Prediction models for musculoskeletal injuries in professional sporting activities: A systematic review

Abstract

The purpose of this systematic review was twofold: (a) identify prediction models for musculoskeletal injuries during the participation of professional sporting activities and (b) evaluate these models by their predictive performances. A systematic review of the PubMed and Embase databases was performed using specific search terms selected according to the PRISMA guidelines. Ten studies met the eligibility criteria and were included. The most commonly employed data component for data pre-processing was body composition data (most commonly, body mass and height), followed by player profile (most commonly, age followed by position). The most common machine learning technique for data processing was the decision tree, followed by logistic regression. The median AUC of the best performing models indicated per study was 0.75 (0.16), median sensitivity/recall was 0.78 (0.15), median specificity was 0.81 (0.27), and median precision was 0.53 (0.13). The performance of prediction models in the literature has been poor, caused by a fundamental difficulty in discovering real effects in small sample sizes with low injury rates. A better understanding of how training and game exposure is associated with the data components for data pre-processing and ultimately associated with injury is vital for the future development of robust injury prediction models.