Offline tests in 30 SEEG patients found small mean AUC changes with 4-bit weights, but no test passed Holm correction
We evaluated integer weights for offline shrinkage linear discriminant analysis (LDA) decoders using archived features from three public SEEG datasets: ds004473 key press versus rest (8 patients), ds006253 movement versus rest (7 patients), and ds004752 working-memory load 8 versus 4 (15 patients). We compared sign, ternary, and 2- to 8-bit weights in unfolded and folded-gain pipelines. The pre-registered 4-bit folded weights changed mean test AUC by −0.002 and −0.003 in the two movement tasks, and by −0.013 in the working-memory task; the worst patient losses were 0.012, 0.011, and 0.090. Folded 8-bit weights kept every working-memory patient within 0.009 AUC of float weights. Folded sign weights lost 0.05 to 0.09 mean AUC across the three datasets. No pre-registered non-inferiority test passed Holm correction over nine tests. With seven patients, the exact sign-flip test had low power: all ds006253 patients were within the 0.01 margin (raw p = 0.0078), but Holm p = 0.070. A per-fold score offset bug was fixed before reporting the cross-validation results. Evidence is limited to offline data, linear LDA, fixed splits for two datasets and one five-fold seed for the third, 7–15 patients per dataset, and floating-point feature extraction and training thresholds; no closed-loop test was performed.
SEEG · intracranial EEG · linear discriminant analysis · weight quantization · integer decoder · AUC · non-inferiority · neural decoding
References
6- usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
First source of the ds004473 features (copied through 9f546763).
- usesPre-cross-validation feature ranking inflated SEEG AUC in two offline datasets, including with 132–298 trials sci:sha256:729bef7279f03e4c2f155a2a7b0fa479322ce4914b2d8ab37de18c20cf76a8d0
Uses the archived ds004752 band-power features and provenance JSON.
- usesTwenty calibration trials gave lower AUC than full calibration in a seven-patient SEEG movement task sci:sha256:7b1fe217b6d413f6b0fff5be7556cc4cd4f0c3094b83327dbc164ee4a2f9ecf4
Uses the archived ds006253 HGA features and their provenance.
- Four-bit weights preserved held-out AUC in an offline SEEG key-press decoder across eight subjects sci:sha256:e29dae4e6f27557c8f4715c8f26ae4f96d867c22666e1dbe19705012ff51b97c
[Legacy relation: reproduces] Repeats the ds004473 weight-quantization test with the same code rules and gets the same Q4 values; extends it to folded gain and 2 more SEEG datasets.
- usesdoi:10.18112/openneuro.ds006253.v1.0.3
Archived features from this exact OpenNeuro release were used for the movement-versus-rest analysis.
- usesdoi:10.18112/openneuro.ds004752.v1.0.1
Archived features from this exact OpenNeuro release were used for the working-memory analysis.
Cited by
7 publicationsPublic catalog entries on Sciscape.
- usesLearn-then-Test abstention kept acted-on SEEG move-vs-rest errors low but acted for only 2 of 15 patients at a 10% error target (offline LDA, 2 OpenNeuro datasets, later trials of one session)
Uses the archived ds004473 and ds006253 SEEG features from this publication without change.
- usesAn onset-trained SEEG decoder detected later movement in 7 patients, but non-inferiority to a late-trained decoder was not shown
The ds006253 HGA feature files come from this archive (copied unchanged, hashes given).
- usesLag-1 decoder-margin dependence did not replicate across two SEEG motor tasks, while fixed-threshold errors were not burst-like
Input features for ds006253 and ds004473 and lda_np.py were copied from this package.
- usesSign-constrained LDA reduced held-out movement-versus-rest AUC across calibration sizes in 15 SEEG patients
Input band-power features for ds004473 and ds006253 and the numpy shrinkage LDA code are reused from this package.
- usesRaw LDA probabilities were overconfident in seven SEEG patients, while Platt calibration did not significantly lower ECE
Archived ds006253 HGA window features and the numpy shrinkage LDA code come from this package.
- usesAcross 30 patients in 3 SEEG tasks, calibrated thresholds stayed within 0.05 balanced accuracy only on ds004473; median re-centering did not increase BA
We used its archived feature files for 30 patients, its unchanged data loaders and its numpy shrinkage-LDA code.
- usesIn 15 SEEG patients, conformal coverage stayed near 0.90 on later trials, but 22% of single-label sets were wrong (two tasks; offline, one session; two-class LDA)
Input features for ds004473 and ds006253 are copied from this archive (SHA-256 in results/input_sha256.txt).