Key result
Adding total mechanical work and session RPE to past HRV history produced only modest reductions in RMSE for forecasting next-day HRV compared to HRV-only models in competitive cyclists.
Why the study?
The extent to which next-day heart rate variability can be forecast from readily available external and internal training-load metrics remains unclear.
Does adding training load metrics to machine learning models improve the prediction of next-day HRV in competitive cyclists compared to HRV-only models?
Observational (n=7)
Does adding training load metrics to machine learning models improve the prediction of next-day HRV in competitive cyclists compared to HRV-only models?
Routinely collected training load descriptors (total work and RPE) add limited information beyond recent HRV history for forecasting next-day HRV in competitive cyclists.
Answers are grounded in this paper and cite their sources.
ML models may forecast next-day HRV from training loads in cyclists; leaves open utility for guiding recovery.
Introduction: Day-to-day fluctuations in heart rate variability (HRV) are widely used to infer autonomic recovery in endurance athletes. However, the extent to which HRV can be forecast one day ahead from readily available external and internal training-load metrics remains unclear. In this study, we evaluated whether machine learning models can predict next-day HRV in competitive cyclists using the two load descriptors most commonly collected in practice: external load quantified as total mechanical work in kilojoules (kJ) and internal load quantified as session rating of perceived exertion (RPE). Methods: Seven male competitive endurance cyclists were monitored daily for sixteen weeks, yielding 590 athlete-days of longitudinal data (seven independent time series). Two machine learning approaches—support vector regression (SVR) and extreme gradient boosting (XGBoost)—were compared with a conventional autoregressive model with exogenous inputs (ARX) as a traditional time-series benchmark. Each model was trained individually per athlete under two predictor scenarios (using past HRV-only or past HRV plus kJ and RPE) and across multiple lag orders (1, 4, 7, 10 and 14 days), with forecasting accuracy expressed as root mean squared error (RMSE). Results: Across all athletes, adding kJ and RPE to the past HRV produced only modest reductions in RMSE relative to HRV-only models. XGBoost achieved the lowest one-step-ahead RMSE at short lag, while all models converged at longer lag orders. Predictive accuracy differed markedly between athletes, reflecting the well-known individual nature of autonomic responses. Conclusions: These findings suggest that the two routinely collected load descriptors examined here—total work (kJ) and RPE—add limited information beyond recent HRV history for forecasting next-day HRV, and that broader contextual variables are likely required to meaningfully improve athlete monitoring.
No takes yet. Share an insight, caveat, or question.
Barsumyan et al. (2026) conducted an observational in competitive endurance cyclists (n=7). Machine learning models using past HRV plus total mechanical work (kJ) and session RPE vs. Past HRV-only models was evaluated on Forecasting accuracy of next-day HRV expressed as root mean squared error (RMSE). Adding total mechanical work and session RPE to past HRV history produced only modest reductions in RMSE for forecasting next-day HRV compared to HRV-only models in competitive cyclists.
Synapse has enriched 10 related papers. Consider them for comparative context: