Rapidly and accurately estimating brain strain and strain rate across head impact types with transfer learning and data fusion

TL;DR

This work tackles real-time brain injury risk estimation by predicting the maximum principal strain $MPS$ and maximum principal strain rate $MPSR$ from head impact kinematics. It introduces data fusion and transfer learning to adapt a large, simulated-head-impact dataset to on-field impacts, yielding substantial gains in accuracy across diverse datasets. Transfer-learning-based estimators achieve MAEs below $0.03$ for $MPS$ and below $7$ s$^{-1}$ for $MPSR$ across HM, CF, MMA, NFL, and NHTSA, and provide region-specific brain maps suitable for immediate risk assessment. The approach dramatically reduces computation time relative to FEM, enabling near real-time monitoring and informed decision-making in sports and clinical settings.

Abstract

Brain strain and strain rate are effective in predicting traumatic brain injury (TBI) caused by head impacts. However, state-of-the-art finite element modeling (FEM) demands considerable computational time in the computation, limiting its application in real-time TBI risk monitoring. To accelerate, machine learning head models (MLHMs) were developed, and the model accuracy was found to decrease when the training/test datasets were from different head impacts types. However, the size of dataset for specific impact types may not be enough for model training. To address the computational cost of FEM, the limited strain rate prediction, and the generalizability of MLHMs to on-field datasets, we propose data fusion and transfer learning to develop a series of MLHMs to predict the maximum principal strain (MPS) and maximum principal strain rate (MPSR). We trained and tested the MLHMs on 13,623 head impacts from simulations, American football, mixed martial arts, car crash, and compared against the models trained on only simulations or only on-field impacts. The MLHMs developed with transfer learning are significantly more accurate in estimating MPS and MPSR than other models, with a mean absolute error (MAE) smaller than 0.03 in predicting MPS and smaller than 7 (1/s) in predicting MPSR on all impact datasets. The MLHMs can be applied to various head impact types for rapidly and accurately calculating brain strain and strain rate. Besides the clinical applications in real-time brain strain and strain rate monitoring, this model helps researchers estimate the brain strain and strain rate caused by head impacts more efficiently than FEM.

Figures (6)

• Figure 1: The distribution of the example kinematics features, MPS, and MPSR across various datasets. The distribution of the peak angular velocity (Ang. Vel.) magnitude (A), the peak angular acceleration (Ang. Acc.) magnitude (B), the 95th percentile MPS (C) and the 95th percentile MPSR (D).
• Figure 2: The illustration of the model development and assessment processes. The processes to develop the basis models developed for the dataset HM (A), and for the on-field datasets (B). The two proposed strategies in adapting for different on-field impacts and the baseline models we compared against (B). The datasets and models related to the on-field impacts are in dotted gray boxes, while those related to the simulated impacts (dataset HM and basis models) are in solid white boxes. The models are represented with oval boxes and the datasets are represented with rectangular boxes.
• Figure 3: The model accuracy of predicting MPS and MPSR on the dataset CF and dataset MMA with different ratios of training data. The mean absolute error (MAE) of the MPS prediction models on dataset CF (A) and dataset MMA (B). The MAE of the MPSR prediction models on dataset CF (C) and dataset MMA (D). Due to the huge variation among different models, the MAE ranges were selected to clearly show the most accurate models. Note that the training set ratio refer to the percentage of on-field data used for training purposes, which is the "1-2X" in Fig. \ref{['pipeline']} and in Section 2E.
• Figure 4: The region-specific mean absolute error of the transfer learning models on dataset HM, dataset CF and dataset MMA with 70% training data and 15% test data (mean and STD over 20 parallel experiments). BS: brainstem, CC: corpus callosum, CL: cerebellum, GM: gray matter, MB: midbrain, TH: thalamus, WM: white matter.
• Figure 5: The model accuracy of predicting MPS and MPSR on the dataset NFL and dataset NHTSA with 50% training data and 25% test data for each dataset. The mean absolute error (MAE) of the MPS prediction models on dataset NFL (A) and dataset NHTSA (B). The prediction MAE of the MPSR prediction models on dataset NFL (C) and dataset NHTSA (D). Due to the large variation among different models, the MAE ranges were selected to clearly show the most accurate models. Note that the models trained on NFL dataset only failed to predict the MPS and MPSR for the test impacts due to numerical errors.