580 MEP069 – Injury risk estimation using machine learning based on athletes’ physical and mental states perception monitoring: preliminary study on 93 high-level athletes during 18 months

AbstractBackground Machine learning is an artificial intelligence technique, consisting of learning from data and making predictions (such as classifications), which could provide access to an injury risk estimation for an athlete. While such approach may offer powerful decision-making processes to stakeholders, it lacks of proof of concepts applied to risk estimation in elite sport performance.Objective To estimate the injury risk of track and field athletes practicing sports involving sprints.Design Prospective cohort study over 532 days.Setting High-level athletes.Participants 93 elite athletes practicing athletics disciplines including running sprints (sprints, hurdles, jumps, combined events).Main Outcomes Measures The risk factors measured daily (input data of the model) were the perceptions of the physical and psychological states of the athletes, self-reported each morning and evening in a dedicated application. The output data corresponded to the injuries identified by the athletes. A decision tree model was trained with a splitting strategy of 70–30%, then a cross validation was performed with an hyper-parameter optimization.Results Model performance was evaluated on the test dataset (30%) that achieved the following performance scores: accuracy=0.86 (0.02); precision=0.06 (0.02); recall=0.16 (0.06); specificity=0.89 (0.02); ROC AUC= 0.52 (0.03). Conclusions The present machine learning approach allowed to estimate the injury risk of athletes. The performance of the model to predict injury was low mainly because of the imbalance between the number of injuries and the number of observations. The interpretation of the prediction model showed that the most influential factors were physical factors and positive emotions.