A synthetic data-driven machine learning approach for athlete performance attenuation prediction.

Introduction: Athlete performance monitoring is effective for optimizing training strategies and preventing injuries. However, applying machine learning (ML) frameworks to this domain remains challenging due to data scarcity limitations. This study extends previous research by evaluating Tabular Variational Autoencoders (TVAE) for generating synthetic data to predict performance attenuation in Gaelic football athletes. Methods: This study assesses synthetic data quality through a comprehensive evaluation framework combining column shape similarity metrics and Hellinger distance analysis, quantifying distributional fidelity across individual variables. Our ML implementation follows a two-phase approach. In the first phase, we evaluated models trained on hybrid datasets with varying synthetic proportions (10%-100%). In the second phase, we examined models trained exclusively on synthetic data and tested them on real data to analyze the utility of the synthetic data. Results: Our results demonstrate that TVAE-generated synthetic data closely replicates original distribution patterns, achieving 85.53% column shape similarity and a Hellinger distance of 0.169. Models trained with additional synthetic data or exclusively on synthetic data outperformed real-data baselines across multiple metrics, particularly for neuromuscular parameters. Our findings emphasize that this approach increased data availability and improved model performance in specific scenarios. Discussion: Several limitations remain: (1) there is limited framework transferability to sports with different physiological demands; (2) the Synthetic Data Generation (SDG) does not currently enforce feature constraints, and future implementations must ensure the procedure respects domain-specific feature limits; and (3) TVAE faced data fidelity challenges with certain variables, such as VO