In 15 SEEG patients, z-clipping did not improve movement-decoder AUC for simulated 5-SD events but reduced loss from events of 10 SD or more

We assessed a linear movement detector in archived high-gamma (70–150 Hz) SEEG window features from 15 patients performing a random-dot motion task. We simulated additive events on one random channel in 20% of test windows and compared Ledoit–Wolf shrinkage LDA with no clipping or clipping of calibration-z-scored features at ±3. For +5 SD events, mean AUC was 0.828 without clipping and 0.828 with clipping; the paired mean difference was −0.0002 (10/15 patient differences positive; one-sided exact sign-flip p=0.51). On clean data, clipping reduced mean AUC by 0.0026; non-inferiority within a 0.01 AUC margin was not shown after Holm correction (p=0.058). In descriptive analyses, clipping at ±3 recovered 0.010 AUC for 10-SD events and 0.030 for 20-SD events, while the un-clipped detector lost 0.015 and 0.035 AUC, respectively. Clipping at ±2 performed best in exploratory analyses. One patient with 148 channels and 72 calibration trials accounted for most of the clean-data cost. Events were simulated, affected one channel, and occurred only in test data; conclusions are limited to this task, split, dataset, and linear decoder.

stereo-EEG · high-gamma activity · movement decoding · shrinkage LDA · transient artifacts · feature clipping