Largest-loss fitting raised correct decisions across all simulated contact losses with mean replacement in later windows from eight SEEG records. Intact accuracy fell.
We tested whether training for the largest loss improves SEEG decoding when one contact fails. We used public OpenNeuro ds004473 version 1.0.2 and preserved features from eight people. Each model used eight selected bipolar pairs and 24 band powers. Selection and fitting used earlier trials. The final third of trials formed the test set. A known missing contact removed every pair that used it. Training means replaced the removed band powers. Largest-loss fitting raised correct decisions across the intact case and every loss case from 50.39% to 55.09%. The paired increase was 4.70 percentage points (person-bootstrap 95% interval: 1.13 to 8.61). The exact two-sided sign-flip p value was 0.0625. However, intact balanced accuracy fell from 71.87% to 69.25%. Two models always predicted rest. A later descriptive audit found lower recall for the weaker class, from 47.35% to 35.32%. Seven control models also returned constant classes. Stable decisions can therefore include stable errors and loss of one class. We provide exact linear programs, score bounds, all predictions, raw checks, figures, and reproduction code. This reused cohort gives exploratory evidence under a simulated failure model. It does not show live interface benefit.
SEEG · robust decoding · contact loss · feature deletion · linear programming · brain-computer interfaces
References
7- usesLoss training raised mean accuracy by 0.34 points under simulated contact loss in eight SEEG decoders sci:sha256:884a92921286ae95392544128af6a4f343937c3b760e03bce164eec4599853d4
Its mean-replacement contact-loss training test informed our average-loss comparison. We changed the loss and penalty for a matched comparison; we did not directly reproduce its numerical effect.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
We used its eight-person SEEG feature arrays, task labels, chronological split, pair ranking, and scaling. Its raw retrieval code and audit ranges enabled source checks and reproduction.
- usesTraining for ±6 dB gain on one shaft raised correct classes with a guarantee in eight SEEG records under a linear gain model sci:sha256:5b0d343335c2da00c225c25a050de434c99e47b5020cc51baebb2b35ef408f94
Its earlier worst-case shaft-gain training test informed the novelty assessment. Our test uses physical contact loss, mean replacement, and a different split and objective.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed features and source records are from OpenNeuro ds004473 version 1.0.2.
- doi:10.1088/1741-2552%2Facae0a
- doi:10.1145/1143844.1143889
- doi:10.1038/ncomms13749
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesTraining for fixed contact loss avoided constant models in eight SEEG records under simulated loss with mean replacement. The minimum accuracy gain was uncertain.
We used its checked SEEG features, physical contact masks, training inputs, chronological split, largest-window loss code, and saved parent models. We changed only the aggregation rule for the new fixed-contact fit. All 96 repeated parent models matched their coefficients, objectives, intercepts, and test scores.
- usesWhole-shaft training showed no clear gain over matched pair omission in lowest balanced accuracy after known shaft loss with mean replacement on later trials from eight SEEG recordings
The cited class-recall audit motivates constant-class checks and reports of both class recalls.