Using machine learning to predict concussion recovery time: The importance of psychological and symptomatic factors.

Abstract

Objectives: The objectives were threefold: 1) To utilize machine learning (ML) to create a model for predicting concussion recovery time using routine clinical metrics, 2) To compare predictive factors within a ML model to previously identified risk factors, and 3) To compare predictive ability of ML models to traditional logistic regression.

Methods: North Texas Concussion Registry (ConTex) data were prospectively collected during an initial post-injury clinic visit and 3-month follow-up. ML models classified 1000 participants with sport- or recreation-related injuries, ages 6-59, into ordinal recovery time groups. Models were trained on an 80-20 train-test split with 5-fold cross validation. Performance was evaluated using area under the curve (AUC). Feature predictive importance was measured using Leave One Feature Out (LOFO) metrics and Permutation Feature Importance (PFI).

Results: A CatBoost binary ML model classified participants into ≤14-d or >14-d recovery with an AUC of 0.79, similar to the logistic regression AUC of 0.77. In contrast, the multiclass model for recovery time had a lower AUC of 0.69. Time to clinic, symptom severity, and factors related to self-reported depressive symptoms, anxiety, and sleep quality had the largest feature importance values in the CatBoost model.

Conclusions: Post-injury depressive symptoms, anxiety, and sleep had a stronger influence in predicting prolonged recovery time than previously identified injury-related variables (e.g. loss of consciousness, headache). While promising, ML may not outperform traditional models depending on the simplicity and linearity of the predictor variables.