Spectral features rank artifact-labeled SEEG contacts in most runs but fail when many contacts are artifacts

We asked whether simple, label-free signal features identify SEEG contacts labeled ‘artifact’ by clinicians. We analyzed 60 seconds from the middle of each EDF in 21 NREM-sleep runs (128 SEEG contacts per run) from OpenNeuro ds007118, ds007119, and ds007120. Sixteen runs had both artifact and other contacts (620 contacts, 110 labeled artifact). For each contact, we calculated log RMS, line-noise ratio, 150–250 Hz / 2–30 Hz power ratio, maximum absolute correlation with another SEEG contact, and flat-sample fraction. Each feature was converted to a within-run robust z-score; their largest absolute value formed a label-free composite. Median within-run AUC was 1.00 for line-noise ratio (15/16 runs above 0.5; exact two-sided sign-test p = 0.0005), 0.99 for the high-frequency ratio (16/16; p = 3×10⁻⁵), and 0.97 for the composite (16/16; p = 3×10⁻⁵). RMS and neighbor correlation were not significant. At composite |z| > 5, median per-run sensitivity was 1.00 and false-alarm rate was 0.036, while pooled sensitivity was 0.37 and pooled false-alarm rate was 0.061. Two runs contained 83 of 110 artifact contacts and had composite AUCs of 0.65 and 0.62; within-run robust normalization therefore failed when many contacts were artifacts. Limits include 60 seconds per run, unknown clinician label criteria, 16 evaluable runs, concentration of artifact labels in two runs, likely average referencing, and missing SEEG channels in two EDFs. The labels were not used to choose feature parameters. These findings support a first-pass detector when bad contacts are few; they do not establish performance for other recordings or reference schemes.

SEEG · intracranial EEG · channel quality · artifact detection · line noise · high-frequency power · robust z-score · AUC · NREM sleep · BCI