Advances in the Leiden Power-Heart Rate model

Abstract

Purpose: To better model riders’ physiological response to prolonged exertion beyond the non-linear relationship between power and heart rate, with the aim of assessing fitness in the field. Methods: The Leiden PHR model captures the non-trivial relationship between power (external load) and heart rate (internal load), shown to provide insights into a rider’s fitness level ^1,2. Because heart rate is influenced by factors beyond recent power history, we modeled the remaining unexplained variance using the machine learning model XGBoost. Additional explanatory variables included cumulative effort (in five alternative ways), atmospheric conditions, geographic location, and current altitude and slope. Feature importance methods were used to assess the relative contribution of variables. Results: As in the earlier models, convolved pedal power was the strongest predictor, followed by a mixture of local parcours properties and cumulative effort indicators. Although atmospheric conditions and geographic location offered some minimal contribution, the same information can be obtained from other variables. The most accurate model used a mixture of variables (R² = 0.81, compared to R² = 0.70 for the existing model, with the average error being