Lesion-specific EEG signatures in stroke: a multi-scale framework integrating oscillations, connectivity, and asymmetry for machine learning decoding

Lesion location is a major source of post-stroke neurophysiological heterogeneity, yet most electroencephalography (EEG) studies analyze patients as a single group, limiting lesion-specific biomarkers and translation. We proposed a lesion-centric, multi-scale EEG framework integrating local oscillations, inter-regional connectivity, and hemispheric asymmetry with machine learning to characterize and decode basal ganglia (P1), fronto-temporal/centrum semiovale (P2), and brainstem (P3) lesions. Five-minute eyes-open, 128-channel resting EEG ($$1\,\text {kHz}$$) was recorded in 57 subacute stroke patients (P1 = 22, P2 = 18, P3 = 17) and 22 matched controls. From artifact-minimized $$90\,\text {s}$$ segments, ROI-averaged power spectral density (PSD) ($$\theta $$: 4–$$7\,\text {Hz}$$; $$\alpha $$: 7–$$12\,\text {Hz}$$; $$\beta _{1}$$: 12–$$16\,\text {Hz}$$; peak $$\alpha $$ frequency), current source density (CSD)-based magnitude-squared coherence, and directional BSI (dirBSI) were computed. Between-group and subgroup differences were assessed using t-tests/Wilcoxon and ANOVA/Kruskal–Wallis with Benjamini–Hochberg FDR ($$q=0.05$$). EEG–behavior associations were examined with Spearman correlations. For machine learning, common spatial patterns (CSP) features were classified using linear discriminant analysis (LDA) with leave-one-subject-out cross-validation. To align with clinical workflow, we report HC vs P as “stroke detection/screening” and patient-only P1/P2/P3 classification as “lesion subtype decoding for stratification” (along with pairwise P1 vs P2, P1 vs P3, and P2 vs P3 models). An EEGNet baseline was evaluated for comparison. Increased $$\alpha $$ power and a leftward peak $$\alpha $$ shift were observed in patients (HC: $$9.93 \pm 1.09\,\text {Hz}$$; P: $$8.75 \pm 1.02\,\text {Hz}$$; $$p = 6.82 \times 10^{-5}$$). Pre-FDR, $$\theta $$-band frontal–motor connectivity was strengthened, while posterior P–O connectivity in $$\alpha $$/$$\beta _{1}$$ was weakened. Ipsilesional dominance in $$\theta $$ was indicated by dirBSI (HC: $$-0.026 \pm 0.101$$; P: $$0.061 \pm 0.122$$; $$q=0.012$$). Across lesions, $$\beta _{1}$$ power differences in central/parietal/occipital ROIs were detected pre-FDR, with higher parietal $$\beta _{1}$$ in P3; $$\alpha $$-band asymmetry was stronger in P1/P2 and more symmetric in P3 ($$q=0.028$$). EEG–behavior correlations did not survive FDR. Using CSP+LDA, accuracies of 92.41% (HC vs P), 94.87% (P1 vs P3), 85.71% (P2 vs P3), and 82.50% (P1 vs P2) were achieved; all binary AUCs exceeded 0.85; three-class accuracy reached 85.96%. This multi-scale EEG framework identifies lesion-associated neurophysiological signatures and demonstrates feasible lesion subtype decoding, supporting the potential of EEG biomarkers for objective stratification and precision neurorehabilitation.