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

Area of Science:

• Sports Science
• Data Science
• Machine Learning

Background:

• Athlete performance monitoring is crucial for optimizing training and preventing injuries.
• Data scarcity poses a significant challenge for applying machine learning in sports science.

Purpose of the Study:

• To evaluate Tabular Variational Autoencoders (TVAE) for generating synthetic data to predict performance attenuation in Gaelic football athletes.
• To assess the quality and utility of synthetic data for athlete performance prediction.

Main Methods:

• A two-phase machine learning approach was used, evaluating models trained on hybrid and exclusively synthetic datasets.
• Synthetic data quality was assessed using column shape similarity and Hellinger distance analysis.

Main Results:

• TVAE-generated synthetic data closely replicated original data distributions (85.53% column shape similarity, 0.169 Hellinger distance).
• Models trained with synthetic data outperformed real-data baselines, especially for neuromuscular parameters.
• The approach increased data availability and improved model performance in specific scenarios.

Conclusions:

• Synthetic data generated by TVAE is effective for predicting performance attenuation in Gaelic football.
• This method addresses data scarcity and enhances athlete monitoring across various metrics.
• The findings open avenues for using synthetic data in sports performance analysis.