An Explainable Machine Learning Approach to Explain the Effects of Training and Match Load on Ultra-Short-Term Heart Rate Variability in Semi-Professional Basketball Players

Understanding how training and match load influence autonomic recovery is essential for optimizing athlete monitoring. This proof-of-concept study aimed to examine the impact of training and match load on next-day heart rate variability (HRV) and to explain how different load measures influenced the internal response, using SHapley Additive Explanations (SHAP) to interpret machine learning models. Five semi-professional basketball players (23 ± 5 years; 191 ± 7 cm; 90 ± 11 kg) were monitored throughout a competitive season. HRV and load metrics were recorded daily. Differences in the natural logarithm of the root mean square of successive differences (LnRMSSD) across Non-Training, Training, and Match days were analyzed using linear mixed models. Additionally, a Gradient Boosting Machine model was developed to examine next-day HRV responses, with SHAP analysis providing both global and individual insights into feature importance. Next-morning LnRMSSD values were significantly lower on Match days compared to both Training and Non-Training days (<i>p</i> < 0.001). SHAP results identified rate of perceived exertion (RPE), days since last match, minutes played, and recent training load as the most influential variables associated with HRV changes. Pre-session heart rate and the root mean square of successive differences (RMSSD) values also demonstrated notable individual relevance. The ranking and magnitude of influential variables varied across players, highlighting the heterogeneity of physiological responses in team sports. While these findings are specific to this cohort, they illustrate the potential of explainable machine learning to enhance transparency and support individualized monitoring strategies. Importantly, they underscore the value of integrating both subjective and objective load measures to inform training decisions. Future research involving larger, multi-team samples is needed to validate the generalizability of these results.