Concussion classification via deep learning using whole-brain white matter fiber strains

Area of Science:

• Biomechanics
• Neuroscience
• Machine Learning

Background:

• Current traumatic brain injury (TBI) prediction relies on limited scalar metrics, potentially losing crucial information.
• Existing methods often lack rigorous cross-validation, impacting reliability.
• There is a need for advanced injury prediction models in sports and accident research.

Purpose of the Study:

• To develop and evaluate a deep learning model for concussion classification using voxel-wise white matter fiber strains.
• To compare the deep learning approach against traditional scalar injury metrics and other machine learning classifiers.
• To assess the predictive performance using leave-one-out cross-validation.

Main Methods:

• A deep learning model was developed to analyze implicit features of white matter fiber strains across the entire brain.
• Performance was evaluated using reconstructed American National Football League (NFL) injury cases with leave-one-out cross-validation.
• Comparisons were made against support vector machine (SVM), random forest (RF), and scalar metrics like Brain Injury Criterion (BrIC).

Main Results:

• Feature-based machine learning classifiers, including deep learning, SVM, and RF, consistently outperformed scalar injury metrics.
• Deep learning achieved superior leave-one-out accuracy (0.828-0.862) compared to scalar metrics (0.690-0.776).
• Deep learning demonstrated the best cross-validation accuracy, sensitivity, AUC, and .632+ error, indicating robust predictive power.

Conclusions:

• Deep learning models utilizing voxel-wise white matter fiber strains show significantly improved concussion prediction performance.
• This approach overcomes limitations of traditional scalar metrics by leveraging comprehensive strain data.
• The findings highlight the potential of deep learning for advancing biomechanical studies of traumatic brain injury.