INTRODUCTION: Acute, exercise related injuries and their negative consequences for player health and team performance are an important issue in football. Identifying players or circumstances associated with an increased risk of injury is fundamental for successful risk management. So far, a considerable number of risk factors have been identified. However, univariable predictive accuracy is low and from the few multivariable prediction models none has been scientifically verified outside the original setting. Of note, time-constant (screening) and volatile (monitoring) risk factors are so far generally considered separately resulting in a restricted set of explanatory variables. Consequently, improvements in predictive accuracy may be expected when screening- and monitoring-data are combined, especially when analysed with current machine learning (ML) techniques able to account for interactions within a "web of determinants". Expectably, screening parameters will set a personal level for injury risk while monitoring parameters account for short term variations around it. METHODS: This trial was designed as a prospective observational cohort study aiming to predict non- contact time-loss injuries in male professional soccer on a daily basis. Considering the expectable number of injuries, a panel of 12 explanatory variables (covering basic player characteristics, screening, monitoring and exposure characteristics) had been specified a priori. Injuries were registered according to the Fuller consensus. Gradient boosting was used for data analysis. Respecting the nesting of timepoints within players, cross-validation was performed on the level of players (not datapoints). Upsampling was implemented with the training set to account for the imbalance between days with and without injury occurrence, respectively. Different splits of the original dataset were used to probe the robustness of results. RESULTS: Data of 88 players from 4 teams (German 3rd and 4th league) could be analysed. 51 non-contact, time-loss injuries could be included in the final analysis. The cross-validated (ROC area under the curve 0.66) and test set performance (ROC area under the curve 0.64) of the gradient boosted model in forecasting injury occurrence was promising and superior to comparator models without integration of screening features. However, the variation of predictive accuracy and feature importance with different splits of the original dataset reflects the relatively low number of events. CONCLUSION: It is concluded that injury prediction based on the integration of screening and monitoring data in combination with ML is promising. However, external prospective verification and continued model development with accumulating numbers of injuries are required before application in sports practice may be envisaged.
© Copyright 2022 27th Annual Congress of the European College of Sport Science (ECSS), Sevilla, 30. Aug - 2. Sep 2022. Veröffentlicht von Faculty of Sport Science - Universidad Pablo de Olavide. Alle Rechte vorbehalten.