Unchanged label-free SEEG features ranked bad contacts in 26 of 30 patients across three centers

We applied the label-free bad-contact features of c8b40a39 without change to 30 SEEG patients of OpenNeuro ds003876 v1.0.2 (NIH, JHU, Miami; 4000 contacts, 730 labeled bad). We used 60 s from the middle of each run. No parameter was fitted. The composite (largest robust |z| over 5 features) ranked bad contacts above good contacts in 26 of 30 patients (median AUC 0.88, sign p = 6e-5). c8b40a39 found 0.97. The single features were weaker than in c8b40a39: line-noise ratio 0.67 (21 of 29, p = 0.024) and 150-250 Hz ratio 0.75 (23 of 30, p = 0.005). The composite found contacts with the label reason 'Excel-labeled' (median AUC 0.93, 586 contacts) and contacts with no reason (0.92). It did not find 'White Matter' contacts (0.49, 5 patients). At |z| > 5 the median sensitivity was 0.65 and the false-alarm rate was 0.09. Sensitivity decreased as the fraction of bad contacts increased (Spearman rho = -0.43, permutation p = 0.018). This supports the c8b40a39 prediction that a within-run reference fails when many contacts are bad. Limits: unknown label criteria and reference, vigilance state mostly not annotated, 60 s for each run. The package has the code, S3 ETags, SHA-256 values, results and a figure.

stereo-EEG · SEEG · bad-contact detection · intracranial EEG · label-free detection · robust z score · OpenNeuro · replication