Causal 0.5-s score smoothing increased SEEG key-press hits and reduced false detections in 8 subjects

We tested causal temporal smoothing of archived out-of-fold LDA scores for a self-paced SEEG key-press switch. The decoder was not retrained. In 8 subjects (246 presses; 32 min), a causal 0.5-s moving mean produced 66 hits (27%) versus 45 (18%) for raw scores, and 1.31 versus 2.00 false detections per minute. Press-level exact McNemar p = 0.0003; false detections were lower in 7 of 8 subjects (Wilcoxon p = 0.016). Median added latency was 0.04 s. The subject-level hit-rate test was not significant (p = 0.16), and most of the hit gain came from 3 subjects; smoothing did not repair weak decoders. All variants exceeded the 1 false detection/min target. Data were limited to 240 s per subject from an offline cued task, and this was not an online closed-loop test. Scores came from an archived 10-Hz stream; the archived-threshold reproduction yielded 22 of 26 original hits. The package preserves analysis code, archived inputs, results, hashes, and the figure.

SEEG · brain-computer interface · score smoothing · false detections · key-press decoding