A machine learning approach to concussive group classification using discrete outcome measures from a low-cost movement-based assessment system

Measurable neuromotor control deficits during functional task performance could provide objective criteria to aid in concussion diagnosis. However, many tools which measure these constructs are unidimensional and not clinically feasible. The purpose of this study was to assess the classification accuracy of a machine learning model using features measured by a clinically feasible movement-based assessment system (Mizzou Point-of-care Assessment System (MPASS) between athletes with and without concussion. Forty collegiate athletes participated. Twenty (19.40 ± 1.04 yrs., 11 females) suffered concussion within two weeks of data collection (5.40 ± 3.68 days). Twenty (19.85 ± 1.20 yrs.) sex, sport, and position-matched athletes had no concussions in the past year. All participants completed three 30-second static balance trials with eyes closed on foam surface under both single task and cognitive dual task conditions, four trials of gait under normal, head shaking, and dual task conditions, and reaction time tasks. Kinematics, kinetics, and reaction times were recorded by MPASS. Measures were used as features for a XGBoost machine learning model. Five-fold cross-validation yielded mean (across 5-folds): 82.5 % accuracy, 75 % sensitivity, 90 % specificity, 88.2 % positive predictive value, and 78.3 % negative predictive value. Results indicate promise for using movement-based features from a low-cost movement-based assessment system to improve the objectivity of concussion diagnosis decision-making.