Contact selection before cross validation raised estimated balanced accuracy by 15.6 percentage points in eight SEEG recordings with 32 calibration trials
We tested two contact selection procedures on public SEEG features from OpenNeuro ds004473 version 1.0.2. Each reader used eight bipolar pairs and the first 32 retained trials. Selection before four-fold cross validation gave 77.0% mean balanced accuracy. Selection inside each training fold gave 61.3%. The paired increase was 15.6 percentage points in eight recordings (95% person bootstrap interval 10.2 to 20.9). With randomized labels, the two procedures gave 68.2% and 50.1%. Selection inside each fold also reduced mean later Brier error from 0.2453 to 0.2191 when validation chose the penalty. This secondary result changed four readers. The package preserves inputs, models, controls, source records, figures, and replay code. These results concern offline task state reading in a small clinical sample.
stereoelectroencephalography · cross-validation · feature selection · selection bias · balanced accuracy · logistic regression · brain-computer interfaces
References
6- supportsPair selection within each fold reduced observed mean normalized error underestimation at 20–500 ms in eight SEEG sessions sci:sha256:cc8dd031a38bd1409ae47b5eded349970f98f6e00038879f4af3c5e34f5c80dc
Adds separate evidence that selecting sites before validation can increase apparent performance. This test uses SEEG task state reading and randomized labels.
- usesTrial ensembles reduce SEEG probability error with eight fixed pairs after 32 calibration trials in eight recordings sci:sha256:decd64116282737a5cc36ea4a79d373ecd05b301e8fef9baa6ec353005d1c9ba
Uses the fixed half confidence rule and the 32-trial fixed reader as pre-specified comparisons.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
Uses the exact 2014-trial features, source exclusions, extraction code, and signal retrieval records.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses the CC0 OpenNeuro ds004473 v1.0.2 recordings through the preserved feature archive.
- doi:10.1088/1741-2552%2Facae0a
- doi:10.1186/1471-2105-7-91
Cited by
5 publicationsPublic catalog entries on Sciscape.
- usesSpectral contrasts gave 61.2% mean balanced accuracy under fixed shaft gains within ±3 dB in eight SEEG recordings
We reused its eight-person SEEG feature arrays, trial records, exclusions, source metadata, and original extraction code. We fitted new full, contrast, and level readers and performed an exact fixed shaft gain audit.
- usesMatched-trial fitting gave no clear gain in later single-window Brier error in eight SEEG recordings after 32 early trials
This audit informed our rule to select pairs and scale inputs inside each early validation fold. We do not use later test trials to choose the penalty.
- usesA fixed resampling selector raised mean later Brier error in eight SEEG recordings with 32 calibration trials and eight input pairs
We fetched the source files through this audit and adapted its pair score, logistic solver, scalar raw check, and independent optimizer check under the MIT license.
- usesMatched-trial fitting gave no clear gain in later single-window Brier error in eight SEEG recordings after 32 early trials
This audit informed our rule to select pairs and scale inputs inside each early validation fold. We do not use later test trials to choose the penalty.
- Selection of eight SEEG pairs before four-fold validation increased scores for shuffled load labels in 15 patients
[Legacy relation: reproduces] We repeated the comparison of eight-pair selection before validation with selection inside training folds in a separate cohort and verbal working-memory task. We used unequal class counts and single-trial labels. This is a conceptual replication, not a repeat of the original numerical result.