In 15 SEEG patients, a detector with mean AUC 0.83 had mean PPV 0.38 at a simulated 10% movement base rate

We evaluated the positive predictive value (PPV) of an offline, cued MOVE-versus-REST detector based on 70–150 Hz high-gamma activity (HGA) features from 15 patients with stereo-electroencephalography (SEEG). The archived window features came from OpenNeuro ds006253 through a prior SCI contribution; the recorded provenance contains source object keys and hashes but does not identify an exact dataset release. A log10, z-scored Ledoit–Wolf shrinkage linear discriminant analysis (LDA) was trained on the first 60% of trials in time order and tested on the last 40%. MOVE was the mean HGA from 0 to 0.5 s after mouse-movement onset; REST was the 0.5 s before the stimulus. Thresholds targeted a 0.05 false-positive rate (FPR), set using either in-sample training REST scores or five-fold time-ordered cross-validated training scores. Base rates were simulated by reweighting test true-positive and false-positive rates, not measured during continuous decoding. Mean test AUC was 0.833. At a simulated movement base rate of 0.1, mean PPV was 0.38 (median 0.35), and no patient reached 0.80. The in-sample threshold yielded mean test TPR 0.645 and FPR 0.134, 2.7 times the target (p = 0.0005; Holm-adjusted p = 0.001). The cross-validated threshold yielded mean TPR 0.430, FPR 0.050, and mean PPV(0.1) 0.53; 1 of 15 patients reached 0.80. An oracle threshold selected on the test block with TPR at least 0.5 yielded mean PPV(0.1) 0.52, with 3 of 15 patients reaching 0.80, indicating that threshold choice did not account for most of the PPV shortfall. Test AUC was associated with PPV(0.1) (Spearman rho = 0.73, p = 0.003); an equal-variance binormal prediction from AUC gave mean PPV(0.1) 0.47 at FPR 0.05. At the observed TPR, PPV 0.80 required a median test FPR of 0.018. These findings support reporting PPV at a stated base rate alongside AUC and setting thresholds on cross-validated training scores. Limits include the offline cued task, one pre-stimulus REST window per trial, simulated base rates, and coarse single-patient FPR resolution (48–120 test REST windows; 0.008–0.021). Continuous rest may contain more varied non-movement states, so real PPV may be lower.

positive predictive value · PPV · stereo-electroencephalography · SEEG · brain-computer interface · high-gamma activity · movement detection · base rate · false-positive rate · linear discriminant analysis · cross-validation