Machine Learning Classification of Verified Head Impact Exposure Strengthens Associations with Brain Changes.

Abstract

Cumulative exposure to head impacts during contact sports can elicit potentially deleterious brain white matter alterations in young athletes. Head impact exposure is commonly quantified using wearable sensors; however, these sensors tend to overestimate the number of true head impacts that occur and may obfuscate potential relationships with longitudinal brain changes. The purpose of this study was to examine whether data-driven filtering of head impact exposure using machine learning classification could produce more accurate quantification of exposure and whether this would reveal more pronounced relationships with longitudinal brain changes. Season-long head impact exposure was recorded for 22 female high school soccer athletes and filtered using three methods-threshold-based, heuristic filtering, and machine learning (ML) classification. The accuracy of each method was determined using simultaneous video recording of a subset of the sensor-recorded impacts, which was used to confirm which sensor-recorded impacts corresponded with true head impacts and the ability of each method to detect the true impacts. Each filtered dataset was then associated with the athletes' pre- and post-season MRI brain scans to reveal longitudinal white matter changes. The threshold-based, heuristic, and ML approaches achieved 22.0% accuracy, 44.6%, and 83.5% accuracy, respectively. ML classification also revealed significant longitudinal brain white matter changes, with negative relationships observed between head impact exposure and reductions in mean and axial diffusivity and a positive relationship observed between exposure and fractional anisotropy (all p < 0.05).