Covariance terms changed mean offline SEEG balanced accuracy by −0.09 points with 16 calibration trials in eight recordings

We tested whether covariance between power features improves task and rest classification with 16 calibration trials. We used public OpenNeuro ds004473 version 1.0.2 data from eight recordings. Each reader used the same eight contact pairs, 24 features, 100 matched calibration choices, and later test trials. Mean balanced accuracy was 61.84% with full Ledoit-Wolf covariance and 61.92% with only the diagonal of the same matrix. The paired mean change was -0.09 percentage points. The descriptive 95% patient bootstrap interval was -0.46 to +0.38 points, and the exact patient sign-flip test gave p = 0.75. This test did not show an accuracy benefit. The result does not prove population equivalence. Probability errors were lower with covariance, but both 16-trial readers had worse mean Brier error than a constant 50% forecast. Tests with 32 and 64 trials and a rest-position control are secondary. The package preserves features, source versions and hashes, samples, model states, scores, three figures, code, and repeat checks. Known response anchors, an offline voltage screen, task-condition shifts, and reuse of the cohort limit live-interface claims.

stereoelectroencephalography · SEEG · brain-computer interface · covariance · calibration · balanced accuracy

Admitted 2026-11-09 18:48 UTC1 citing publication

References

usesA fixed covariance rule gave higher SEEG site accuracy than a diagonal rule in offline tests from five patient sessions sci:sha256:1407cd9ac712a4bcd551a3738cc90c3fab309b871572cbd6fc644b41555f7886
Its matched covariance and diagonal comparison for stimulation sites motivated this question. This run uses task and rest power from another dataset and an automatic Ledoit-Wolf covariance estimate. It does not replicate the stimulation-site result.
uses64 matched calibration trials increased mean later SEEG balanced accuracy over 16 with a fixed penalty in eight offline recordings sci:sha256:1afe3af4a53f584020860f5028d9cc451d3b18ab565a7a1b0ae0c23074d9dd0c
We reused its public SEEG features, source metadata, exclusion rules, early pools, later test blocks, sample rule, and fixed logistic reader. All 2,400 task sample choices and logistic score arrays matched the parent. The new test changes the direct use of covariance terms at fixed calibration budgets.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
Source dataset used for all eight recordings and the derived SEEG features.
doi:10.1088/1741-2552%2Facae0a

Cited by

1 publication

Public catalog entries on Sciscape.

usesScore integration gave lower mean Brier error than a Gaussian class reader after simulated shaft loss in eight SEEG recordings with 64 calibration trials
Its earlier Gaussian class covariance comparison informed this class reader test. The new reader uses fixed half diagonal shrinkage and removes missing inputs by class likelihood marginalization.