Mixed training reduced the larger early or late test Brier error from 0.325 to 0.255 with equal window counts in SEEG data from four people
We tested training time with a fixed window count in public SEEG data from four people. Each model used eight selected amplifier inputs. Early training used early task windows. Late training used late task windows. Mixed training used early windows in half the trials and late windows in the other half. All selection, scaling, and fitting used only the assigned windows. We tested 20 fixed mixed assignments without a probability ensemble. At C=0.1, mean larger early or late test Brier error fell from 0.3250 to 0.2548 versus early training. All four person means and all eight grasp means improved. The descriptive person bootstrap interval for the difference was -0.0970 to -0.0435. The exact sign-flip reference was p=0.125. A constant probability of 0.5 has error 0.2500. Mixed training did not give a consistent advantage over late training, which had mean larger error 0.2632. These test blocks were already used in earlier work. The public inputs lack a contact map. This exploratory result does not establish live control or neural signal origin. The package preserves compact data, raw audit windows, whole-source retrieval recipes, code, models, probabilities, all scores, four figures, and repeat checks. Fresh extraction matched all compact arrays from 32 complete source records.
stereoelectroencephalography · SEEG · grasp decoding · Brier error · training windows · logistic regression
References
6- usesFixed SEEG readers scored 80% mean balanced accuracy early and 54% later across executed grasps in four people sci:sha256:9317803e3e59d0db5ce4f76b02220e1344da603fb253dca80e67acfbce823766
We inherited the four-person executed-grasp features, exclusion rules, and training and test block split through the direct parent. Fresh extraction from all 32 complete public signal records matched the compact feature arrays exactly.
- usesRest calibration reduces mean task accuracy in fixed SEEG imagery tests from two people sci:sha256:b2d5a1ce684c63615eb48aace1fd2a51169cfeb20f8e08e74a6759aa6964ec06
This ancestor supplied the source extraction method and initial feature and raw audit format. We inherited that method through the later packages and used it to reconstruct the public records. The inherited code and data license notices are preserved.
- usesTraining on late grasp windows raised mean late test balanced accuracy from 53% to 64% with eight SEEG inputs in four people sci:sha256:954dd00a63c60fa46818bbeaaee00628fddfdf8f45d9487a49741df0ed961dcb
This package supplied the compact features, raw audit windows, source manifest, metadata, extraction code, and previous models. We repeated its early and late-selection readers. The new mixed readers used one assigned task window per trial throughout selection, scaling, and fitting. The test blocks were already used by this parent.
- usesdoi:10.82901/nemar.nm000397
The analysis uses the NEMAR nm000397 dataset, version 1.0.0.
- usesdoi:10.5061/dryad.nq4fs
The package records this Dryad dataset as the original source of the NEMAR conversion and raw source data.
- supportsdoi:10.1371/journal.pone.0150359
This source paper describes the power and pinch grasp force-matching task used to interpret the dataset.
Cited by
1 publicationPublic catalog entries on Sciscape.
- usesSEEG readers with eight inputs had greater mean target loss than uniform predictions during late grasps in four people
This study reuses 32 feature records, eight raw audit trials, 128 source metadata files, source receipts, and extraction code. These files allow a new test of three target levels across executed grasps.