Data-driven computational framework for personalized training plans in cycling

Abstract Context: The widespread adoption of wearable technology in sports has generated massive datasets, yet the transition from descriptive monitoring to data-driven prescription remains largely unrealized. While predictive modeling of athletic performance has advanced, training plans are still predominantly designed through manual, expert interpretation rather than mathematical optimization.

Methods: This study presents a modular computational framework designed to automate the generation of personalized training plans in cycling. First, we introduce a continuous performance metric (PO@HRtarget), derived from submaximal longitudinal sensor data (power output and heart rate), bypassing frequent, invasive laboratory testing. Second, we extract features regarding training load, intensity distribution, and decay rates to model the dose-response relationship using four regression architectures: Lasso, Random Forest, XGBoost, and the novel Random Forest Featuring Linear Extensions (RaFFLE). Finally, we employ a differential evolution algorithm to invert these predictive models, generating optimized training prescriptions (duration, intensity, and distribution) that maximize future performance.

Results: Validation against laboratory tests confirms the metric effectively tracks physiological trends. Our analysis reveals inter-subject variability; while linear models suffice for some athletes, complex non-linear ensembles are required for others, validating the necessity of a flexible, modular architecture. The prescriptive optimization demonstrated a potential average performance increase of 5.5%.

Conclusion: We propose a framework that bridges the gap between predictive analytics and prescriptive decision-making in sports. While validated on cycling data, the methodology is sport-agnostic and scalable, offering a pathway to leverage big data for individualized performance optimization in broader endurance disciplines.