Harnessing advanced AI, researchers uncover the complexities of athlete recovery, revealing that one-size-fits-all approaches fall short in optimizing performance.
Study: Predicting daily recovery during long-term endurance training using machine learning analysis
In a recent research paper published in the European Journal of Applied Physiology, researchers evaluated whether machine learning (ML) could predict morning perceived recovery status (AM PRS) and heart rate variability changes (HRV change) in endurance athletes using training, diet, sleep, HRV, and well-being data. The study tracked and analyzed data from 43 athletes over 12 weeks, comparing global and individualized models. The ML models showed lower prediction errors than the baseline, with significant variation in accuracy across individual participants.
Background
Past work has shown that coaches and athletes monitor training load, HRV, sleep, diet, and subjective well-being to understand training responses. While internal training load, reflecting physiological strain, plays a crucial role in training outcomes, interactions between training load, sleep, and diet remain underexplored, especially in endurance sports. The relationship between self-selected nutrition intake and daily recovery during endurance training is not well characterized, though increased energy and carbohydrate intake during intense training can reduce overreaching symptoms.
Study Design Overview
This observational study tracked 55 endurance athletes (final analysis included 43 athletes) over 12 weeks to monitor daily nutrition, exercise, sleep, HRV, and subjective well-being. Participants were excluded if they did not log at least 85% of the required data points or trained an average of less than 6 hours per week. Participants were allowed to follow their usual routines, recording data on metrics like AM PRS and HRV change. The study was open to male and female athletes aged 18 or older who trained at least 7 hours per week and were already using smartphone apps to track diet, HRV, and sleep. The researchers focused on understanding individual and group-level responses by analyzing participants' data using models to predict AM PRS and HRV change.
The team used data from wearable devices and tracking apps to assess participants' exercise, dietary intake, sleep, and subjective well-being. Exercise sessions were recorded in Training Peaks, with details on modality, duration, and perceived exertion using the Borg CR100® scale, which offers additional precision compared to the commonly used CR10 scale.
Researchers tracked dietary intake through apps like MyFitnessPal, removing incomplete tracking days from the analysis to maintain data integrity. Resting HRV and sleep were monitored using consumer-grade devices like Oura rings and Whoop straps, which provided daily measures. Subjective well-being was assessed each morning using scales for recovery, stress, sleep quality, muscle soreness, and weekly body mass measurements.
Data preparation involved transforming time-series data into independent observations using a process called Markov unfolding, which allowed the application of various modeling techniques.
Multiple models were developed for group- and individual-level AM PRS and HRV change predictions. These included models using key variables identified as actionable by athletes and coaches. Leave-one-subject-out cross-validation was performed to validate the model's performance and ensure that predictions were tested on data not utilized for training.
Multiple models were developed for group- and individual-level AM PRS and HRV change predictions. These included models using key variables identified as actionable by athletes and coaches. Leave-one-subject-out cross-validation was performed to validate the models' performance and ensure that predictions were tested on data not utilized for training.
Analysts assessed variable importance in identifying the most influential factors in the projections and created partial dependence profiles to aid in interpreting the model's predictions. The analysis revealed that while group models generally showed improved accuracy over baseline models, the performance of individual models varied widely, highlighting the need for personalized approaches in monitoring and adjusting training and recovery strategies.
Results and Insights
The study analyzed data from 3,572 days of tracking across 43 endurance athletes, with each participant contributing an average of 83.1 days. The participants had a mean sleep duration of 7.5 hours per night, ingested 39.6 kcal/kg daily, and trained for an average of 11.8 hours weekly. The change in primary outcome variables, AM PRS and HRV averaged 0.0, with density plots of their distribution available in supplemental materials.
Group models showed improved accuracy over baseline models, with individual models also outperforming baseline models. However, the accuracy varied significantly across participants (RMSE range 5.5–23.6 for AM PRS and 0.05–0.44 for HRV change). Key variables influencing AM PRS and HRV change predictions were identified, with scatterplots comparing predicted vs. actual values and partial dependence plots illustrating how predictions change with these variables.
The top variables from individual models for AM PRS and HRV change were highlighted, showing that individual model performance varied widely. These findings suggest that a small subset of variables provides most of the predictive power at the group level, but the key variables may vary at the individual level, indicating the importance of a personalized approach.
The models' average root mean squared error (RMSE) and R-squared values were calculated across 500 bootstrap resamples, providing a comprehensive view of model accuracy. The variability in individual model performance underscores the complexity of predicting recovery and HRV changes and suggests that additional data may be needed to improve individual-level predictions.
Additionally, the analysis revealed that models using the top five variables from the MAIN group and individual MAIN models provided the most accurate predictions for both AM PRS and HRV change. While group models generally offered improved predictive accuracy compared to baseline models, the substantial variability in individual models reflects the unique responses of each athlete. The variability underscores the importance of tailoring predictions to individual athletes for more precise monitoring and adjustment of training and recovery strategies.
Conclusion
The study demonstrated that ML models effectively predicted group-level recovery measures using common variables, reducing prediction error compared to baseline models. However, individual-level predictions exhibited significant variability, suggesting that personalized approaches and additional data may be necessary to enhance accuracy. The study highlights the potential of ML models in sports science while emphasizing the need for individualized monitoring to better support athlete training and recovery.
Machine Learning Models Predict Arsenic Contamination