Predictive model to prevent injuries in athletes using Machine Learning and sensitivity analysis

Abstract

This study presents a predictive system based on Machine Learning and Deep Learning techniques aimed at supporting early prevention of sports injuries in athletes through anthropometric and morphological indicators. The CRISP-DM methodology was adopted to structure the analytical process, incorporating data preprocessing steps such as cleaning, Min–Max normalization, ordinal encoding, class balancing, and stratified train–test splitting. Six predictive models were developed and evaluated: Random Forest, XGBoost, Multilayer Perceptron (MLP), Support Vector Machine, K-Nearest Neighbors, and a Deep Learning model implemented in Keras. In addition, an Ensemble Voting approach was integrated to improve robustness by combining model predictions through majority voting. Model performance was assessed using accuracy, precision, recall, F1-score, and Precision–Recall curves to address class imbalance. The results indicate that the MLP model achieved the highest overall performance, reaching an accuracy of 87.5%, followed by Random Forest and Keras models with competitive results. Finally, the proposed system was deployed as a web-based application that allows sports professionals to input athlete data and obtain real-time injury risk predictions, facilitating preventive decision-making in load management and reducing the likelihood of muscle injuries.