A Machine Learning Approach to Concussion Risk Estimation Among Players Exhibiting Visible Signs in Professional
Jared M Bruce1,2,3, Kaitlin E Riegler4, Willem Meeuwisse5
1Department of Biomedical and Health Informatics, University of Missouri-Kansas City School of Medicine, Kansas City, MO, 64108, USA. brucejm@umkc.edu.
Background:
The identification of concussion risk factors, such as visible signs and mechanisms of injury, improves concussion identification. Exploring individual risk factors, such as concussion history, may help to improve existing concussion risk models and algorithms.
Objectives:
The primary aim of the current study was to use machine learning techniques to develop a comprehensive, prospectively coded concussion risk model in professional hockey among players exhibiting visible signs. The secondary aim was to examine whether including concussion history improves model performance.
Methods:
Data from the National Hockey League (NHL) spotter program, including coded visible signs and mechanisms of injury associated with possible concussive events, were extracted from the 2018-2019 to the 2021-2022 seasons. Each unique spotter event was matched with data extracted from the medical record to determine whether the event was associated with a subsequent physician diagnosed concussion. We compared the ability of three machine learning-based approaches to identify the likelihood of physician diagnosed concussion: conditional inference tree, conditional inference random forest, and logistic regression.
Results:
A total of 1563 unique events with visible signs were identified by spotters (183 leading to a concussion diagnosis). A randomly selected training sample had 1250 events (146 concussions) and the remaining set-aside test sample had 313 events (37 concussions). The obtained models performed at a high level with large effects in the training [area under the receiver operating characteristic curve (AUC) = 0.79] and set-aside test data (AUC = 0.82). Concussion history was retained in the tree and logistic regression models, with each additional prior concussion associated with a 1.32 times increased odds of concussion diagnosis.
Conclusions:
We present simple tree and logistic algorithms for concussion screening and as diagnostic aids. Our results show that player concussion history can explain additional risk above and beyond that explained by visible signs and mechanisms of injury alone.
