Cortical Grey-White Contrast Distinguishes Collegiate Athletes Based On Sport And Concussion Status: A Machine Learning Approach To Staging Athlete Brain Health

Injury models suggest that the boundary between brain grey and white matter is particularly sensitive to repetitive blunt trauma, like the impacts sustained in contact and collision (C&C) sports that sometimes result in a diagnosed concussion. PURPOSE: Borrowing from studies designed to stage (un)healthy brain aging, we employed supervised machine learning to learn features that distinguish patterns of a grey-white matter boundary sharpness contrast (BSC) between healthy athletes engaged in C&C sports, athletes engaged in non-C&C sports, and C&C athletes 24-48 hours after a sport-related concussion (SRC). METHODS: High-resolution structural brain imaging was retrieved from FITBIR for 340 collegiate athletes recruited through two studies: 116 athletes were scanned 24-48 hours after a diagnosed sport-related concussion (post-SRC) and 222 athletes (118 C&C athletes) served as controls. Individual pial and white matter surfaces were reconstructed (Freesurfer 7.1.0). Vertex-wise BSC was computed by fitting a sigmoidal curve to the normalized T1w intensities sampled at 8 points across the grey-white boundary at 59 k vertices. Given that BSC is known to covary with curvature, BSC was regressed against vertex-wise curvature and participant age before being averaged between 200 cortical parcels comprising the Schaefer atlas. We sought to (1) distinguish healthy non-C&C athletes from healthy C&C athletes (2) distinguish non-C&C from post-SRC athletes, and (3) distinguish healthy C&C athletes from post-SRC athletes. A boosted classifier (Python, 3.10.4) and hyperparameter tuning was performed using Bayesian search with 5-fold cross-validation. Classifier performance was determined by precision, recall, and F1-score on the hold-out data. Feature importance was determined using Shapley values. RESULTS: Cortical BSC distinguished non-C&C from post-SRC athletes (precision = 0.60, recall =0.71, F1 = 0.65) and from C&C athletes (precision = 0.59, recall =0.74, F1 = 0.65). However, BSC did not distinguish between C&C and post-SRC athletes (precision = 0.41, recall =0.52, F1 = 0.46). Areas driving this effect were predominantly found in unimodal (somatomotor, visual) cortical regions. CONCLUSIONS: BSC may be a biomarker of brain health among collegiate athletes engaged in contact and collision sports.