Robust multi-centre interictal EEG biomarker for distinguishing epilepsy from mimickers.

Abstract

To develop and validate an interpretable multi-centre interictal EEG biomarker for distinguishing epilepsy from mimickers, addressing the critical gap of single-centre studies with limited sample sizes in prior literature. We analysed routine interictal EEG from 448 subjects across two tertiary centres (IHBAS: N = 230; MAX: N = 218), encompassing diverse epilepsy subtypes and mimicker categories. A 13-dimensional Interictal Clinical Signature (ICS) encoding spectral slowing, posterior dominant rhythm, complexity, and network synchrony was computed from 10-second epochs. A two-stage logistic regression framework aggregated epoch-level features into subject-level predictions. Recording-length confounding was assessed through predictor exclusion and fixed-duration truncation analyses. Within-centre AUCs were 0.723 (95% CI: 0.655-0.783) for IH-BAS and 0.790 (95% CI: 0.722-0.845) for MAX. When recording length was excluded as a predictor, cross-centre generalisation was symmetric (MAX→ IHBAS: 0.725; IHBAS→MAX: 0.725). Fixed-duration truncation (10-30 minutes) confirmed consistent cross-centre AUCs of 0.70-0.72. Performance remained stable between 20-125 Hz sampling rates. Low-dimensional, clinically interpretable EEG features support robust epilepsy-versus-mimicker classification within and across centres when recording-length confounds are controlled. This multi-centre validation demonstrates that the ICS framework provides an accessible decision-support tool for epilepsy diagnostics with centre-invariant performance, while the systematic confound analysis offers a methodological template for EEG biomarker studies.