SEEG movement decoding changes with latency, while a pooled decoder stays within 0.007 AUC of time-specific decoders

In an offline analysis of archived 70–150 Hz SEEG high-gamma features from 15 patients in OpenNeuro ds006253, we tested movement-versus-pre-stimulus-rest decoding across ten 0.2 s bins from −1.0 to +1.0 s around mouse-movement onset. Using a time-ordered 60/40 trial split, Ledoit–Wolf shrinkage LDA, and temporal generalization, AUC declined as train–test latency increased. A pre-registered comparison found a +0.098 AUC difference (95% bootstrap CI +0.068 to +0.129; 15/15 patients; Holm p = 0.00009) between the 0.2 s and at least 1.0 s train–test distances. An exploratory comparison against each test bin’s diagonal AUC found losses of −0.010 at 0.2 s and −0.086 at 1.4 s. Before movement, time-specific decoders exceeded a decoder trained at [0, 0.2] s by +0.020 AUC (11/15; Holm p = 0.017), a small effect with median +0.012 and a +0.115 contribution from patient 03. A decoder trained on all ten bins differed from time-specific decoders by −0.007 AUC (95% CI −0.017 to +0.003; Holm p = 0.214); this is no detected difference, not evidence of equivalence. The result supports latency-dependent changes in the movement-related pattern, while a detector trained across the full latency range performed nearly as well as time-specific decoders in this task. Limits include offline single-session data from 15 of 22 patients, pre-movement bins that overlap stimulus and decision processing, and an AUC that is not a continuous-data false-positive rate.

SEEG · high-gamma activity · brain-computer interface · temporal generalization · movement decoding · shrinkage LDA · AUC · OpenNeuro ds006253