Developing a multivariate model for the prediction of concussion recovery in sportspeople: a machine learning approach
Louise C. Yates, Elliot Yates, Xuanxuan Li, Yi‐Ping Lu, Kamal Makram Yakoub, David Davies, Antonio Belli, Vijay Sawlani
Abstract
Background: Sportspeople suffering from mild traumatic brain injury (mTBI) who return prematurely to sport are at an increased risk of delayed recovery, repeat concussion events and, in the longer-term, the development of chronic traumatic encephalopathy. Therefore, determining the appropriate recovery time, without unnecessarily delaying return to sport, is paramount at a professional/semi-professional level, yet notoriously difficult to predict. Objectives: To use machine learning to develop a multivariate model for the prediction of concussion recovery in sportspeople. Methods: Demographics, injury history, Sport Concussion Assessment Tool fifth edition questionnaire and MRI head reports were collected for sportspeople who suffered mTBI and were referred to a tertiary university hospital in the West Midlands over 3 years. Random forest (RF) machine learning algorithms were trained and tuned on a 90% outcome-balanced corpus subset, with subsequent validation testing on the previously unseen 10% subset for binary prediction of greater than five missed sporting games. Confusion matrices and receiver operator curves were used to determine model discrimination. Results: 375 sportspeople were included. A final composite model accuracy of 94.6% based on the unseen testing subset was obtained, yielding a sensitivity of 100% and specificity of 93.8% with a positive predictive value of 71.4% and a negative predictive value of 100%. The area under the curve was 96.3%. Discussion: In this large single-centre cohort study, a composite RF machine learning algorithm demonstrated high performance in predicting sporting games missed post-mTBI injury. Validation of this novel model on larger external datasets is therefore warranted. Trial registration number: ISRCTN16974791.