Loss training increased mean balanced accuracy by 0.34 points under simulated loss of one contact with mean replacement in eight SEEG decoders
We tested simulated contact loss in eight fixed SEEG task-state decoders. The inputs came from 2014 selected trials in OpenNeuro ds004473 version 1.0.2. Each decoder used eight bipolar pairs. Early trials set the pairs, scaling, and model. Later trials formed the test. A known missing contact removed all pairs that used it. Missing features were replaced with their training means. Standard mean balanced accuracy was 70.69% without loss and 70.00% across equally weighted single-contact losses. The planned average loss was 0.68 points, with a 95% participant interval from -0.29 to 1.44 and p = 0.203. The mean minimum accuracy over contacts was 65.51%. Loss training gave a planned secondary gain of 0.34 points, with an interval from 0.12 to 0.56 and Holm-adjusted p = 0.0469. The minimum-accuracy gain was uncertain. The package supplies code, features, all failure predictions, physical contact maps, source checks, and two figures. This is an offline simulation with perfect failure detection. It does not establish online control, hardware reliability, or transfer to another day.
stereoelectroencephalography · brain-computer interfaces · contact loss · balanced accuracy · robustness
References
4- 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 use the preserved SEEG features, source records, and extraction code. We reproduce the standard eight-pair model before the new contact loss and loss training tests.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analyzed features and physical contact identities came from OpenNeuro ds004473 version 1.0.2.
- supportsdoi:10.1088/1741-2552%2Facae0a
The cited source study describes the SEEG recordings and task used in the dataset.
- doi:10.1038/ncomms13749
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesRidge weights from training data gave mixed error changes under known input loss in eight SEEG recordings
Its known contact-loss test and mean replacement assumption informed the comparator. This test uses a different fixed ridge model and version 1.0.1 features.
- supersedesLoss training raised mean accuracy by 0.34 points under simulated contact loss in eight SEEG decoders
This version corrects the package hash checker to exclude AutoSci download metadata. The analysis code, data, results, and figures are unchanged.