Summary

"This study developed a machine learning model to predict lower limb sports injuries. 120 athletes underwent screening evaluations, and the K Nearest Neighbour algorithm with Random Oversampling showed the best results. The model can help identify at-risk athletes and inform injury prevention programs."

Abstract

Introduction

The burden of sports-related injuries is very significant due to the work years lost due to these injuries in the youth. Primary prevention of these injuries is an important public health goal in recent years. The development of an effective prevention model is difficult because of the complexity of the factors that influence these injuries. Preventive approaches should take into consideration, all the risk factors and help us identify at-risk individuals. Artificial intelligence with advanced machine learning tools can be applied for this purpose. This study aims to build the best-performing injury prediction model for lower limb sports injuries. Methods: 120 male university-level football, cricket, and basketball players were recruited for the study from two different universities. They underwent an elaborate screening evaluation which included personal (anthropometric measurements, playing surface), psychological (sleep deprivation and athlete burnout scores), and neuromuscular measures (range of motion, muscular strength of lower limb, balance and core stability measurement) accounting for 44 variables. Their previous season's injury details were recorded in detail. Using the data available, different predictive machine learning models were tried to identify the at-risk individuals for injury. Predictive ability of several models built by applying a range of learning techniques was analysed and compared. Results: 38 injuries were recorded, 20 (52.6%) of which corresponded to ankle injuries, 9 to hamstring injuries (23.6%), four to knee injuries (10.5%), and 5 to foot injuries (13.3%). 8 injuries occurred during training and 30 during competition. 6 players were injured twice, so the first injury was used leaving 32 injuries that were used to develop the predictive models. The model generated by the K Nearest Neighbour algorithm (KNN) with a Random Oversampling (ROS) as oversampling technique reported the best evaluation criteria (area under the receiver operating characteristic curve score = 0.87, true positive rate = 100%, true negative rate = 62.5%) and hence was considered the best for predicting injuries. Conclusions: The machine learning model - KNN with ROS showed good accuracy for identifying athletes prone to injuries. The identified athletes can be rehabilitated to prevent future injuries. This model would be helpful in injury prevention programs. Larger studies are needed to validate the model.