Prediction of running injuries from training load: a machine learning approach

The prediction of the running injuries based on selfreported training data on load is difficult. At present, coaches and researchers have no validated system to predict if a runner has an increased risk of injuries. We aim to develop an algorithm to predict the increase of the risk of a runner to sustain an injury. As a first step Self-reported data on training parameters and injuries from high-level runners (duration=37 weeks, n=23, male=16, female=7) were used to identify the most predictive variables for injuries, and train a machine learning tree algorithm to predict an injury. The model was validated by splitting the data in training and a test set. The 10 most important variables were identified from 85 possible variables using the Random Forest algorithm. To predict at an earliest stage, so the runner or the coach is able to intervene, the variables were classified by time to build tree algorithms up to 7 weeks before the occurrence of an injury. By building machine learning algorithms using existing self-reported training data can enable prospective identification of high-level runners who are likely to develop an injury. Only the established prediction model needs to be verified as correct.