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

Area of Science:

• Sports Medicine
• Computational Neuroscience
• Artificial Intelligence

Background:

• Concussion diagnosis is challenging due to subtle symptoms, subjective assessments, and limited real-time data integration.
• Existing computational models often lack transparency (black boxes) and the robustness needed for critical applications.

Purpose of the Study:

• To develop a novel, explainable, hybrid deep learning framework for early concussion risk prediction.
• To improve the accuracy and reliability of concussion detection by integrating multimodal data.

Main Methods:

• A symmetry-aware, multiscale deep learning framework combining kinematic sensor data, EEG, contextual information, and synchronized video.
• Utilized modality-specific feature extractors and an attention-based multimodal fusion layer.
• Incorporated explainability using SHAP value analysis and attention visualizations, evaluated via subject-independent 5-fold cross-validation.

Main Results:

• Achieved high performance with 0.91 accuracy, 0.90 macro F1-score, and 0.95 AUC, outperforming single-modality and traditional multimodal approaches.
• Effectively recalled high concussion risk events (F1-score 0.89) and predicted moderate events (F1-score 0.87).
• Demonstrated the efficacy of symmetry-aware multiscale fusion and explainable AI in concussion prediction.

Conclusions:

• The developed framework provides a robust and explainable tool for concussion risk prediction.
• This approach shows significant promise for clinical decision support in managing sports-related head injuries.
• Further clinical validation in diverse populations is required for real-world implementation.