MACHINE LEARNING CLASSIFICATION OF VERIFIED HEAD IMPACT EXPOSURE REVEALS ASSOCIATIONS WITH LONGITUDINAL WHITE MATTER CHANGES IN FEMALE HIGH SCHOOL SOCCER PLAYERS

BACKGROUND: Longitudinal changes in white matter (WM) integrity have been reported following cumulative exposure to sub-concussive head impacts (SCI) incurred during sports. SCI exposure is typically quantified using accelerometers that use an arbitrary g-force threshold for impact detection; however, this approach does not differentiate true impacts from non-physiological events, such as high-frequency cutaneous vibrations. This leads to a high false positive rate and the tendency to overestimate SCI exposure. HYPOTHESIS/PURPOSE: To examine whether machine learning classification trained on video-verified impacts can produce more accurate quantification of SCI exposure and whether this can reveal associations between cumulative exposure and diffusion tensor imaging (DTI)-measured longitudinal WM changes in athletes. METHODS: Pre- and post-season brain MRI scans were collected from 46 female high school soccer athletes. Athletes’ SCI exposure was recorded during their competitive season using head-mounted accelerometers. 24 athletes were assigned to a “Treatment Group” (TG) and were also video recorded during 17 of their games, while the remaining 22 athletes were assigned to a “Control Group” (CG) and were unrecorded. The TG video was used to verify whether the sensor-recorded impacts during those games were true head impacts or false positives. This dataset, along with the corresponding true/false labels, was used to train a machine learning classifier to learn a mapping between the accelerometer data and their respective labels. The trained model was then used to classify and filter the impacts recorded for the CG athletes by removing their predicted false positives. The CG athletes’ SCI exposure in both the unfiltered and filtered conditions was correlated with their pre- to post-season changes in DTI measures of mean diffusivity (MD), axial diffusivity (AD), radial diffusivity (RD), and fractional anisotropy (FA). RESULTS: During the TG athletes’ 17 video-recorded games, 5146 impacts were recorded and 1128 were confirmed as true positives (22.0% accuracy). After training the machine learning classifier, the model was able to achieve 83.5% accuracy on this dataset. Associating the CG athletes’ unfiltered SCI exposure to pre- to post-season WM changes revealed no significant associations based on voxel-wise analysis over the whole brain WM network; however, after removing predicted false positives, the filtered SCI exposure revealed significant associations with changes in MD, RD, and FA (all p < 0.05). CONCLUSION: Machine learning classification of sensor-recorded SCI exposure exhibits superior accuracy and sensitivity to threshold-based detection used in standard accelerometry. Accurate quantification of SCI exposure also reveals associations with longitudinal WM changes.