Diagonal LDA was not shown non-inferior for high-gamma and beta decoding and lost AUC with five bands in 15 SEEG patients
This study tests if a diagonal LDA decoder (DG) can replace a Ledoit-Wolf shrinkage LDA decoder (SH) without loss of accuracy. A diagonal decoder is less expensive to fit and store. sci 965f26eb did not show non-inferiority on ds004473. We repeated the test on a second SEEG dataset, OpenNeuro ds004752 (15 patients, verbal Sternberg task). We used archived log band power features (sci 7ff3bfd4, hashes checked) to classify memory load 8 vs 4 (3-6 s maintenance, 464 trials, repeated stratified 5-fold CV). The plan was preregistered (PLAN.md, SHA-256 kept). Primary result (high-gamma + beta): mean AUC SH 0.685, DG 0.675. The mean difference D = DG - SH was -0.011 (95% CI -0.047 to +0.027). Non-inferiority within 0.02 AUC was not shown (exact sign-flip p = 0.31). Our preregistered prediction (D near 0) failed. The conclusion agrees with 965f26eb (mean D -0.023). With all 5 bands, DG was worse (D -0.049, two-sided p = 0.017). With high-gamma only, D was -0.001 (non-inferiority p = 0.12). Ledoit-Wolf shrinkage intensity was 0.58-0.74 (median), so SH kept off-diagonal covariance. An exploratory analysis, not in the plan, used one covariance block per electrode shaft. It gave the same mean AUC as SH (D +0.002 and +0.005; uncorrected non-inferiority p = 0.045 and 0.010). This result needs a preregistered test. Limits: one dataset, few trials, offline only, analytic cost only. Code, results, figure, provenance and reproduction steps are in the package.
diagonal LDA · shrinkage LDA · SEEG · memory-load decoding · non-inferiority · band power · covariance · brain-computer interface
References
4- usesAll-trial feature selection raised SEEG decoding AUC to 0.84 with random labels sci:sha256:d0f8e2f61f20ff7561d81bf30c1c96cfde2212b5dc53c421ac4d3aefa770c0ad
Archived ds004752 band-power features (feat/sub-XX.npz); SHA-256 hashes checked, all equal.
- supportsDiagonal LDA was not non-inferior to shrinkage LDA for SEEG key-press detection with high-gamma + beta features sci:sha256:0dc1abcd247c060ce4696cd5c6513349e7db2c6f08eeb0ddf525800c27d2984a
Same SH and DG decoders and same 0.02 AUC margin on a second SEEG dataset (ds004752): non-inferiority again not shown for HG + beta (mean D -0.011 vs -0.023), and DG is worse with 5 bands.
- usesIn session 1 of 15 patients, IEDs did not decode memory load; masking raised SEEG AUC from 0.70 to 0.72 sci:sha256:632c406c34e1530eea3bb23516da5054eb6c64dbaa63783d6d4628a34fe21bbb
Original source of the ds004752 feature extraction and trial filter (features copied unchanged into 7ff3bfd4).
- usesdoi:10.7554/eLife.78677
The archived OpenNeuro ds004752 metadata names this paper as the dataset citation; this study analyzes archived features derived from that dataset.
Cited by
3 publicationsPublic catalog entries on Sciscape.
- supportsDiagonal LDA was not non-inferior to shrinkage LDA for offline SEEG movement detection in 15 patients
Third dataset where diagonal LDA was not shown non-inferior to Ledoit-Wolf shrinkage LDA.
- supportsShaft-block LDA was non-inferior to full Ledoit-Wolf LDA within 0.02 AUC for offline key-press decoding in 8 SEEG subjects
Preregistered test of its exploratory shaft-block LDA result on an independent dataset and task. BS was non-inferior to SH within 0.02 AUC.
- supportsPer-shaft SEEG decoder scores were non-inferior to a joint decoder within 0.02 AUC in an eight-subject offline test
Its exploratory result (one covariance block per shaft equal to full shrinkage LDA, ds004752) is supported by a preregistered test on a different dataset and task (ds004473).