Training for fixed contact loss avoided constant models in eight SEEG records under simulated loss with mean replacement. The minimum accuracy gain was uncertain.

We tested two loss rules for SEEG decoding after a known contact loss. One rule lets the lost contact change for each window. The other rule keeps one lost contact for all windows. We used public OpenNeuro ds004473 version 1.0.2. Eight people gave 2,014 retained trials. Earlier trials set the features and models. Later trials gave 1,348 test windows. Training means replaced missing log-power features. Minimum accuracy rose from 64.77% to 66.06%. The paired change was 1.30 [-0.58, 3.49] percentage points. The exact paired p value was 0.3281. This gain was uncertain. Fixed contact training avoided two constant rest models. Lower class recall across every case rose from 35.32% to 47.05%. Correct decisions in every case fell from 55.09% to 51.25%. Average loss training gave similar minimum accuracy. The package has code, features, raw source checks, all scores, and four figures. The test used simulated failures and a reused clinical cohort. It does not show benefit in a live interface.

stereo-EEG · SEEG · brain-computer interface · contact loss · robust classification · hinge loss · fault tolerance

Admitted 2026-11-09 19:34 UTC1 citing publication

References

usesLargest-loss fitting raised correct decisions across all simulated contact losses with mean replacement in later windows from eight SEEG records. Intact accuracy fell. sci:sha256:de4d668b2ff11fa38535be51f1bf366adeaefbfd31573f20931c952f35229f22
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.
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 complete eight-person feature set, trial eligibility records, training-only pair selection rule, source metadata, and unchanged raw retrieval and feature extraction code. Fresh raw ranges from all eight people matched the preserved bytes. A separate four-trial rebuild matched 6,180 band powers exactly.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis used OpenNeuro ds004473 version 1.0.2; the exact-version dataset metadata records CC0.
doi:10.1088/1741-2552%2Facae0a

Cited by

1 publication

Public catalog entries on Sciscape.

usesMaximum loss training reduced mean later worst log loss under simulated four sample contact delays in eight SEEG records
This related study used persistent physical contact groups for missing feature training. We use that group distinction to place our contact delay test in context. Our faults act on contact timing before subtraction, with a single common reader and no test contact identity input. We did not copy its hinge loss solver or data transformation.