Using Machine Learning to Predict Concussion Recovery Time: The Importance of Psychological and Symptomatic Factors

Research Objectives: To (i) utilize machine learning (ML) to create a model for predicting concussion recovery time among athletes, (ii) compare predictive factors in a ML model to previously identified risk-factors.

Design: Binary and multiclass classification ML models applied to prospective data collected during an initial clinic visit and 3-month follow-up (North Texas Concussion Registry; ConTex).

Setting: Specialty concussion clinics.

Participants: 1296 ConTex participants ages 6 to 64 with sports-related concussion.

Interventions: Not applicable.

Main Outcome Measures: Self-reported pre-morbid depression (PHQ-9), anxiety (GAD-7), quality of sleep (PSQI), and concussion symptoms (SCAT-5) were collected at initial visit. A model’s classification ability was evaluated using Area Under the Curve (AUC). Feature predictive importance was measured using Leave One Feature Out metrics and Permutation Feature Importance. Early recovery was defined as less than or equal to 14 days, and late recovery was greater than 14 days.

Results: A CatBoost ML model with 5-fold cross validation and an 80-20 train-test split classified participants into early vs. late recovery with an AUC of 0.80 on testing data. Factors related to time to clinic, depression, anxiety, sleep quality, and symptom severity have the largest feature importance values in the predictive model. Loss of consciousness, post-injury headache, and other variables previously significantly associated with recovery time were not found to have significant predictive power in our model. In contrast to the binomial model, multinomial classification of participants' recovery time has low AUC (0.69).

Conclusions: Our findings suggest that ML may assist in estimating athletes at risk for prolonged recovery. Initial mood, sleep, and concussion symptoms had stronger influence in ML classification than injury-related variables previously identified by significance testing, which underscores differences in prediction vs. association. Results highlight the need for multifaceted player health assessments and further research into treatments for athlete mental health and sleep to potentially improve concussion recovery time.

Author(s) Disclosures: The authors declare no competing interests.