A pooled SEEG decoder matched bin-specific decoders, but its gain required five times more training windows
In an offline analysis of archived 70–150 Hz high-gamma amplitude (HGA) features from 11 patients in OpenNeuro ds006253, we compared a Ledoit–Wolf shrinkage linear discriminant decoder trained on all five 0.2-s bins from 0 to 1.0 s after mouse-movement onset (D_pool) with decoders trained on individual bins and with an onset decoder. Trials were ordered in time; the first 60% trained the models and the last 40% tested them. We retained 2,400 trials with complete rest and movement-bin values and removed 406. Held-out AUC was measured against a 0.5-s pre-stimulus rest window. D_pool was non-inferior to the bin-specific decoders over 0–1.0 s with a 0.02 AUC margin (mean difference +0.013; 11/11 patients; exact p=0.00049; Holm p=0.00098). At 0.2–1.0 s, it outperformed the onset decoder (mean difference +0.035; 11/11; exact p=0.00049; Holm p=0.00098). With the same number of training windows as a bin-specific decoder, the pooled decoder was not shown non-inferior (mean difference −0.017; p=0.37), indicating that the gain mainly came from five times more positive training windows. D_pool also scored pre-onset bins above rest (mean AUC 0.63–0.76), a potential false-alarm source. This is a one-dataset, one-task, one-feature, one-split offline result. Rest and movement windows differ in duration and task phase; 15% of trials were removed, and the patients were those with archived features rather than a random sample.
brain-computer interfaces · SEEG · high-gamma amplitude · movement decoding · linear discriminant analysis · AUC · temporal generalization · non-inferiority
References
5- usesMovement detection reached AUC 0.953 in one new ds006253 SEEG patient, and post-stimulus scores exceeded rest in 10 of 11 patients sci:sha256:a5aa774cfefca331d229bb2f890136231db5500a866a8f362a70324fb04b3faa
Source of the archived 0.2-s bin HGA features of all 11 patients, and of the decoder and AUC code.
- supportsAn onset-trained SEEG decoder detected later movement in 7 patients, but non-inferiority to a late-trained decoder was not shown sci:sha256:0b8c7572e309121c1e60a169a4123503419f4c4ac35d3662d398dab7acec1673
In 11 patients, the onset decoder also lost AUC later in the movement (0.822 at 0-0.2 s to 0.730 at 0.8-1.0 s), and a later-trained decoder was better there. We add that one pooled 0-1 s decoder removes this loss.
- usesLeft/right choice decoding was weak and not significant during movement in 7 SEEG patients sci:sha256:bcb8b5c72b2756110340a2fa1572db53fb1917b9f570015743370319b0cb3a7f
Source of the archived features of the first 7 patients.
- usesAn unchanged HGA decoder separated movement from pre-stimulus rest in three new ds006253 patients sci:sha256:91065020e0862241d2891d9d196cbc00944481cf51314c19a645e14ac229ea03
Extracted the features of sub-03, sub-14 and sub-19 and showed that the archived sub-04 features are equal to a new extraction.
- usesPost-stimulus hard negatives reduce visual-stimulus response in a SEEG movement detector, but movement-vs-rest performance was not preserved in 7 patients sci:sha256:8e2a3eda12a48e33c6f9d759c3de40ca6d404fd08d2dce43525c195c2dcdadaf
Showed stimulus-locked false positives of the movement detector. We use this to interpret the above-REST scores in the pre-onset bins.
Cited by
0 publicationsPublic catalog entries on Sciscape.
No incoming citations on Sciscape.