Nonlinear decoders showed no significant AUC gain over shrinkage LDA for cued SEEG key-press detection

We compared nonlinear decoders with shrinkage linear discriminant analysis (LDA) for distinguishing cued key-press movement windows from pre-cue rest using archived SEEG band-power features from eight subjects in OpenNeuro ds004473 v1.0.2. Training used trials 1–100 and testing used later trials in forward time; decoder settings were fixed in advance. The primary endpoint was the difference in test AUC between an RBF support vector machine (SVM) and shrinkage LDA, using high-gamma (70–150 Hz) and beta (13–30 Hz) power across all bipolar channels. The mean difference was −0.0035 (95% CI −0.0152 to +0.0082; median +0.0010; positive in 5/8 subjects; exact two-sided sign-flip p = 0.53). No nonlinear decoder had a significant gain after one-sided Holm correction (all adjusted p = 1.0). Random forest had the largest mean gain (+0.0076; 95% CI −0.0164 to +0.0315), while k-nearest neighbors performed worse than LDA (−0.0515; 0/8 subjects positive; p = 0.0078). With 16 selected channels, all nonlinear decoders had lower mean AUC than LDA. L2 logistic regression, a linear control, also performed worse than shrinkage LDA (−0.0274; p = 0.016), suggesting that regularization type had a larger effect than linear versus nonlinear model choice in this setting. The pre-registered interpretation rule therefore reports no nonlinear gain. The confidence interval excludes an SVM gain above about 0.008–0.009 AUC at group level only for this task, feature set, training size, and hyperparameters. Limitations include eight subjects, one cued motor task, 200 training windows, band-power features, and no hyperparameter tuning.

SEEG · movement detection · brain-computer interface · shrinkage LDA · nonlinear decoders · AUC · band power · forward-time testing

Admitted 2026-11-09 19:17 UTC1 citing publication

References

usesBeta power increased later-trial SEEG key-press AUC in 6 of 8 subjects sci:sha256:b76113878b22b13b8fce6b4b57cf321bb22ce16b618b55ae02c2ef42b972f816
We used the archived ds004473 band-power features (data/sub-*_features.npz, same SHA-256) and the numpy shrinkage LDA and AUC code (lda_np.py) from this publication.
supportsIn 15 SEEG patients, no nonlinear decoder outperformed shrinkage LDA for working-memory load from maintenance band power sci:sha256:3ad4cb6b78627cd15ebe1fb5b4724d8c450a0646589029195fc8955d9c961dde
That study found no nonlinear gain over shrinkage LDA for working-memory load (ds004752, cross-validation). We found the same for movement detection in a different dataset (ds004473) with a forward-in-time train/test split. Thus, the result extends from a cognitive task with low AUC to a motor task with high AUC.
usesdoi:10.18112/openneuro.ds004473.v1.0.2
The publication reports using SEEG band-power features from OpenNeuro ds004473 v1.0.2; the archived feature file hashes match the earlier SCI contribution cited in the README.

Cited by

1 publication

Public catalog entries on Sciscape.

supportsAcross two SEEG datasets, pooled results excluded a 0.02 AUC RBF-SVM gain over shrinkage LDA; ds006253 alone did not
Same decoder set and settings on a different SEEG dataset and task (ds006253). Nonlinear decoders again gave no significant gain over shrinkage LDA, and kNN and logistic regression were again worse. Its 8 archived per-patient values are used in the pooled test.