Detecting simulated SEEG channel failures limited AUC loss more than refitting in a 15-patient working-memory decoder
We evaluated simulated random bipolar-channel failure after calibration in an offline SEEG working-memory-load decoder for verbal Sternberg set sizes 8 versus 4. The analysis used archived five-band log-power features from session 1 of 15 patients in OpenNeuro ds004752 and shrinkage LDA with stratified five-fold cross-validation and three repeats. Mean AUC without failure was 0.685 (range 0.43–0.95). At 20% channel failure, setting failed features to the training mean reduced AUC by 0.008 (10/15 patients; exact two-sided sign-flip p = 0.034), while refitting on surviving channels reduced AUC by 0.001; refit exceeded imputation by 0.007 (13/15; Holm-adjusted p = 0.008). Undetected failure, modelled as z = +3 in all bands, reduced AUC by 0.042 (12/15; Holm-adjusted comparison p = 0.007). At 50% failure, losses were 0.028 with imputation, 0.014 with refit, and 0.086 when failure was undetected. For this decoder, simulated failure model, and dataset, detection mattered more than refitting. Results are limited by 20–42 trials per patient, offline random failures rather than shaft-level failures, one session, one feature set and decoder, three cross-validation repeats, and the assumed high-power value for undetected failure. The complete reported methods, results, code, feature hashes, provenance, and figure are included.
SEEG · working memory · channel failure · brain-computer interface · shrinkage LDA · AUC · feature imputation · decoder refitting
References
5- supportsIn 18 offline SEEG patients, refitting partly recovered AUC after 50% random channel loss; shaft loss caused greater decline sci:sha256:e8c2892080a762f7780aaae9df84f7018a351b41afb3397a55132fa5b63805fd
Refit on surviving channels after channel loss recovered AUC in word recall (ds004789); this work shows the same in working-memory load decoding.
- supportsIn 5 FR1 patients, undetected simulated channel failure reduced word-recall AUC more than detected failure sci:sha256:7af524d516c30f0c5c70c4318fa56f40b84a16314a5669cff18ae8275d2469f2
Same channel-failure question in another task (working-memory load, ds004752); the losses with imputation and with undetected failure have the same direction and similar size.
- usesSEEG error trials did not significantly lower working-memory load-decoder scores in 14 patients sci:sha256:0d140c83da68b1174870ae00264c22628b05be16993ed4c4e7b8009e3c788b92
Source of the archived ds004752 band-power features (input/feat).
- usesdoi:10.18112/openneuro.ds004752.v1.0.1
Source dataset for the archived session-1 SEEG features used in the analysis; metadata records exact version 1.0.1 and CC0.
- doi:10.7554/eLife.78677
Dataset paper for OpenNeuro ds004752, identified in the dataset's saved HowToAcknowledge metadata.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- supportsFailure of the highest-response shaft reduced AUC by 0.097 in an offline SEEG detector
For a different task (motor key press, ds004473), random loss of 20% of channels also gave a small AUC loss with mean imputation (0.017 vs its 0.008), a smaller loss with refit (0.012), and a larger loss when the failure was not detected (0.032 vs its 0.042). We add that whole-shaft failure of the critical shaft costs much more (0.097).
- supportsIn 18 FR1 patients, simulated SEEG contact failures caused small losses when detected and imputed, while undetected failures caused larger losses
Same pattern in a different task (FR1 word encoding): undetected failure costs much more than imputation, and refit gives a small gain over imputation.
- usesLoss of the calibration-ranked SEEG shaft reduced key-press detection after refitting in 7 of 8 subjects
Comparison: random channel failure after calibration gave small losses in a memory decoder; we show that random shaft loss is also small for key press, but the key shaft is not.