Adaptive Machine Learning Head Model Across Different Head Impact Types Using Unsupervised Domain Adaptation and...

Area of Science:

• Biomechanics
• Computational Neuroscience
• Machine Learning

Background:

• Machine learning head models (MLHMs) estimate brain deformation for traumatic brain injury (TBI) detection.
• Current MLHMs suffer from overfitting to simulated data and reduced accuracy due to dataset shifts, limiting clinical use.

Purpose of the Study:

• To develop an improved MLHM integrating unsupervised domain adaptation for accurate prediction of whole-brain maximum principal strain (MPS) and MPS rate (MPSR).
• To enhance TBI detection capabilities through more reliable brain deformation estimation.

Main Methods:

• A deep neural network was configured with unsupervised domain adaptation techniques, including domain-regularized component analysis (DRCA) and cycle-GAN.
• The model was trained and validated using extensive simulated and real-world head impact datasets (12,780 simulated, 302 college football, 457 MMA impacts).
• Performance was evaluated by comparing MPS and MPSR estimation accuracy against baseline models and TBI thresholds.

Main Results:

• The DRCA method demonstrated superior performance in MPS and MPSR estimation accuracy across different impact datasets (e.g., MPS MAE of 0.017 for CF, 0.020 for MMA).
• The DRCA model significantly outperformed the baseline model on hold-out test sets of college football and boxing impacts.
• Achieved estimation errors for MPS/MPSR were substantially lower than previously reported TBI detection thresholds.

Conclusions:

• The proposed MLHM with DRCA unsupervised domain adaptation effectively addresses overfitting and distributional shift issues in MLHM development.
• This approach enables accurate brain deformation estimation, crucial for the early clinical detection of TBI.
• The findings support the potential for widespread clinical application of advanced MLHMs in TBI diagnostics.