Modelos de inteligência artificial para predição de indicação e do risco de lesões em atletas de futebol feminino

Abstract

Physical training plays a key role in enhancing athletes’ performance in the pursuit of high-level results; however, this context can trigger adverse effects that harm both these professionals and sports organizations. Sports injuries emerge from a complex network of factors, however, their occurrence is strongly associated with inadequate training load volumes. The aim of this work is to predict the indication and risk of injuries in professional female soccer athletes, where the injury indication is a binary value (0 or 1) and the risk is a value between 0 and 1. In this study, 11,612 daily records from athletes of two professional female soccer clubs in Norway were used, collected via the Player Monitoring System (pmSys) application during the 2020 and 2021 seasons. The database used includes variables related to training load, well-being, and injury occurrence. Training load variables were collected using the Rating of Perceived Exertion (RPE) scale, a tool that quantifies exercise intensity based on the athlete’s perception. In the dataset, injuries were recorded on the day they occurred, serving as the model’s output variable. Injury indication prediction models were developed using Decision Tree and Random Forest machine learning methods. The metrics used to evaluate model performance were sensitivity (success rate of positive predictions) and accuracy (success rate of all predictions). Random Forest outperformed Decision Tree, presenting a sensitivity rate of 64.52% and an accuracy rate of 98.71%. To predict injury risk, neuro-fuzzy methods were applied: the Dynamic Evolving Neural Fuzzy Inference System (DENFIS) and the Subtractive Clustering Method with Fuzzy C-Means (SBC/FCM). Each model employs four training and well-being–related input variables to generate a Fuzzy Rule-Based System (FRBS) aimed at predicting the athlete’s injury risk. In the test set, the DENFIS model presented a sensitivity rate of 75% and an accuracy rate of 85.38%, while the SBC/FCM model obtained a sensitivity of 75% and an accuracy of 91.70%. Using k-fold cross-validation with k=5, the DENFIS model achieved higher percentages across the folds in the sensitivity metric. In the injury risk models, to address the imbalance of the “injury” attribute, since its occurrence is generally of low frequency, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to generate synthetic data for the minority class (injury). This study also developed a computational interface using R software that calculates the user’s injury risk based on physical training and well-being data, utilizing the model obtained by DENFIS.