Multimodal Machine Learning for Predicting Reduced Participation in Elite Youth Football

STUDY DESIGN AND COHORT:

This is a prospective, longitudinal, single-site observational cohort study conducted across one complete 10-month competitive season (2026/2027) within the U16 and U19 squads of an elite youth football academy (SK Sigma Olomouc, Czechia). The cohort comprises approximately 30 elite male youth football players. The unit of analysis is the at-risk player-day in full participation (longitudinal person-period array), yielding an estimated 6,000 to 8,000 observed player-days.

PRIMARY OUTCOME (TARGET VARIABLE):

The primary outcome is the staff-adjudicated transition event from full participation into any lower category of the daily participation gradient (modified training with >20% volume reduction, reduced training intensity, or complete absence due to medical/health reasons). Conditional on full participation on day t, the primary modeled quantity is the one-step discrete-time onset hazard. Operational cumulative onset probabilities are read coherently across 24-, 48-, and 72-hour horizons. To ensure outcome adjudication reliability, an inter-rater agreement analysis (Cohen's kappa) is performed between independent staff members.

MULTIMODAL PREDICTOR DOMAINS:

Candidate predictors span eight pre-specified canonical domains (~31 candidate variables; ~33-36 parameters after categorical expansion):

1. Player Profile \& Maturation: Baseline anthropometry, body composition (InBody 230: FFM, % body fat), and biological maturation offset (Mirwald equation).
2. External Load: Catapult Vector GNSS/IMU metrics (total distance, high-speed running >20 km/h, sprinting >25.2 km/h, accelerations/decelerations >3.0 m/s², PlayerLoad).
3. Internal Load: Daily heart rate telemetry and session-RPE derived metrics (training monotony and training strain).
4. Autonomic Recovery \& Sleep: Continuously recorded overnight via WHOOP 5.0 wrist-worn wearable bands (resting heart rate, HRV RMSSD during slow-wave sleep, sleep architecture, and skin temperature).
5. Neuromuscular Capacity: Repeated weekly neuromuscular screening on VALD Systems (ForceFrame hip adduction/abduction and NordBord eccentric/isometric hamstring strength and bilateral asymmetries).
6. Performance Capacity: Periodic physical testing (sprint times, change of direction, and jump mechanics).
7. Contextual Factors: Microcycle schedule (days to/from match day), match congestion, and playing position.
8. Daily Subjective Wellness: Morning perceptual ratings on a 1-5 Likert scale (Hooper index: sleep, fatigue, stress, soreness). An objective-only sensitivity model will explicitly exclude subjective wellness to prevent construct overlap and circularity.

RESAMPLING AND LEAKAGE DISCIPLINE:

Model evaluation operates via a calendar-aligned rolling-origin cross-validation scheme with an expanding training window and temporally subsequent held-out test blocks. To prevent data leakage across the train/test boundary, an embedded temporal embargo of >=72 hours is implemented. All data-dependent preprocessing steps (imputation via MICE/PMM, scaling, dimensionality reduction) are fitted strictly within the training partition of each resample and applied to held-out test data.

STATISTICAL MODELLING AND EVALUATION:

Pre-specified candidate models include baseline logistic regressions, Generalised Additive Mixed Models (GAMM) with player-level random intercepts, and tree-based learners (Survival XGBoost with functional monotonicity constraints). Model evaluation follows a calibration-first hierarchy (Brier score, calibration slopes/curves), followed by discrimination under class imbalance (Precision-Recall AUC, Matthews Correlation Coefficient). Model behavior and domain contributions are interpreted using explainable artificial intelligence (XAI) frameworks (SHAP and LIME).