Abstract / Summary
The wide range of physiological demands in trail running makes performance modelling challenging, and existing models require resource intensive laboratory testing and lack individual predictive resolution. This study aimed to develop a two-layer physics-informed machine learning (ML) pipeline that predicts pre-race finish times without laboratory assessment – a physics-based simulation layer, whose parameters were calibrated from athlete's historical Global Positioning System (GPS) race records, combined with an XGBoost stacking layer that learned systematic residuals of the physics model. The pipeline was evaluated under leave-one-athlete-out (LOAO) cross-validation across 579 trail running races from 62 competitive athletes (44 males, 18 females; age: 32.3 ± 7.1 years; ITRA Performance Index: 739 ± 109) spanning five different race formats. The XGBoost stacking layer achieved a mean absolute percentage error (MAPE) of 11.2% (MAE = 22.0 min, RMSE = 43.9 min, R 2 = 0.938), significantly outperforming the physics layer alone (MAPE = 11.7%; W = 73.5, p = 0.010). Feature importance analysis revealed the physics-layer prediction (37.6%) and log-transformed race distance (48.3%) together accounted for 85.9% of the XGBoost model's explanatory power, confirming the physics model as a structured informative prior. Accuracy varied by format (MAPE: 9.9–21.9%) and the stacking layer reduced the physics model's proportional bias from r = −0.564 to r = −0.206 and corrected elevation-dependent underestimation by up to 78.6% in races >3000 m elevation gain. This framework is the first population-validated, laboratory-free digital twin for trail running performance prediction, providing both athlete-specific accuracy and mechanistic interpretability from GPS data alone.