Diagonal LDA was not non-inferior to shrinkage LDA for offline SEEG movement detection in 15 patients

We compared diagonal linear discriminant analysis (DG), Ledoit-Wolf shrinkage LDA (SH), and nearest class mean (NM) for offline classification of movement versus pre-stimulus rest from 70–150 Hz high-gamma activity features in 15 patients with SEEG. The archived features came from OpenNeuro ds006253 via an earlier SCI contribution; the exact OpenNeuro release for these files is uncertain. A pre-registered, time-ordered split used the first 60% of trials for calibration and the last 40% for testing. DG was not shown non-inferior to SH within a 0.02 AUC margin: mean test AUC was 0.812 for DG and 0.833 for SH, with a mean difference of −0.021 (bootstrap 95% CI −0.036 to −0.008; exact one-sided sign-flip p = 0.58); DG was better in 3 of 15 patients. With 20 calibration trials, non-inferiority was also not shown (difference −0.013, CI −0.034 to +0.006; Holm-adjusted p = 0.75). NM was not shown non-inferior in either condition. Exploratory analyses found equal DG and SH AUC with 10 calibration trials (difference −0.001), a larger DG loss with more channels (Spearman rho −0.42), and about 88 times fewer multiply-adds and 29 times less stored data for a DG fit at these channel counts. These are offline, task-locked results from one clinical session per patient; operation counts are analytic, not hardware measurements.

intracranial EEG · SEEG · high-gamma activity · linear discriminant analysis · shrinkage LDA · decoder calibration · movement detection · non-inferiority · OpenNeuro ds006253