The risk of injury in the first five years of an Australian football career – Can it be predicted without using player load?

Abstract

Background: The concept of predicting injury risk has been criticised (Bahr, 2016), however research using machine learning methods has demonstrated that it is possible to predict injury risk in basketball (Talukder 2016a) and soccer (Talukder 2016b). Player load is by far the strongest injury risk factor (Rossi 2018), however player load data may not always be available for analysis and the prediction of risk. The aim of this study was to investigate whether player characteristics alone, could be used to predict injury risk in the first five seasons of an Australian football player’s professional career. Methods: The data represented 1033 unique AFL players, who sustained 4762 injuries in seasons 1997-2016. Player characteristics (height, body mass, age, playing position, indigenousness & natural kicking foot) and the injury characteristics of the previous season, were used in models to predict injury risk. Results: A Naïve Bayes model that was based on player characteristics alone was able to classify injury risk category (low, medium, high) over 5 years, 1.3 times better than random chance. The characteristics of; playing position, height and body mass contributed to 97% of the prediction of injury risk category. A Decision Tree model that classified injury risk category in a single season, based on the characteristics of a player and their injuries in the previous season, performed 1.4 times better than random chance. When the models were used to predict whether a player would be in the highest risk category, the performance of both models increased to 1.9 and 2.2 times better than random chance. Taller and heavier key position players who sustained hamstring strain and groin strain/osteitis pubis injuries in the previous season had a higher injury risk than shorter and lighter non-key position players who remained injury-free in the previous season. Discussion: The injury prediction models reported here did not perform as well previously reported models that were based on training load. Nevertheless, they demonstrate how player characteristics affect injury risk and the models themselves can be used “live” to stratify injury risk, which can then be used to help clinicians make decisions about risk mitigation strategies such as prehabilitation. Conflict of interest statement: My co-authors and I acknowledge that we have no conflict of interest of relevance to the submission of this abstract.