Machine learning to predict sports-related concussion recovery using clinical data

Area of Science:

• Sports Medicine
• Neurology
• Data Science

Background:

• Sport-related concussions (SRCs) pose significant risks to high school athletes.
• Predicting SRC recovery time is crucial for effective clinical management but challenged by complex data.
• Traditional methods struggle with the multifaceted nature of concussion assessment data.

Purpose of the Study:

• To evaluate the utility of machine-learning (ML) algorithms in predicting SRC recovery time.
• To assess the ability of ML models to predict protracted recovery (recovery > 21 days).
• To compare ML model performance against human-driven models for concussion recovery prediction.

Main Methods:

• Retrospective case series of 655 athletes (aged 8-18) diagnosed with SRC.
• Utilized pre-injury risk factors, injury severity, and post-concussion symptoms including Vestibular Ocular Motor Screening (VOMS), King-Devick Test, and C3 Logix Trails Test data.
• Trained and evaluated multiple sex-stratified ML models, comparing their predictive accuracy (AUC) for protracted recovery.

Main Results:

• Gradient boosting decision-tree algorithms showed the best performance in predicting SRC recovery time and protracted recovery for both males and females.
• Model accuracy improved significantly when VOMS data were combined with King-Devick and C3 Logix data.
• The ML models achieved higher AUC values (0.84/0.78 for males/females) compared to statistical models (0.74/0.73) for predicting protracted recovery.

Conclusions:

• Machine-learning models successfully managed the complexity of vestibular-ocular motor system data in SRC assessment.
• ML models demonstrate significant clinical utility for informing prognostic evaluations of SRC recovery.
• These findings support the integration of ML into clinical practice for predicting athlete recovery timelines.