Using Eye And Head Movements During The VOMS In VR To Predict Concussion Diagnosis
Nicholas G. Murray, Brian J. Szekely, Madison T. Fenner, Kristen G. Quigley, Addie Jane Heithecker, Zackary C. Buck, Christian Rojas, Stella K. Thornton, Kimia Faghfoori, Julia Shray, Julia Shray, Philip K. Pavilionis, Nora L. Constantino · Medicine & Science in Sports & Exercise · 2025
Virtual reality (VR) systems can be used to track the eye and head movements during the Vestibular/Ocular Motor Screening (VOMS) assessment. These data, independent of symptoms, may provide insight into concussion diagnosis. PURPOSE: Use machine learning (ML) to predict concussion diagnosis based on eye and head movements recorded during the VOMS in VR. METHODS: 279 Division I athletes with sport-related concussion (SRC) (female = 120, male = 159; average age = 20 ± 1 years) along with 279 randomly sampled controls (female = 184, male = 95; average age = 20 ± 1 years) participated in this study. All controls were evaluated at pre-season baseline testing and SRC were within 3 to 12 days (avg. = 3.5 ± 3.12 days) of their injury. All SRC were diagnosed by a team physician and confirmed using a multifaceted test battery. The VOMS stimuli were presented using a VR head-mounted display with integrated eye tracking capabilities (HTC Vive Pro Eye HMD: 90 Hz; 2880x1600 pixels) developed using Unity3D engine (v2019.1.6). Participants reported symptoms of headache, dizziness, nausea, and fogginess before the test and after each of the seven subtests: smooth pursuit, horizontal/vertical saccades, NPC, horizontal/vertical vestibulo-ocular reflex, and visual motion sensitivity. The eye and head movements were tracked and extracted during each of the subtests. An Extreme Gradient Boost (XGB) classifier was tuned and trained using a 70/30 training/test split. Hyperparameters were tuned on the training data with 5-fold cross validation. The model predicted SRC probabilities, with the performance assessed using the Receiver Operator Characteristic curve, sensitivity, and specificity. RESULTS: Overall accuracy of the model to predict SRC was 0.85 (Area Under the Curve = 0.88) with sensitivity = 0.87 and specificity = 0.61. Among the features, horizontal VOR explained 35% of the variance while vertical saccade duration explained 15% in determining SRC diagnosis. CONCLUSION: Eye and head movements, excluding symptom provocation data, predicted concussion diagnosis with an accuracy of 85%. This is an important first step in using eye and head movements during the VOMS to determine SRC diagnosis rather than relying on subjective symptom reporting. Supported by: P30GM145646