A sleep-signal power model weakly identifies white-matter SEEG pairs in unseen patients in an 18-patient UCLA cohort

A preregistered leave-one-patient-out study tested whether seven sleep-signal features could identify white-matter (WM) stereoelectroencephalography (SEEG) bipolar pairs in unseen patients, using archived tables from 18 UCLA patients in OpenNeuro ds005398 v1.1.3. The tables contain 1,285 pairs; 50 atlas-labelled Outside pairs were excluded, leaving 1,235 pairs, including 251 WM pairs. A seven-feature L2 logistic model had mean patient AUC 0.734 (95% CI 0.646–0.809; 17/18 patients above 0.5; Holm p=0.0007), below the preregistered 0.75 useful-detector threshold. It did not improve on low 80–300 Hz power alone (AUC difference +0.002, 95% CI −0.033 to +0.040; Holm p=0.92). At a threshold retaining 90% of gray-matter pairs, it detected 30.6% of WM pairs. Removing detector-flagged pairs reduced the SOZ AUC for 80–300 Hz power to 0.682, below both all pairs (0.728) and atlas-labelled gray matter (0.700); low power is also a non-SOZ marker. Transfer to SOZ pairs was not testable because only three patients had at least two WM and two GM SOZ pairs. In this pediatric, single-centre cohort, the power rule found WM pairs above chance but is not a substitute for imaging labels and should not be used to clean pairs before power-based SOZ analyses. Atlas labels, short sleep windows, and a single untuned model limit interpretation.

stereoelectroencephalography · white matter · bipolar pairs · machine learning · interictal sleep · seizure onset zone · power features · leave-one-patient-out validation