Machine Learning for Classification of Pediatric Concussion Recovery Stages

Abstract

Mild traumatic brain injury (mTBI), or concussion, results from sudden acceleration or deceleration of the brain and subsequent complex tissue propagation of shock waves that disrupt structure and function. Concussions can cause many symptoms including headache, dizziness, and difficulty concentrating. These can be detrimental to children, a ecting their participation in school, sport, and social activities. Therefore, return to school (RTS) and return to activity (RTA) protocols have been developed to help safely return children to these activities without risking further injury. The goal of this study was to develop machine learning (ML) algorithms to predict RTA and RTS stages, that can easily be incorporated into a smartphone application (APP). Ideally this would assist children in tracking and determining their RTA and RTS progression leading them to a safe and timely return. Support vector machine classi er (SVC) and random forest (RF) algorithms were developed to predict RTA/RTS stages. Both were modeled on previously acquired data, and on newly acquired data, and results were compared. Models were trained and tested using accelerometry and symptom data from pediatric concussion patients. A sliding window technique and feature extraction were performed on raw acceleration data to extract suitable features, which were combined with yes/no symptom recordings as ML inputs. The dataset consisted of 67 participants aged 10 to 18, 42 female and 25 male, with a total of 844408 samples. The best results for RTS prediction showed average accuracy of 83% for RF and 66% for SVC. For RTA predictions, the best results had average accuracy of 60% for RF and 58% for SVC. For new data, RTS predictions showed an accuracy of 45% for RF and 41% for SVC. RTA predictions had an accuracy of 35% for RF and 30% for SVC. RF models had superior performance on all data. These results show that predicting RTA/RTS is possible with ML. However, improvements to these models can be made by training on more data prior to APP implementation. More data is needed, as recruitment during this study was limited due to Covid-19 restrictions.