Athlete modeling for personalized race performance predictions in endurance sports

The work of this thesis relies on modeling fatigue in endurance sport as a smooth, athlete-normalized dynamical state that links external workload and internal responses to performance. Fatigue is treated as the progressive loss of capacity to sustain pace or power, emerging from cardiovascular, metabolic, neuromuscular, and perceptual factors. Because no single laboratory metric or rating captures this heterogeneity, we seek a representation that is data efficient, comparable across athletes, and predictive. We adopt an interpretable bounded formulation in which a differentiable state in [0,1] accumulates when intensity exceeds an athlete's specific critical threshold and dissipates during recovery, with a smooth transition between the two regimes. Threshold normalization keeps it interpretable and comparable across periods and sessions while avoiding hard cut-offs and fixed decays that miss transitions near maximal effort. To estimate parameters robustly and enable credible forward simulation for race time prediction, high-resolution, low-noise data are required. We therefore leverage modern wearable technology to collect synchronized streams describing external load (pace, distance, power) and internal responses (heart rate and variability, breathing surrogates), enriched with metadata on perceived dynamics and load to improve model interpretability. To obtain a diverse and harmonized corpus, we built a centralized enrollment and ingestion portal that time aligns sensors, enforces validity checks, standardizes units, and exposes a stable analytical schema, enabling reproducible modeling and validation. In a small pilot study, calibrated simulations reproduced observed finish times while maintaining low fatigue during easy sessions, offering face validity for both accumulation and recovery dynamics. An initial session labeling module that assigns effort levels has been cross-checked against manual labels to validate the approach and tune thresholds. While preliminary, these results include single-digit race-time precision on a small sample of recreational athletes; scaling to larger, denser datasets is expected to support external validation, assess generalizability across athletes and contexts, and refine parameter estimation.