Adaptive Machine Learning Head Model Across Different Head Impact Types Using Unsupervised Domain Adaptation and Generative Adversarial Networks
Abstract
Machine-learning head models (MLHMs) have shown promise in estimating traumatic brain injury (TBI)-related brain deformation from head kinematics, but overfitting to simulated impacts and performance degradation under distribution shift hinder clinical deployment. This study introduces an adaptive MLHM combining unsupervised domain adaptation and generative adversarial networks to address dataset distributional differences across simulated impacts, college football, mixed martial arts, and boxing datasets. Using domain regularized component analysis (DRCA) and cycle-GAN-based adaptation, the proposed models markedly improved maximum principal strain (MPS) and strain-rate (MPSR) prediction accuracy. The DRCA-based approach achieved the best performance with MPS mean absolute errors of 0.017 (college football) and 0.020 (MMA), and MPSR MAE of 4.09 s⁻¹ and 6.61 s⁻¹, outperforming baseline MLHMs across multiple hold-out test sets. These results demonstrate that unsupervised domain adaptation effectively reduces cross-domain error and moves MLHMs toward clinically reliable TBI detection.
