Rapid Prediction of Brain Tissue Response in Pig TBI Models Using Machine Learning

Introduction: : Traumatic brain injury (TBI) is a significant health burden worldwide, leading to long-term cognitive, behavioral, and neurological consequences. TBI typically results from kinematic loading of the head that causes rapid brain tissue deformation and stretching of axons. Pig models of TBI serve as valuable surrogates in research due to their anatomical similarities to the human brain. These models offer a unique opportunity to isolate and investigate the underlying mechanisms behind specific TBI outcomes and to access neuropathological data that cannot be obtained from living humans. On the other hand, brain finite element models (FEMs) are powerful tools for predicting neuropathological outcomes in TBI by providing estimates of brain tissue responses based on head kinematics [1]. Studies have shown that FEM-derived brain tissue and axonal strains in pig TBI models produce injury outcomes similar to those observed in humans, confirming the utility of brain FEMs as translational tools in neurotrauma research [2]. However, FEMs are computationally intensive, time-consuming, and require specialized expertise and resources, making them less practical for widespread use in animal TBI experiments. To address these limitations, we developed a machine learning (ML) framework that estimates key biomechanical injury metrics—maximum principal strain (MPS) and maximum axonal strain (MAS)—directly from input kinematic traces in pig TBI models. This approach aims to provide an accessible and efficient alternative to FEMs, enabling researchers in the TBI community to estimate tissue deformation responses and potential severity of pathology directly from experimental kinematic data, without the need for complex computational simulations.

Materials and

Methods: : Finite element simulations of uniaxial rapid head rotation (25–250 rad/s) about the axial (n=142) and sagittal (n=197) planes were conducted using a validated pig brain FEM [3]. Input kinematics were obtained from (1) synthetic velocity traces and (2) experimental data from a well-established non-impact rapid head rotation pig TBI model [4]. For each simulation, two output measures were extracted: MPS95 and MAS95, representing the 95th percentile of maximum principal strain in brain tissue and the 95th percentile of maximum strain in axonal fibers, respectively. Figure 1 illustrates examples of axonal strain distributions and kinematic traces for axial and sagittal experimental cases [1]. Fifteen input features were extracted from the angular velocity traces. Data were randomly split into training (70%), validation (15%), and test (15%) sets. Five models were developed in MATLAB: Lasso and Ridge regressions served as baseline linear models incorporating L1 and L2 regularization to prevent overfitting. Random Forest was employed as a nonlinear ensemble method, with a variant using curvature-based predictor selection to better capture nonlinear feature interactions. A Deep Neural Network (DNN) was also trained using MATLAB’s FitNet architecture to evaluate performance in learning complex, high-dimensional patterns. All models were trained separately for axial and sagittal impacts to account for directional effects. Performance was assessed using R², MAE, and RMSE, averaged over 100 runs for all models except the DNN, which was limited to 10 runs due to its higher computational cost. A repeated-split approach was used to mitigate bias from random data partitioning.

Results, Conclusions, and Discussions:: Overall, the Random Forest and DNN models outperformed the linear methods in predicting tissue deformation, with MAE values indicating superior precision, as shown in Figure 2. Among all models, the default Random Forest achieved the best overall performance. For MPS95, it reported MAEs of 0.0114 (sagittal) and 0.0127 (axial), while for MAS95, the MAEs were 0.0067 (sagittal) and 0.0087 (axial). These low error margins, compared to the previously determined TBI injury thresholds for MPS95 (0.3) and MAS95 (0.13) [3], suggest high reliability in strain prediction. The model also achieved strong R² values: 0.9853 and 0.9788 for MPS95 and 0.9749 and 0.9720 for MAS95 in the sagittal and axial directions, respectively. DNNs using the FitNet architecture showed competitive MAE performance, particularly for axial MAS95 estimation (MAE = 0.0065, R² = 0.9699), but demonstrated greater variability across repetitions, as evidenced by higher standard deviations. While Lasso and Ridge regressions did not reach the accuracy of the nonlinear models, they still produced acceptable MAEs, highlighting their practicality under resource-constrained scenarios.
These results demonstrate that ML models—particularly Random Forests—can effectively serve as practical surrogates for FEM in estimating brain deformation metrics [3]. This significantly reduces computational demands and technical barriers for researchers in the TBI community. Our framework allows those conducting experimental animal studies to estimate MPS95 and MAS95 without requiring full FEM pipelines. We plan to publicly release these models as a user-friendly tool to encourage widespread adoption in TBI research.
Future work will focus on extending the model’s capabilities beyond single-value strain metrics to enable spatial prediction of kinematic responses throughout the brain. This enhancement aims to improve understanding of spatial tissue deformation, offering insights into localized pathology and injury mechanisms.

Acknowledgements and/or References (Optional):: NSF #2138719

References:
1. Hajiaghamemar, M., Seidi, M., & Margulies, S. S. (2020). Head rotational kinematics, tissue deformations, and their relationships to the acute traumatic axonal injury. Journal of biomechanical engineering, 142(3), 031006.
2. Wu, T., Hajiaghamemar, M., Giudice, J. S., Alshareef, A., Margulies, S. S., & Panzer, M. B. (2021). Evaluation of tissue-level brain injury metrics using species-specific simulations. Journal of neurotrauma, 38(13), 1879-1888.
3. Hajiaghamemar, M., Wu, T., Panzer, M. B., & Margulies, S. S. (2020). Embedded axonal fiber tracts improve finite element model predictions of traumatic brain injury. Biomechanics and modeling in mechanobiology, 19, 1109-1130.
4. Margulies, S. S., Kilbaugh, T., Sullivan, S., Smith, C., Propert, K., Byro, M., ... & Duhaime, A. C. (2015). Establishing a clinically relevant large animal model platform for TBI therapy development: using C yclosporin a as a case study. Brain Pathology, 25(3), 289-303.