RISK FACTOR IDENTIFICATION FOR ADDUCTOR STRAIN IN A PROFESSIONAL US MLS COHORT THROUGH DESCRIPTIVE ANALYSIS AND PREDICTIVE MACHINE LEARNING MODELS
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Soccer players experience a multitude of injuries, however, one of the most common injuries endured is hip adductor strains. The adductor muscles play a key role in the adduction of the thighs and stabilization of the pelvis. In soccer, athletes experience repeated kicking, change of direction, acceleration, and deceleration movements that can put repeated strain on their adductor musculature. These injuries can cause athletes anywhere from days to months of time away from play and have been found to put athletes at a higher risk of reinjury in the future. Risk factors of adductor injury have been studied in an effort to assess what may be related to increased risk of injury. In an effort to better understand risk factors of these injuries within an MLS team cohort, a descriptive analysis was conducted. Descriptive features such as player position, history of injury, injury type, and mechanism of injury were investigated. Generalized Estimating Equations were utilized to assess how previous injury, position type, and season were related to injury outcomes between the 2016 to 2022 seasons. A season was defined by the calendar year. From this study, we found adductor injury made up 15.5% of all non-contact lower extremity injuries. A total of 30 non-contact adductor injuries occurred and reinjury accounted for 20% of these injuries. Overuse, change of direction, running, and kicking made up 80% of all mechanisms of injury. From the GEE results, previous injury was found to be significantly related to greater odds of injury in a future half season. Position and season were not significantly related to odds of injury. In addition to descriptive qualities, machine learning models and techniques have been deployed to predict an athlete’s risk of injury. A predictive model focusing on adductor strain injuries in a specific professional soccer team cohort in Major League Soccer was constructed to further the understanding of how we can utilize commonly collected assessments for the mitigation of adductor injury. Injury, game, inertial measurement unit/global positioning system, strength, and countermovement jump assessment data was collected over a period of four seasons. Random forest machine learning algorithms were used with chosen features to predict soft tissue, non-contact adductor injuries. Weekly and monthly structured models using IMU/GPS, injury, and game data from 2019 to 2022 were created. Weekly and monthly models using IMU/GPS, injury, game, and strength data from 2021 to 2022 were also created. Finally, weekly and monthly models using IMU/GPS, injury, game, strength, and countermovement jump data from 2022 were created. Synthetic minority sampling and random undersampling were utilized in an attempt to balance the data as injury data points proved to be rare events. Results showed the models were able to predict adductor injury with misclassification of the injured class ranging between 5 to 15%. All models were able to predict injury with differing variables of significance and percent misclassification. Position, history of previous injury, change of direction efforts, acceleration efforts, maximum force production of the ab/adductors and hamstrings, and the adductor-to-abductor ratio were identified as significant variables (p<0.05) in at least one of the four models. These variables give an indication of what metrics are related to soft-tissue, non-contact adductor injury and should be further studied in the identification of risk factors.