Brain Deformation Estimation With Transfer Learning for Head Impact Datasets Across Impact Types.

Area of Science:

• Biomechanics
• Computational modeling
• Neuroscience

Background:

• Machine-learning head models (MLHMs) accelerate brain strain and strain rate calculations, key predictors for traumatic brain injury (TBI).
• MLHM accuracy declines with diverse head impact types (e.g., car crashes, football), limiting their real-world application.
• Small datasets for specific impact types hinder the development of accurate, specialized MLHMs.

Purpose of the Study:

• To develop robust MLHMs applicable across various head impact types.
• To enhance the prediction accuracy of maximum principal strain (MPS) and maximum principal strain rate (MPSR).
• To overcome data limitations for specific impact types using advanced machine learning techniques.

Main Methods:

• Proposed data fusion and transfer learning strategies to create a series of MLHMs.
• Tested models on diverse datasets including American football, mixed martial arts, and reconstructed car crash impacts.
• Evaluated model performance based on mean absolute error (MAE) for MPS and MPSR predictions.

Main Results:

• MLHMs developed with transfer learning demonstrated significantly higher accuracy in estimating MPS and MPSR across all tested impact datasets.
• Achieved MAE below 0.03 for MPS and below [Formula: see text] for MPSR.
• Transfer-learning-based models showed high performance in concussion detection using estimated MPS and MPSR.

Conclusions:

• Developed MLHMs using transfer learning are applicable to various head impact types for rapid and accurate brain strain/strain rate calculation.
• This approach enables MLHM development even with limited data for specific head impact types.
• Accelerates the broader application of MLHMs in TBI research and prevention.