Learn-then-Test abstention kept acted-on SEEG move-vs-rest errors low but acted for only 2 of 15 patients at a 10% error target (offline LDA, 2 OpenNeuro datasets, later trials of one session)

We tested whether a confidence threshold with a finite-sample risk bound (Learn-then-Test, fixed-sequence binomial tests; epsilon = 0.10, delta = 0.10) could limit the error of acted-on SEEG move-vs-rest decisions on later trials from the same session. The archived features cover 15 patients in OpenNeuro ds004473 and ds006253. At the 10% target, Learn-then-Test acted for 2/15 patients, with mean test coverage 0.059 and mean error 0.073; it abstained on every window for 13/15. A plug-in threshold acted for 8/15 (mean coverage 0.319, mean error 0.135), and 5/8 exceeded 10% error. Conformal single-label decisions acted for all 15 and had mean error 0.223. On test data, only 6/15 patients had any threshold with error at most 0.10 and coverage at least 0.20. The pre-registered tests at epsilon = 0.10 had only two patients with actions and did not support or reject the hypotheses. At epsilon = 0.20, Learn-then-Test acted for 7/15 patients (mean coverage 0.298, mean error 0.139); these are secondary results. This offline study used one session per patient, two classes, and one linear decoder. Windows from the same trial may not be independent, although the binomial test treats them as independent. The dataset release versions used for the archived features are not established.

Learn-then-Test · selective classification · abstention · risk control · SEEG · brain-computer interface · linear discriminant analysis · conformal prediction