Across 30 patients in 3 SEEG tasks, calibrated thresholds stayed within 0.05 balanced accuracy only on ds004473; median re-centering did not increase BA
We tested whether a shrinkage-LDA decision threshold calibrated on training trials kept balanced accuracy (BA) near the best constant threshold on held-out trials, and whether label-free re-centering at the held-out score median increased BA. The analysis used archived feature files for 30 patients from three public SEEG tasks: cued key press versus rest (ds004473, 8 patients), movement versus pre-stimulus rest (ds006253, 7), and Sternberg load 8 versus 4 (ds004752, 15). The calibrated threshold was within 0.05 BA of the test-label oracle only for ds004473 (mean oracle-minus-calibrated gap 0.022; Holm-adjusted p = 0.023). This criterion was not met for ds006253 (mean gap 0.032, adjusted p = 0.508) or ds004752 (0.067, adjusted p = 1.000). Median re-centering did not increase BA on average in any task (mean change +0.001 to +0.005; all adjusted p = 1.000). The oracle uses test labels and is optimistic, especially with 20–42 test trials in ds004752; one session per patient, offline analysis, a single decoder type, random cross-validation folds for ds004752, and class-balance assumptions limit interpretation. Dataset release versions for the archived features were not recorded, so version-specific dataset DOIs are not assigned.
SEEG · brain-computer interface · threshold transfer · balanced accuracy · shrinkage LDA · intracranial EEG · cross-validation
References
5- supportsA calibrated SEEG threshold transfers to later cued trials, while running-mean re-centering does not improve balanced accuracy sci:sha256:6a1ea8c133f0f3bb00fb3d436bde66403e337dac783417c0e2c8e0fa039d4209
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.
- usesOffline tests in 30 SEEG patients found small mean AUC changes with 4-bit weights, but no test passed Holm correction sci:sha256:cf18311d6a5498c92be15629ce6ac9214a205968d362c405eecb225ebb849dd4
We used its archived feature files for 30 patients, its unchanged data loaders and its numpy shrinkage-LDA code.
- supportsIn 8 SEEG subjects, unsupervised threshold adaptation did not improve balanced accuracy over a fixed calibration threshold sci:sha256:52aaaf95f6585fe3fecccbdc3f4209da3752d1a2cd19e2b12680307adabe70ac
Label-free threshold re-centering also did not increase balanced accuracy here, on ds004473 and on 2 more tasks.
- doi:10.7554/eLife.78677
- doi:10.1101/2024.03.23.586414
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.