Using machine learning to discover traumatic brain injury patient phenotypes: national concussion surveillance system Pilot

Abstract

ABSTRACT Objective The objective is to determine whether unsupervised machine learning identifies traumatic brain injury (TBI) phenotypes with unique clinical profiles. Methods Pilot self-reported survey data of over 10,000 adults were collected from the Centers for Disease Control and Prevention (CDC)’s National Concussion Surveillance System (NCSS). Respondents who self-reported a head injury in the past 12 months (n = 1,364) were retained and queried for injury, outcome, and clinical characteristics. An unsupervised machine learning algorithm, partitioning around medoids (PAM), that employed Gower’s dissimilarity matrix, was used to conduct a cluster analysis. Results PAM grouped respondents into five TBI clusters (phenotypes A-E). Phenotype C represented more clinically severe TBIs with a higher prevalence of symptoms and association with worse outcomes. When compared to individuals in Phenotype A, a group with few TBI-related symptoms, individuals in Phenotype C were more likely to undergo medical evaluation (odds ratio [OR] = 9.8, 95% confidence interval[CI] = 5.8–16.6), have symptoms that were not currently resolved or resolved in 8+ days (OR = 10.6, 95%CI = 6.2–18.1), and more likely to report at least moderate impact on social (OR = 54.7, 95%CI = 22.4–133.4) and work (OR = 25.4, 95%CI = 11.2–57.2) functioning. Conclusion Machine learning can be used to classify patients into unique TBI phenotypes. Further research might examine the utility of such classifications in supporting clinical diagnosis and patient recovery for this complex health condition.