In 59 SEEG patients, a label-free outlier score ranked clinical SOZ pairs above chance but below high-frequency activity power

Archived interictal bipolar-pair features from 59 patients in OpenNeuro ds004100 (30 patients) and ds003876 (29 patients) were used to test whether a label-free, within-patient multivariate outlier score ranked clinically labeled seizure onset zone (SOZ) pairs above other pairs and better than 80–150 Hz high-frequency activity (HFA) power. The primary score was the Mahalanobis distance of rank-based normal scores across 10 features, with covariance shrinkage of 0.1; clinical labels were not used to form scores. Mean per-patient AUC was 0.556 (95% CI 0.507–0.604; Holm-adjusted p=0.037) in ds004100 and 0.582 (0.529–0.636; p=0.028) in ds003876. The outlier score was below HFA in both cohorts: OUT−HFA was −0.097 (95% CI −0.169 to −0.022; Holm p=0.037) and −0.100 (−0.167 to −0.034; p=0.028), respectively. Removing spike rate reduced OUT AUC to 0.523 and 0.556 and increased the loss against HFA to −0.130 and −0.125 (Holm p=0.005 in both cohorts). PC1 was closer to HFA, but differences were not significant and did not meet the prespecified equivalence margin of ±0.05. The direction-free score can mark pairs that differ from the patient’s other pairs, including pairs with low power or spike rate, without identifying SOZ pairs. Limits include clinical rather than true epileptogenic-zone labels, one 120-s clip per patient, only 10 archived features, and no evaluation of other unsupervised scores.

SEEG · seizure onset zone · outlier detection · interictal features · high-frequency activity · AUC · epilepsy