Multimodal Machine Learning for Predicting Reduced Participation in Elite Youth Football

This study protocol describes a prospective longitudinal observational study designed to evaluate the accuracy of machine learning models in predicting the short-term (24-72 h) risk of reduced participation (training modification or absence due to health complaints, injury, or illness) among elite youth football players (categories U16 to U19).

Full participation in training and matches is a key determinant of athletic development and squad availability in elite youth football. Conventional injury surveillance typically reduces athlete availability to a binary time-loss state and overlooks the continuum on which athletes remain active despite emerging fatigue, minor musculoskeletal complaints, or functional limitations.

In this study, continuous multimodal monitoring integrates external training and match locomotion (Catapult Vector 10 Hz GNSS / 100 Hz IMU), nocturnal autonomic nervous regulation (WHOOP 5.0 wrist-worn photoplethysmography evaluating resting heart rate, heart rate variability RMSSD, and sleep architecture), daily subjective wellness (AMS Hooper Index: sleep quality, stress, fatigue, mood, and muscle soreness), weekly neuromuscular screening (VALD ForceDecks countermovement jumps and NordBord hamstring strength), and anthropometrics with biological maturation offset.

Participation status is adjudicated daily by designated coaching and medical staff. The primary objective is to develop, compare, and internally validate short-term prediction models for the onset transition from full participation to reduced participation, formulated as a single one-step discrete-time hazard read coherently across 24-, 48-, and 72-hour operational horizons. The study aims to evaluate explainable machine learning models (Survival XGBoost, GAMM with SHAP/LIME interpretation) to support training load management and injury risk mitigation within elite academies.