Early detection of drug-resistant epilepsy via multi-domain EEG fusion with prefrontal candidate biomarkers

Drug-resistant epilepsy (DRE) is associated with poor clinical outcomes. Interictal scalp EEG-based DRE identification faces a methodological limitation: segment-level models often capture individual-specific EEG patterns rather than group-level pathological differences. In this retrospective study of 42 patients (21 DRE, 21 non-DRE), we constructed an EEG-based DRE recognition framework. Multi-domain features (temporal, spectral, nonlinear) were extracted, and a mutual validation strategy integrating patient-level statistical tests and machine learning was applied to verify the pathological discriminative validity of features. A multi-domain support vector machine was built with ReliefF feature selection, evaluated via leave-one-out cross-validation and hold-out validation, and further analyzed for interpretability and clinical correlations. Significant between-group differences were found in clinical and electrophysiological features, with prominent alterations in prefrontal temporal statistics, θ/β band power and entropy features. The multi-domain model consistently outperformed single-domain models; prefrontal β/γ power and entropy features were identified as core candidate biomarkers, significantly correlated with disease duration and seizure frequency. This study identifies prefrontal EEG features as candidate DRE biomarkers, and the combined validation framework supports the pathological basis of classification performance. Nevertheless, the single-dataset retrospective design and limited benchmark comparisons restrict conclusion generalizability. Further multicenter external validation is warranted to assess the framework's generalization capacity and clinical translational potential.