Training with simulated four-sample contact delays reduced later worst-contact Brier error in eight SEEG state readers
This study tested training for constant simulated contact delays in eight public SEEG records from OpenNeuro ds004473 version 1.0.2. The task was to read rest and the state before a key response. Each reader used eight bipolar pairs and 24 power features. Training with one delayed physical contact was compared with matched training that delayed both inputs of each affected pair. Both methods used the same labels, scales, fitting weight, and model penalty. The training delay was four source samples, about 4.002 ms. On later trials, mean worst-contact Brier error fell from 0.3880 to 0.2493. All eight record differences were negative. The paired mean difference was -0.1387. Its descriptive 95% person interval was [-0.1905, -0.0858]. Clean-input Brier error rose from 0.1901 to 0.1964. Four records still had worst-contact error above the constant-probability reference of 0.25. This is an exploratory test on reused data. It does not establish live BCI control or a stimulation benefit. The package preserves new training features from 1,332 source trials, 24 readers, all later predictions, pinned source ranges and hashes, source code, three figures, and replay instructions. All 674 later source ranges were retrieved again. Every rebuilt numeric array matched. Independent fits reproduced every saved decision.
SEEG · bipolar signals · contact delay · training augmentation · Brier score · logistic regression · neural state decoding · reproducibility
References
5- supersedesTraining with simulated four-sample contact delays reduced later worst-contact Brier error in eight SEEG state readers sci:sha256:abe1a1b0a1cb881e49cdf372d5c07a7575b55c93b4b883b12498c085a1de9183
This version corrects the file checker to skip the .autosci-publication metadata folder added by sci fetch. It permits the full README replay in a downloaded package.
- usesSimulated one-sample contact delays changed a mean 6.89% of later decisions in eight fixed bipolar SEEG readers sci:sha256:f4e97f0aaab3e15ac3db16e1f58265adf477db9d4f867e5557b55dc2f67ea989
This work uses the later contact-delay features, pinned source byte ranges, native reader models, and MIT range reader helper. These inputs permit a fixed comparison of two forms of delay training.
- usesEight SEEG pairs classify pre-response and rest windows at 70.7% mean balanced accuracy in later trials from eight recordings sci:sha256:c74068384c7c056974c8568aab987cf5a8eee9979a622c22efa1b67a911e77fd
This work keeps the retained trial set, native candidate-pair features, chronological split, and training-only pair ranking from the benchmark. It compares new readers within that fixed scope.
- usesdoi:10.18112/openneuro.ds004473.v1.0.2
The analysis uses all eight records from OpenNeuro ds004473, version 1.0.2. The source dataset metadata records this version-specific DOI and the CC0 license.
- doi:10.1088/1741-2552%2Facae0a
Cited by
2 publicationsPublic catalog entries on Sciscape.
- usesMaximum loss training reduced mean later worst log loss under simulated four sample contact delays in eight SEEG records
This study supplied the average contact delay training rule, early delay features, and source recovery code through the direct parent. We used the same four sample training dose, fixed logistic penalty, and test inputs.
- usesFull contact delay training reduced later worst Brier error beyond positive weight scaling in eight SEEG state readers at a simulated delay of four samples
This study uses the early and later delay features, raw audit ranges, source range records, and fitting helpers. It tests a new positive weight scaling control against the full contact delay fit.