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

Xianghao Zhan, Yuzhe Liu+5 more|arXiv (Cornell University)|Aug 31, 2021

Automotive and Human Injury Biomechanics56 references4 citations

TL;DR

This study proposes a transfer learning and data fusion approach to train machine learning head models (MLHMs) that rapidly and accurately predict maximum principal strain (MPS) and strain rate (MPSR) across diverse head impact types—such as football, MMA, and car crashes—achieving a mean absolute error (MAE) of less than 0.03 for MPS and 7 s⁻¹ for MPSR, significantly outperforming models trained solely on simulation or on-field data.

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.

Motivation & Objective

• To overcome the high computational cost and limited real-time applicability of finite element modeling (FEM) for brain strain and strain rate estimation.
• To address the poor generalization of machine learning head models (MLHMs) when applied to on-field impact data different from their training data.
• To improve model accuracy and robustness across diverse impact types—such as American football, MMA, and car crashes—using data fusion and transfer learning.
• To enable real-time, region-specific TBI risk monitoring by accurately predicting both MPS and MPSR from head kinematics.

Proposed method

• The authors trained MLHMs using a combination of 13,623 simulated head impacts and real-world on-field impact data from American football, MMA, and car crashes.
• They employed transfer learning by first pre-training the model on large-scale simulated data and then fine-tuning it on smaller on-field datasets for each impact type.
• Data fusion was implemented by jointly training the model on both simulated and on-field datasets to improve generalization.
• The input features included linear acceleration, angular velocity, angular acceleration, and angular jerk, engineered to capture temporal and spectral characteristics of impact kinematics.
• Model performance was evaluated using mean absolute error (MAE) for MPS and MPSR across all impact types.
• The KTH finite element model was used as a reference for ground truth strain and strain rate values.

Experimental results

Research questions

• RQ1 Can transfer learning improve the accuracy of MLHMs in predicting brain strain and strain rate across diverse on-field impact types?
• RQ2 How does data fusion between simulated and on-field impact data affect model generalization and prediction accuracy?
• RQ3 Can MLHMs trained with transfer learning achieve real-time, accurate prediction of both MPS and MPSR, outperforming models trained only on simulation or on-field data?
• RQ4 To what extent do engineered kinematic features (acceleration, angular velocity, etc.) support accurate strain and strain rate estimation in low-data regimes?

Key findings

• The MLHMs trained with transfer learning achieved a mean absolute error (MAE) of less than 0.03 in predicting maximum principal strain (MPS) across all impact types.
• The same models achieved an MAE of less than 7 s⁻¹ in predicting maximum principal strain rate (MPSR), demonstrating high accuracy for strain rate estimation.
• Transfer learning-based models significantly outperformed models trained only on on-field data or only on simulated data, especially on unseen impact types.
• The fusion of simulated and on-field data enhanced model robustness and generalization, enabling accurate predictions across diverse impact scenarios.
• The proposed MLHMs enable rapid computation of whole-brain strain and strain rate in real time, overcoming the computational burden of finite element modeling.
• The models provide region-specific strain and strain rate information, supporting more detailed biomechanical analysis of TBI risk than traditional brain injury criteria.