Interpretable Machine Learning for Identifying Performance-Critical Flight Phases in Ski Flying Using Wearable Sensor Data

This study investigates the predictive value of different flight sub-phases in ski flying performance using machine learning (ML) and explainable artificial intelligence (XAI). Data from 105 elite-level jumps, captured during the 2022 Ski Flying World Championships and World Cup, were analyzed using wearable motion sensors and environmental metrics. The flight phase was segmented into takeoff, early flight, stable flight, and landing preparation. Three regression models—Linear Regression (LR), Random Forest (RF), and XGBoost(XGB)—were trained on each phase. Results revealed that the stable flight phase offered the highest predictive accuracy (R2 = 0.987, MAE = 1.20 m), while the takeoff phase was least predictive. SHAP (SHapley Additive exPlanations) analysis identified aerodynamic and environmental features during stable flight as key performance drivers. This is the first study to apply ML and XAI to ski flying using real-world competition data. Findings offer novel insights into flight dynamics and provide actionable guidance for optimizing training and in-air technique through data-informed coaching strategies.