Amygdala decoders generalized across 24-s aversive and neutral blocks in 9 patients, with no evidence of within-block code change

We tested whether shrinkage-LDA decoders of aversive versus neutral video blocks transfer across time within 24-s blocks. The analysis used archived features from 9 patients with amygdala SEEG: six z-scored log band powers in 2-s windows, averaged over amygdalae. Leave-one-neutral-aversive-pair-out AUCs were 0.664 for early-to-early, 0.683 for early-to-late, 0.633 for late-to-late, and 0.667 for late-to-early decoding. The pre-registered generalization loss, defined as the mean same-half AUC minus the mean cross-half AUC, was -0.026 (exact two-sided sign-flip p = 0.18; 95% t CI -0.068 to +0.015). One of 9 patients had a positive loss. The interval's upper bound excludes losses larger than 0.05 AUC. A descriptive 3 x 3 matrix with 8-s bins showed no diagonal advantage. Separation along the LDA axis did not decrease across the block (slope -0.006 SD per window, p = 0.72). Under the pre-registered rule, there was no evidence of within-block code change or habituation. The result applies to these tested conditions and the 0-24 s scale. Limitations include 9 patients with one session each, low power for small losses, coarse per-fold AUCs, fixed alternating block order, confounding of valence with video category, features averaged over amygdala contacts, uncorrected secondary analyses, and use of archived features that were not recomputed from raw data.

amygdala · stereo-EEG · affective decoding · band power · temporal generalization · shrinkage LDA · habituation · brain-computer interface