 0.002). Cross-patient minus within-patient block AUC averaged -0.165 (two-sided p = 0.016). Cross-patient window-level median AUC was 0.57 (group p = 0.057). The decoder was strongly inverted in the only patient with a right amygdala alone (AUC 0.18). No band showed a consistent effect across patients. These results indicate partial transfer and substantial patient-specific information, and support short per-patient calibration in this sample. Interpretation is limited by nine patients, one session each, fixed block order, and video content confounded with emotional condition.

amygdala · SEEG · brain-computer interface · cross-patient decoding · band power · aversive video · linear discriminant analysis · patient-specific calibration