Learnable ICA and pathophysiology-informed consensus: a three-stage framework for EEG-based stroke and TBI classification

We present an engineering framework for EEG-based neurological classification that integrates pathophysiology-informed multi-view learning, end-to-end trainable adaptive preprocessing, and temporal-aware data augmentation. Accurate three-class differentiation of stroke and traumatic brain injury (TBI) from normal electroencephalogram (EEG) remains challenging due to overlapping spectral signatures, heterogeneous artifacts, and limited training data. The proposed three-stage framework addresses each bottleneck through complementary mechanisms. Stage 1 introduces pathophysiology-informed consensus learning, using three frequency-band configurations that target TBI-specific delta dynamics, stroke-related hemispheric asymmetry, and shared cross-pathology features, combined through an adaptive consensus function with expertise-weighted probability aggregation. Stage 2 introduces the Dynamic Weight Learning ICA module (DWL-ICA), a compact one-dimensional convolutional neural network (approximately 2500 parameters) that transforms ICLabel probability distributions into sample-adaptive component importance weights, trained end-to-end with the classifier. Unlike conventional binary ICA rejection that discards all non-brain components, DWL-ICA learns to retain diagnostically informative components in a graded manner. Stage 3 applies sliding-window Welch decomposition for temporal-aware data augmentation, capturing intra-recording spectral dynamics that static full-segment averaging discards. Evaluated on the Temple University Hospital EEG Corpus (9276 segments, 1390 subjects, 19 channels), the complete framework achieves 88.22% balanced accuracy and 0.9774 macro-average ROC-AUC. A subject-level split experiment yields 0.748 macro-AUC, indicating that cross-subject generalization remains an open challenge. The segment-level results substantially exceed the prior three-class benchmark of 0.85 micro-average AUC reported on the same corpus. Cross-stage ablation reveals that the three stages contribute through distinct and complementary mechanisms, and analysis of the learned DWL-ICA weights reveals that components conventionally treated as artifacts (notably Heart and Eye) carry diagnostically informative content, providing data-driven evidence against blanket artifact rejection.