Predicting daily recovery during long-term endurance training using machine learning analysis

Abstract

Purpose: The aim of this study was to determine if machine learning models could predict the perceived morning recovery status (AM PRS), training feeling during exercise (exercise TF), and daily change in heart rate variability (HRV change) of endurance athletes based on training, dietary intake, sleep, HRV, and subjective wellbeing measures.

Methods: Self-selected nutrition intake, exercise training, sleep habits, HRV, and subjective wellbeing of 40 endurance athletes was monitored daily for 12 weeks (3,325 days of tracking). Global and individualized models were constructed using nine machine learning techniques and combined into an ensemble model at the group level, and with a single best algorithm chosen for individualized models. Model performance was compared with a baseline intercept-only model.

Results: Prediction error (root mean square error [RMSE]) was lower than baseline for the group models (12.1 vs. 17.5, 13.1 vs. 14.7, and 0.25 vs. 0.30 for AM PRS, exercise TF, and HRV change, respectively). At the individual level prediction accuracy outperformed the baseline model but varied greatly across participants (RMSE range 5.5 to 23.6, 5.7 to 18.2, and 0.05 to 0.52 for AM PRS, exercise TF, and HRV change, respectively).

Conclusion: Daily recovery measures can be predicted based on commonly measured variables, with a small subset of variables providing most of the predictive power. However, at the individual level the key variables may vary, and additional data may be needed to improve prediction accuracy.