Toward more accurate and generalizable brain deformation estimators for traumatic brain injury detection with unsupervised domain adaptation

TL;DR

This work tackles the generalization gap in machine learning head models for traumatic brain injury by coupling unsupervised domain adaptation with a neural network that predicts whole-brain deformation metrics $MPS$ and $MPSR$. The authors compare domain regularized component analysis (DRCA) and Cycle-GAN–based methods, demonstrating that DRCA substantially improves estimation accuracy across on-field datasets (CF1, MMA) and hold-out sets (CF2, Boxing) with reductions in both MAE and RMSE. Key contributions include a large FE-simulated training set (HM, $12{,}780$ impacts), a detailed feature engineering pipeline, and empirical evidence that DRCA outperforms deep translation approaches in small-data regimes. The approach enables rapid, reliable brain deformation estimates in real-world settings, supporting better TBI detection and protection in high-risk populations.

Abstract

Machine learning head models (MLHMs) are developed to estimate brain deformation for early detection of traumatic brain injury (TBI). However, the overfitting to simulated impacts and the lack of generalizability caused by distributional shift of different head impact datasets hinders the broad clinical applications of current MLHMs. We propose brain deformation estimators that integrates unsupervised domain adaptation with a deep neural network to predict whole-brain maximum principal strain (MPS) and MPS rate (MPSR). With 12,780 simulated head impacts, we performed unsupervised domain adaptation on on-field head impacts from 302 college football (CF) impacts and 457 mixed martial arts (MMA) impacts using domain regularized component analysis (DRCA) and cycle-GAN-based methods. The new model improved the MPS/MPSR estimation accuracy, with the DRCA method significantly outperforming other domain adaptation methods in prediction accuracy (p<0.001): MPS RMSE: 0.027 (CF) and 0.037 (MMA); MPSR RMSE: 7.159 (CF) and 13.022 (MMA). On another two hold-out test sets with 195 college football impacts and 260 boxing impacts, the DRCA model significantly outperformed the baseline model without domain adaptation in MPS and MPSR estimation accuracy (p<0.001). The DRCA domain adaptation reduces the MPS/MPSR estimation error to be well below TBI thresholds, enabling accurate brain deformation estimation to detect TBI in future clinical applications.

Figures (4)

• Figure 1: Distribution of MPS and MPSR of HM, CF1, MMA, CF2 and Boxing datasets. The distribution of the peak angular velocity magnitude (peak Ang. Vel. Mag., A), the peak angular acceleration magnitude (peak Ang. Acc. Mag. B), the 95th percentile MPS (C) and the distribution of the 95th percentile MPSR (D). The principal component analysis (PCA) visualization (E) and t-distributed stochastic neighbor embedding (t-SNE) visualization of all types of head impacts (F). Note: 95th percentile is frequently used in TBI biomechanics research to avoid numerical instability in the maximum values.
• Figure 2: The pipeline of the model development. The pipeline consists of a deformation prediction model and domain adaptation performed by DRCA or cycle-GAN. Specifically, it shows the development of basis model on dataset HM (A), the baseline method (B), the DRCA method (C), and the Cycle-GAN method (D).
• Figure 3: The 3D visualization of two example cases. Upper: MPS. Lower: MPSR. The reference values, the prediction of the baseline method and the DRCA method are shown.
• Figure 4: The accuracy of DRCA and baseline model on the two hold-out test sets. The distribution of mean absolute error over all test impacts in MPS estimation (A) and in MPSR estimation (B) on hold-out test datasets CF2 and Boxing. The MAE is computed for each test impact. Statistical significance of pair-wise comparison with the baseline method (paired t-test) is shown: *: $p<0.05$, *: $p<0.01$, ***: $p<0.001$