Rapid Prediction of Brain Injury Pattern in mTBI by Combining FE Analysis With a Machine-Learning Based Approach

• Published in: IEEE access 2020, Vol.8, p.179457-179465
• Main Authors: Shim, Vickie B., Holdsworth, Samantha, Champagne, Allen A., Coverdale, Nicole S., Cook, Douglas J., Lee, Tae-Rin, Wang, Alan D., Li, Shaofan, Fernandez, Justin W.
• Format: Article
• Language: English
• Subjects: Brain damage; Brain modeling; Computation; Computational modeling; Diagnostic software; Diagnostic systems; Diffusion tensor imaging; Finite element analysis; Finite element method; Head; Head injuries; Impact damage; Injury analysis; Machine learning; Magnetic heads; Magnetic resonance imaging; mild traumatic brain injury; partial least squares regression; Regression analysis; Regression models; Statistical analysis; Statistical models; Strain; Traumatic brain injury
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.3026350

Mild traumatic brain injury (mTBI) is a significant issue worldwide. Public awareness of the dangers of mTBI has increased sharply in recent years, yet there is no easy-to-use tool available for early detection and post injury management. Computational models of the head impact, usually in the form of finite element analysis, are a method of choice for characterizing how mechanical impacts lead to brain damage by causing high strains in certain regions of the brain. However, those models require a prohibitively large amount of computational power as well as pre and post processing expertise, making them unrealistic to be used in clinical settings. In this study, we propose a framework that combines finite element analysis with a machine learning based approach where a large number of pre-computed FE results are used to train a statistical model. We analyzed a number of different head impact scenarios in which a football player would sustain a minor brain injury and computed brain internal strain patterns. These pre-computed strain patterns were then used to train a partial least squares regression model to be able to predict the general strain pattern and the location and magnitude of peak strains. Our models were able to predict the overall distribution pattern, including the location of the peak strain, with an average error of 3%. The peak strain magnitudes were also predicted accurately with the average error of 9% at almost real time speed (less than 10 seconds). This model may play an important role in developing a diagnostic tool for mTBI that can predict the severity of head impacts.