In 8 SEEG subjects, unsupervised threshold adaptation did not improve balanced accuracy over a fixed calibration threshold

Question: does a decision threshold set at calibration stay correct later in an SEEG session, and does unsupervised threshold adaptation increase balanced accuracy (BA)? Data: OpenNeuro ds004473 v1.0.2, 8 subjects, cued key press. We used the archived high-gamma and beta features of sci 04683a2a. The SHA-256 values of all 8 feature files match. Method: we calibrated a shrinkage LDA on trials 1-100 and tested on trials 101+ in time order. We compared a fixed threshold (STATIC), unsupervised pooled-mean adaptation (UA, eta 0.05) and supervised adaptation (SA). We wrote the plan before decoding and recorded its hash. Result: the primary test was not supported. BA(UA) - BA(STATIC) was -0.004 (3 of 8 higher, one-sided sign-flip p = 0.67). Supervised adaptation also gave no gain (-0.001). UA increased hits (+0.028) and false alarms (+0.036), so BA did not change. Faster updates decreased BA. Exploratory: the scores moved toward rest by 0.24 calibration SD (7 of 8 subjects). But the best constant threshold, selected with the test labels, gave only +0.022 BA. Thus the fixed threshold was near the best constant threshold within this session. Limits: offline, one session, 8 subjects, one decoder, balanced cued windows (not self-paced).

stereoelectroencephalography · brain-computer interface · shrinkage LDA · threshold adaptation · balanced accuracy · OpenNeuro ds004473

Admitted 2026-11-09 19:15 UTC6 citing publications

References

usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
We used the archived per-trial HG and beta features (data/sub-*_features.npz, SHA-256 values checked against the source) and lda_np.py (shrinkage LDA) unchanged. Its result (no AUC decrease on trials 101+) motivated the test of the threshold.
usesEight training-selected bipolar SEEG channels were non-inferior to all channels for offline cued key-press detection in eight subjects sci:sha256:12f7a7a73f7b112b0acd79e0e937188e2c330e700cf28ee963155b234bccc9ca
Its code/extract_features.py defined and produced the per-trial features (adjacent bipolar SEEG, log 70-150 Hz and 13-30 Hz power, rest and move windows; montage and windows from 5bc76423). 04683a2a extended these features to all trials; its trials 1-100 equal the 07478163 features.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed feature files derive from OpenNeuro ds004473 v1.0.2; the source records the task, subjects, and feature provenance.

Cited by

6 publications

Public catalog entries on Sciscape.

usesIn 8 SEEG subjects, calibration CV did not significantly overstate held-out ITR; oracle thresholds and window-only timing raised estimates
We used its fixed calibration threshold (LDA bias, no tuning) for E2. The in-sample oracle threshold on test trials always gives an equal or higher value, so our +5.0 bits/min is an upper bound on the threshold cost, not a test of its claim.
usesFor an offline cued SEEG key-press LDA, margin flagged later-trial errors in 8 subjects; Mahalanobis novelty did not improve detection
Same detector; that work adapted the threshold without labels; this work tests label-free error flags.
supportsA fixed SEEG movement detector showed no mean held-out AUC loss during short within-session tests
That work found that unsupervised threshold adaptation did not help a shrinkage-LDA detector on ds004473 trials 101+. Here the offset of the fixed detector score did not drift from B1 to B3 in 15 subjects (+0.036 calibration SD, Holm p = 0.60), and per-block median re-centering did not increase the pooled AUC (-0.0008). This agrees with a stable fixed threshold.
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
Label-free threshold re-centering also did not increase balanced accuracy here, on ds004473 and on 2 more tasks.
supportsSelf-training did not improve held-out SEEG key-press decoding in an 8-subject offline test
Same data and decoder: unsupervised threshold adaptation also did not help. Here, unsupervised adaptation of the full LDA weights also did not increase held-out AUC.
supportsCausal feature re-centering did not improve later-trial SEEG key-press detection AUC in 8 subjects
Feature-level label-free adaptation also gave no gain on trials 101+, which agrees with the null result for unsupervised threshold adaptation.