One 48-s calibration pair supports above-chance within-patient decoding of amygdala SEEG video blocks

In nine patients, a within-patient shrinkage LDA using six amygdala log band-power features decoded aversive (fearful-face) versus neutral (landscape) video blocks with mean held-out pair accuracy 0.77 after one 48-s calibration pair (Holm-adjusted p = 0.027). The group curve reached 90% of its seven-pair gain at two pairs (96 s); individual k90 values ranged from one to six pairs in eight patients, while one patient remained near chance. A decoder trained on the other eight patients had accuracy 0.61 (block AUC 0.67), including inverted accuracy for one patient. Adding other-patient data with a fixed 0.5 pooling weight did not significantly improve within-patient accuracy at any calibration length. Within-patient accuracy exceeded cross-patient accuracy numerically from two pairs, but this difference was not significant after Holm correction (smallest adjusted p = 0.11). The study includes nine patients and eight block pairs per patient. Block order was fixed, face versus landscape content was not separated from emotion, and label-free z-scoring used the full session, so the pipeline was not strictly causal online. Calibration subsets overlapped, the pooling weight was not tuned, and results may vary across patients.

amygdala · intracranial EEG · SEEG · calibration · shrinkage LDA · band power · within-patient decoding · aversive video