An Ensemble of Long Short-Term Memory Models to Automatically Detect End-Range Movement Patterns in Men's Professional Hard Court Grand Slam Tennis.

Evaluating end-range movements during tennis match-play can quantify high intensity load exposure and facilitate specific analysis of players' high-end physical capabilities. Currently, the process to evaluate such movement is labour intensive, with an established and efficient process to identify end-range movements lacking. Using three-dimensional pose model data for male competitors in the 2024 Australian Open, we evaluated an ensemble of long short-term memory (LSTM) models to correctly classify coach-identified end-range movement patterns. An ensemble of 10 LSTM models that took the average prediction value and applied a class prediction threshold of 0.63 was the best performing approach, producing an F1-score of 0.944, overall accuracy of 95.9%, precision of 97.8% and recall of 91.2%. From these results, we provide a novel and practical way of using real-world pose model data and machine learning to automatically detect one of the most physically demanding movement tasks in professional men's tennis. This work enhances post-match analysis via an automated analytical pipeline that can quantify high intensity movement exposures and produce descriptive statistics of end-range movement to assist with the load monitoring and management of professional players.