Amygdala units and macro SEEG had no consistent decoding winner; combined median AUC was 0.80 at 2 s and 0.92 at 8 s in five patients

We tested if sorted amygdala units on micro-wires identify aversive (fearful-face) video blocks against neutral (landscape) blocks better than macro bipolar SEEG band power from the same depth electrodes. Data: OpenNeuro ds003374 v1.1.1, the 5 patients with sorted units (42 units, 3-16 per patient), 8 neutral-aversive pairs of 24-s blocks. Features: sqrt(spike count) per unit, or log power in 6 bands (1-150 Hz) per bipolar channel, in 1-8 s windows. Decoder: shrinkage LDA. CV: leave-one-pair-out with AUC inside each held-out pair. Null: all 256 within-pair label swaps. At 2 s, units minus macro AUC was -0.29 to +0.20 (3 of 5 patients positive, mean -0.025, sign-flip p = 0.81). The patient with no macro read-out (AUC 0.46) had a unit read-out (0.65; 0.75 at 8 s). The 2 patients with the strongest macro read-out (0.85, 0.86) had the weakest unit read-out (0.60, 0.57). The combined decoder was near the better single source in each patient (mean difference -0.015 at 2 s, +0.019 at 8 s). Its median AUC was 0.80 at 2 s and 0.92 at 8 s, with exact p < 0.05 in 5 of 5 patients at 8 s (macro 3 of 5, units 1 of 5). With reverse pairs (aversive first), the macro and combined results stayed, but the unit read-out decreased (median 0.65 to 0.59 at 2 s), so slow firing drift can explain a part of it. Limits: n = 5 (the group test cannot reach p < 0.05), faces vs landscapes, one session, offline.

amygdala · intracranial EEG · single-unit activity · neural decoding · affective neuroscience · brain-computer interface · OpenNeuro