Rapid atlas-based predictions of regional brain strain simulations from wearable head kinematics using diffusion MRI-informed machine learning.
Authors
Affiliations (7)
Affiliations (7)
- Auckland Bioengineering Institute, University of Auckland, Auckland, New Zealand.
- Auckland Bioengineering Institute, University of Auckland, Auckland, New Zealand; University of the Immaculate Conception, Davao City, Philippines.
- Auckland Bioengineering Institute, University of Auckland, Auckland, New Zealand; Mātai Medical Research Institute, Gisborne, New Zealand; Faculty of Medical and Health Sciences & Centre for Brain Research, University of Auckland, Auckland, New Zealand.
- Auckland Bioengineering Institute, University of Auckland, Auckland, New Zealand; Mātai Medical Research Institute, Gisborne, New Zealand.
- School of Physical Education Sport and Exercise Sciences, University of Otago, Dunedin, New Zealand.
- Mātai Medical Research Institute, Gisborne, New Zealand; Faculty of Medical and Health Sciences & Centre for Brain Research, University of Auckland, Auckland, New Zealand; Brain Health Research Institute, Auckland University of Technology, Auckland, New Zealand.
- Auckland Bioengineering Institute, University of Auckland, Auckland, New Zealand; Mātai Medical Research Institute, Gisborne, New Zealand. Electronic address: [email protected].
Abstract
Sports-related traumatic brain injury (TBI) remains significantly underdiagnosed, with up to 50% of mild TBI cases in sport going undetected. While finite element (FE) simulations can predict brain deformation from head impacts, their computational cost limits clinical applicability for on-field assessments. This study shows a proof-of-principle in using a personalised machine learning framework capable of rapidly predicting regional brain strain simulations over time from wearable sensor kinematics, requiring as few as 150 head acceleration events (HAE) per athlete for training. Subject-specific anisotropic, viscoelastic FE brain models were constructed from T1-weighted and diffusion MRI for ten high school rugby players. Head kinematics were recorded via instrumented mouthguards and used to simulate the FE brain models. The resulting brain strains were parcellated to two commonly used atlases, including the Desikan-Killiany (supplemented with additional structures for whole-brain coverage, totalling 98 regions) and Automated Anatomical Labelling (116 regions) atlases, with the approach readily adaptable to other atlas schemes. Subject-specific multi-output Random Forest regression models were trained on 150 HAEs for every athlete using 70 time-series-extracted kinematic features (using the tsfresh python package) to predict time-dependent parcellated strain profiles. The ML models achieved an R<sup>2</sup> from 75% to 87% and an average RMSE from 0.02 to 0.04 between target and predicted regional strains, with low inter-athlete variability. This framework captures the location, magnitude, and temporal behaviour of regional brain strain simulations, enabling rapid estimation directly from wearable devices and potentially enabling personalised, real-time concussion monitoring in sport.