Proactive concussion prediction using symmetry-aware multiscale explainable hybrid deep learning and multimodal data fusion

Abstract

Abstract Diagnosis of concussion has been hampered by its subtle and late symptoms, the reliance on subjective judgment, and the lack of real-time multimodal data integration. The developed computational models are generally black boxes and lack the robustness required for safety-critical scenarios. We introduce a symmetry-aware, multiscale, explainable, hybrid deep learning framework for early prediction of concussion risk. The system effectively combines kinematic sensor data, EEG, contextual information, and synchronized video using separate modality-specific feature extractors and a multimodal fusion layer that utilizes attention-based multimodality fusion. The framework is evaluated using subject-independent 5-fold cross-validation to evaluate its generalizability. The proposed framework incorporates explainability with SHAP value analysis and attention visualizations. The proposed system achieves an accuracy of 0.91, a macro F1-score of 0.90, and an AUC of 0.95, outperforming baselines that rely on single modalities and traditional multimodal integration. High concussion risk events are recalled with an F1-score of 0.89, and moderate concussion events are predicted with an F1-score of 0.87 to capture subtle injury characteristics. We showed that symmetry-aware multiscale fusion, combined with an explainable decision support system, provides a useful tool for concussion risk prediction. This framework holds promise for clinical decision support, although further clinical trials with larger, more diverse populations are necessary to establish its reliability for real-world implementation.