Author(s): BJØRNDAL, C., PHIPPS, D.J., GUSTAFSSON, H., SAARINEN, M., Institution: NORWEGIAN SCHOOL OF SPORT SCIENCES, Country: NORWAY, Abstract-ID: 2505

INTRODUCTION:
Sport school students face cumulative physical and psychological demands that may increase injury risk. While previous research has identified associations between training load, prior injury, and burnout, most studies rely on cross-sectional or regression-based approaches that may not capture complex, non-linear and temporal interactions. This study examined whether machine learning could predict weekly incident acute and overuse injuries among adolescent sport school students by integrating psychological, workload, and injury history variables.
METHODS:
A longitudinal design was applied to 211 Norwegian lower secondary sport school students (Grades 8–10), contributing 2,135 healthy-week observations. Weekly surveys assessed sport-related injuries, training exposure, sport burnout, and sport engagement. Incident injuries (acute and overuse) were predicted one week ahead using a Random Forest model with balanced sampling and person-level cross-validation. Twenty-seven features capturing acute workload, chronic workload patterns, psychological states and traits, injury history, and interaction terms were included. Model performance was evaluated using cross-validated accuracy, class-specific accuracy, and one-vs-rest AUC.
RESULTS:
The optimized Random Forest achieved 68.3% mean cross-validated accuracy. Discriminative ability was moderate (AUC = 0.57 for acute injuries; 0.66 for overuse injuries). Injury history variables (personal injury rate and weeks since injury) were the strongest predictors. Chronic workload characteristics and workload variability contributed more to prediction than acute weekly load alone. Interaction effects were important: acute load × recent injury and engagement × recent injury ranked among the top predictors. Chronic psychological patterns, particularly long-term engagement and burnout-related inadequacy, were more predictive than weekly fluctuations. The full model significantly outperformed simplified and regression-based comparison models.
CONCLUSION:
Injury risk among adolescent sport school students reflects cumulative and interacting risk processes rather than isolated weekly factors. Machine learning approaches capture these non-linear and conditional patterns more effectively than traditional regression models. Early and continuous monitoring of injury history, workload stability, and psychological functioning may support more targeted and timely prevention strategies in sport school contexts.