Physiological load estimation in athletes using ECG-derived features and gradient-boosted modeling.

Area of Science:

• Sports Science
• Biomedical Engineering
• Data Science

Background:

• Accurate measurement of internal training load is vital for optimizing athlete performance and preventing overtraining.
• Existing athlete monitoring tools often rely on proprietary, non-transparent methods for calculating training load metrics like Training Impulse (TRIMP).

Purpose of the Study:

• To develop and validate a transparent, reproducible machine learning framework for estimating session-level TRIMP using electrocardiogram (ECG) and heart rate variability (HRV) data.
• To provide an open-source pipeline for session-level TRIMP calculation, enhancing the consistency and reliability of sports analytics.

Main Methods:

• A retrospective analysis using the SportDB 2.0 dataset was performed.
• Supervised machine learning models (including XGBoost) were trained using ECG-derived features (heart rate, HRV) to predict TRIMP, excluding TRIMP-related variables to prevent circularity.
• SHAP (SHapley Additive exPlanations) was employed to ensure model interpretability.

Main Results:

• The XGBoost model demonstrated superior cross-validation performance among the evaluated machine learning algorithms.
• ECG-derived features, especially HRV metrics, effectively reconstructed the reference TRIMP, achieving strong predictive accuracy.
• The framework successfully provided a reproducible method for session-level TRIMP estimation.

Conclusions:

• ECG and HRV data provide a robust foundation for calculating session-level TRIMP, offering a viable alternative to traditional methods.
• The developed framework offers a transparent, interpretable, and reproducible solution for sports scientists and coaches to monitor athlete internal load.
• This approach facilitates consistent and reliable sports analytics, aiding in evidence-based training management.