Pair selection before blocked validation increased mean balanced task accuracy by 2.4 percentage points in eight SEEG recordings
We tested eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. Each decoder used eight bipolar pairs and three power bands. Choosing pairs with all calibration labels before blocked validation gave 71.99% mean balanced accuracy. Choosing pairs with training labels inside each fold gave 69.55%. The paired change was 2.44 percentage points, with a 95% person-bootstrap interval of 0.90 to 3.76 points. With random within-trial label swaps, the two scores were 58.64% and 49.94%. A two-rest-window control gave 58.05% and 48.45%. The package preserves input features, exact source records, models, predictions, figures, and tests for label isolation. This result applies to offline task-state classification in one session per person. It does not test continuous intent or brain stimulation.
stereoelectroencephalography · brain-computer interface · feature selection · validation leakage · balanced accuracy · blocked cross-validation
References
4- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
We used all eight archived feature sets, trial records, source metadata, and pinned raw retrieval records. We adapted its raw feature and audit code. The final later-trial model states reproduce its eight-pair task decoder. We introduced the before-validation pair-selection path for the new comparison.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses the source recordings and metadata from OpenNeuro ds004473 v1.0.2.
- usesdoi:10.1088/1741-2552%2Facae0a
This paper is cited as the study describing the movement-related oscillation dataset used here.
- supportsdoi:10.1038/nn.2303
Cited in the publication as prior description of circular feature selection.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesAdding alpha and beta power to high-gamma power at the same eight pairs did not show a clear mean error reduction in later trials from eight SEEG recordings
Its pair-selection test informed the rule to choose pairs within each training fold. We apply that rule in expanding-time folds for model and band choice.
- usesAdding alpha and beta power to high-gamma power at the same eight pairs did not show a clear mean error reduction in later trials from eight SEEG recordings
Its pair-selection test informed the rule to choose pairs within each training fold. We apply that rule in expanding-time folds for model and band choice.