In 8 SEEG subjects, later-trial conformal coverage was near nominal, but single-label key-press decisions had 24% error

We evaluated Mondrian split-conformal prediction for an offline, cued SEEG key-press-versus-rest decoder in 8 subjects. A shrinkage LDA model trained on trials 1–60 was calibrated on trials 61–100 and tested on trials 101 onward; 100 random train/calibration/test assignments served as an exchangeable control. At α = 0.10, mean class coverage on later trials was 0.874 versus nominal 0.90 (p = 0.17), and random-split coverage was 0.902. At α = 0.05 and 0.20, ordered-trial coverage was 0.927 and 0.771, respectively, about 0.02–0.03 below nominal and not significantly different from nominal or random splits. Non-inferiority within a 0.05 margin was not shown (p = 0.063–0.16). At α = 0.10, 60% of later windows received one label, but 24% of those single-label decisions were wrong; the random-split error was 21%. Thus, near-nominal set coverage did not limit errors among decisions that a BCI might act on. The study had 8 subjects and low power; it used balanced, offline, cued fixed windows, and only 40 calibration windows per class. The move window included the key press, so this was not a causal or self-paced detector.

split conformal prediction · SEEG · brain-computer interface · key-press detection · coverage · selective classification · shrinkage LDA