Evaluation of risk factors for sport-related concussion and recovery from concussion: a machine learning approach

Abstract

Background: Concussion is a challenging injury to diagnose and treat. There is no gold standard for diagnosis. Despite how heterogeneous the injury is, management guidelines remain the same for all concussions. Targeted or specialized therapy may be an option for concussion treatment; however, when and for whom that intervention may be beneficial is unknown. The annual incidence of sport-related concussion (SRC) is estimated to be around 1-2 million per year. Adolescents and young adults are among the top age groups at greater risk for concussion. Experts agree that a standard method for classifying concussion severity and predictors of outcomes is needed. The overall goal of this project was to improve understanding of the clinical course of concussion through self-reported symptoms prior to and immediately after a concussion injury and time to recovery after sustaining an SRC.Methods: This dissertation addresses the goal of this study using three unique approaches. Chapter 2 used a publicly available prospective cohort from the Concussion Assessment, Research and Education (CARE) Consortium. This cohort consists of collegiate athletes from 30 institutions in the United States. Pre-season baseline concussion assessments were collected on all participants and then participants were followed after incident concussion injury. Concussion assessments after injury until return to sport were completed using a standardized protocol. To test if any baseline symptoms were associated with concussion injury in the dataset, baseline symptom variables reported on the pre-season concussion assessments were selected using a least absolute shrinkage and selection operator (LASSO) regression, a machine learning model. The final selected symptom variables were fit using a multivariable Poisson regression to estimate risk of concussion. Chapter 3 also uses the CARE Consortium cohort; however, the dataset was limited to varsity collegiate athletes who sustained a concussion during the follow-up period. Symptoms reported at the <6hr post-injury Sport Concussion Assessment Tool-3 assessment were categorized into one of 6 concussion sub-types (headache, ocular, vestibular, cognitive, mood, sleep, or multiple sub-types). Cox proportional hazards regression was used to evaluate concussion sub-types associated with time to becoming asymptomatic. In Chapter 4 we compiled a retrospective cohort of patients ages 10-30 years diagnosed with concussion from 2013-2023 in either a primary care clinic, urgent care, radiology, or sports medicine clinic in a large health system. Cohort participants’ electronic health record data was queried to capture demographics, baseline characteristics, characteristics of the injury, and post-injury variables in the 2 years prior to and up to 2 years after their first concussion. Chapter 4 developed a clinical prediction tool through classification and regression tree machine learning. Results: In the CARE Consortium dataset, the overall rate of concussion among college athletes (Chapter 2) was 0.052 concussions/per person year (95% CI 0.05-0.054). The LASSO model selected 14 symptom variables with an area under the curve of 66%. Poisson regression found severe drowsiness (RR 1.54, 95% CI (1.09 - 2.19)), moderate pressure in head (RR 1.36, 95% CI (1.03 - 1.79)), mild blurry vision (RR 1.33, 95% CI (1.06 - 1.68)), and mild nausea/vomiting (RR 1.33, 95% CI (1.04 - 1.71)) to be significantly associated with greater risk of future concussion. Evaluating sub-types in the sub-cohort who sustained concussions (Chapter 3), headache sub-type was the most common sub-type (53%) and least common was mood sub-type (3%). There was no significant association between sub-type and time to becoming asymptomatic. In the retrospective cohort of 3,601 patients (Chapter 4), the median number of visits after a concussion diagnosis was 2 visits (mean 3.88). A decision tree of 6 terminal nodes selected previous number of visits for a mood disorder, diagnosis during the COVID-19 restrictions, diagnosis department, and race as the branching variables. Conclusions: This project examined pre-injury symptoms as risk factors for concussion, immediate post-injury symptom sub-type and survival time to being asymptomatic, and, finally, pre- and post-injury clinical and demographic predictors for concussion injury healthcare utilization. Select baseline symptoms are significantly associated with concussion and may identify athletes at higher risk for future concussion. Addressing the underlying causes of these symptoms when reported on baseline assessments can be a first step in risk reduction. Future work should identify interventions for the baseline symptoms at higher risk for concussion. No sub-types are significantly associated with longer time to asymptomatic. Differentiating concussion sub-types directly after a concussion injury does not appear to identify those at risk of greater symptom resolution. Identifying individuals where targeted treatment may be beneficial in reducing symptom time but should rely on other risk factors and not early sub-type categories. Replication of findings that concussion sub-types are not associated with symptom resolution is needed in other populations. Previous high number of encounters for a mood disorder before a concussion, department of concussion diagnosis, and race are predictive of concussion related health care visits. The prediction model suffered from overfitting and low prediction accuracy. A clinical prediction tool using easy to capture electronic health record data could be valuable in clinical settings, however, more work is needed to improve accuracy of the tool. Concussion remains a complex injury. Despite significant findings from this dissertation, challenges remain in identifying risk factors for concussion and concussion recovery. More objective tools are needed given the lack of gold standard and reliance on self-report for many concussion recovery outcomes.