Random window splits gave 50.9% mean balanced accuracy for random trial labels in eight SEEG recordings with no shared samples
We tested 2,014 trials from eight public SEEG recordings in OpenNeuro ds004473 version 1.0.2. Each trial supplied four separate 500-sample windows before the response. The windows shared no source samples. We assigned a random binary label to each trial and used 100 label draws per recording. Random window splits gave 50.85 percent mean balanced accuracy. Whole trial splits gave 49.99 percent. The paired difference was 0.87 percentage points. Its 95 percent interval was 0.44 to 1.26 points. The exact sign-flip test gave p = 0.0156 under independent people and symmetric null differences. Thus, removing shared samples did not remove all test bias in this fixed pipeline. The package preserves source hashes, extracted features, fitted models, class decisions, code, and figures. This test measures evaluation bias. It does not measure the ability to decode movement intent.
stereoelectroencephalography · SEEG · brain-computer interfaces · data leakage · evaluation bias · random labels · trial splits · balanced accuracy
References
4- usesRandom window splits gave 57.3 percent mean balanced accuracy for random trial labels in eight SEEG recordings with four overlapping windows sci:sha256:9035d06d835ffd76c67ea3cf7af7a412474f615215115244b44e51a51ad7a757
This study uses the fixed trials, shared-sample features, source hashes, split rows, and random label protocol. It tests the open question with zero shared source samples.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This study uses source metadata, fixed trial rules, and EDF feature code through the direct parent. They permit an exact source range check.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
This is the exact OpenNeuro ds004473 version used for the eight SEEG recordings. The pinned dataset metadata identify this DOI and list the license as CC0.
- doi:10.1088/1741-2552%2Facae0a
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesMixed windows gave 9.1 percentage points higher mean balanced accuracy than separate records for a four-pair seizure reader in 28 SEEG subjects
Its random-label test with disjoint SEEG windows informed the rule that retained windows must share no source samples. This study measures a seizure-state validation gap.