851 MEP161 – Prediction of impact force during scrums using generative machine learning

AbstractBackground The overall peak impact force that front-row players experience during scrummaging events has been estimated to be on average around 9.8 kN. It follows that it is crucial to monitor such high-magnitude forces to ensure an appropriate players' load management. Objective To use machine learning (ML) techniques to predict contact forces that scrum front-row players are exposed to from videos. Design A generative machine learning model was implemented to predict contact forces during staged scrums. Setting On-field biomechanical analysis of scrummaging events measured across eleven professional teams, for a total of 22 forward packs, in the 2012/13 season. Patients (or Participants) 35 staged scrums were selected from the dataset taken from a previous study that measured three-dimensional coordinates and impact forces of the 6 front-row players involved in the scrum. Assessment of Risk Factors The risk factors that were considered in this study were the peak scrum engagement force and the force trace during the sustained push phase of the scrum. Main Outcome Measurements Six impact forces upon the front-row players during the scrummaging trials were predicted from coordinate data that were retrieved from video footages.Results The error between the generative model predictions and the measured forces was averaged per trial and across the entire dataset. The error in the peak engagement force estimation was 585.77 ± 427.73 N which was a 26.81% ± 19.32% error, whereas the average RMSE during the sustained push phase was 81.60 ± 32.81 N, which was an 8.37% ± 11.24% error.Conclusions The approach of generative ML presented in this study may be considered an important step towards taking force predictions during scrums in live competitions from broadcast videos. In the grand scheme, this framework might be a crucial component in supporting evidence-based interventions for players' load management.