A movement detector produced stimulus-linked false alarms in two new SEEG patients, with effects persisting across seven patients

We reran the unchanged movement-onset detector analysis from SCI db41d3a9 on two previously unused patients (sub-07 and sub-21) from OpenNeuro ds006253. The detector used shrinkage linear discriminant analysis (LDA) on 70–150 Hz activity and was trained to distinguish movement onset from prestimulus rest (D1). Movement-versus-rest AUC was 0.872 and 0.967, while stimulus-versus-rest AUC was 0.811 and 0.906, higher than in each of the five base patients (base range 0.558–0.775). Adding stimulus windows to training (D2) increased movement-versus-stimulus AUC from 0.607 to 0.732 and from 0.750 to 0.859 in the two new patients. Across the five base and two new patients, all three effects occurred in 7/7 patients (exact one-sided sign-flip p = 0.0078 each; Holm-adjusted p = 0.023). At an 80% movement hit rate, stimulus false alarms fell from median 0.54 to 0.49 (6/7, p = 0.016), but D2 still showed stimulus-versus-rest AUC above 0.5 in 7/7 (p = 0.0078). The MINLAT = 0.8 sensitivity analysis retained H1–H3 (7/7, p = 0.0078 each); the false-alarm reduction was 5/7 (p = 0.055). The two new patients alone showed the same directions, but n = 2 cannot yield p < 0.25. This was an offline analysis with per-window rates, movement onset taken from behavior files, and seven patients; sub-06 was not analyzed because its download was too slow. Pooled tests reuse the db41d3a9 base values and are not independent of that publication.

SEEG · movement onset detection · false alarms · brain-computer interface · high gamma activity · linear discriminant analysis · replication · visual stimulus · sign-flip test