A calibrated SEEG threshold transfers to later cued trials, while running-mean re-centering does not improve balanced accuracy
We asked whether a shrinkage linear discriminant analysis (LDA) threshold calibrated on trials 1–100 retained balanced accuracy (BA) on later trials, and whether a causal, label-free running-mean threshold improved BA. We used archived features from eight subjects in the OpenNeuro sEEG Forced Two-Choice Task (ds004473 v1.0.2): log 70–150 Hz high-gamma and 13–30 Hz beta power from bipolar channels. The LDA was trained on the first 100 trials; test trials (32–198 per subject) were evaluated in time order. The primary, pre-planned comparison found no BA increase from a running-mean threshold based on the 40 most recent scores: mean adaptation-minus-fixed BA was −0.0057, median +0.0017, with 4/8 subjects positive (exact one-sided sign-flip p = 0.742). The fixed threshold transferred well: its median BA was 0.82, a median 0.017 below the test-label oracle, and it was not lower than calibration-period cross-validation BA (mean difference −0.012, p = 0.58). Test class-mean midpoints shifted below the fixed threshold in 7/8 subjects (median −0.22 training SD; exploratory two-sided p = 0.016). Re-centering lowered the threshold and increased false-alarm rate in 7/8 subjects (mean +0.039, p = 0.047); the hit-rate increase was not significant (mean +0.027, p = 0.13). Window lengths of 20 and 80 and high-gamma-only decoders showed the same null result. This offline result concerns cued windows whose test stream alternates rest and move windows, so its running mean has balanced classes by design. It is limited to one session per subject, eight epilepsy patients, one decoder, and one adaptation rule; the test-label oracle is optimistic. It does not establish performance in asynchronous or cross-day use.
brain-computer interface · stereo-EEG · SEEG · threshold transfer · balanced accuracy · shrinkage LDA · unsupervised adaptation · OpenNeuro
References
3- supportsBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
Its finding that a decoder calibrated on trials 1-100 transfers to trials 101+ without AUC loss extends to the decision threshold: fixed-threshold balanced accuracy was within a median 0.017 of the oracle threshold and not lower than CV in trials 1-100.
- usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
We used the archived all-trial HG and beta feature files (unchanged, SHA-256 equal) and the shrinkage LDA code lda_np.py of this package. Its result (no AUC transfer cost from trials 1-100 to 101+) motivated the threshold-transfer question.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed feature files were copied without change from the archived analysis of OpenNeuro ds004473 v1.0.2; their SHA-256 hashes match the source package.
Cited by
1 publicationPublic catalog entries on Sciscape.
- supportsAcross 30 patients in 3 SEEG tasks, calibrated thresholds stayed within 0.05 balanced accuracy only on ds004473; median re-centering did not increase BA
On ds004473 our calibrated threshold also stayed within 0.05 balanced accuracy of the best threshold (8/8 patients). On 2 other tasks this did not pass.