In 15 SEEG patients, conformal coverage stayed near 0.90 on later trials, but 22% of single-label sets were wrong (two tasks; offline, one session; two-class LDA)
We tested class-conditional split-conformal prediction sets for stereoelectroencephalography (SEEG) move-versus-rest decoders on archived features from two OpenNeuro datasets: cued key press (ds004473, 8 patients) and mouse movement (ds006253, 7 patients). A shrinkage linear discriminant analysis decoder trained on early trials was calibrated on middle trials and tested on later trials. At alpha 0.10, pooled mean class-averaged coverage was 0.886 against nominal coverage of 0.90 (exact sign-flip p = 0.163; Holm-adjusted p = 0.325); it exceeded 0.85 (Holm-adjusted p = 0.008). Coverage was 0.899 in ds006253, and 0.874 in ds004473; one ds004473 patient had coverage of 0.794. Time-ordered coverage was 0.017 below the random-split control (Holm-adjusted p = 0.290). Fifty-eight percent of test windows received a single-label set, and 22.3% of those decisions were wrong, above alpha (Holm-adjusted p = 0.0002). Thus set coverage did not bound the error rate among single-label decisions. The same pattern occurred at alpha 0.05 and 0.20. This was an offline analysis of one session per patient, two classes, one decoder type, and 15 patients; there was no closed-loop test. Dataset release versions for the archived feature inputs were not recorded.
split-conformal prediction · stereoelectroencephalography · brain-computer interface · coverage · single-label error · distribution shift · linear discriminant analysis
References
2- supportsIn 8 SEEG subjects, later-trial conformal coverage was near nominal, but single-label key-press decisions had 24% error sci:sha256:d02249a4772681202f4a0ed182ab560e24d98dce02fe045e976ba57746d057c4
The same method gives the same ds004473 values, and a second dataset (ds006253, 7 patients) shows coverage near nominal with single-label error above alpha.
- usesOffline tests in 30 SEEG patients found small mean AUC changes with 4-bit weights, but no test passed Holm correction sci:sha256:cf18311d6a5498c92be15629ce6ac9214a205968d362c405eecb225ebb849dd4
Input features for ds004473 and ds006253 are copied from this archive (SHA-256 in results/input_sha256.txt).
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesLearn-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)
Same question follow-up: uses its LDA, loader, time-ordered splits and conformal code; reproduces its 22.3% single-label error.
- usesAn onset-trained SEEG decoder detected later movement in 7 patients, but non-inferiority to a late-trained decoder was not shown
The feature files and the shrinkage LDA code were copied from this package.