Six-dimensional predictive analytics for multi-dimensional strength and conditioning monitoring in collegiate athletes: a proof-of-concept study

Collegiate athletes aged 18 to 22 years occupy a transition period between adolescent development and adult performance, yet existing training frameworks rarely address the full range of physical and cognitive capacities that determine competitive success and injury resilience in this population. This proof-of-concept study developed and internally cross-validated an ensemble machine learning model estimating six-dimensional performance changes and injury occurrence over rolling 28-day windows, intended to inform, but not yet to guide, training monitoring in university athletes. Data were collected from 150 collegiate athletes (mean age 20.1 ± 1.2 years; soccer, basketball, track and field) across a six-month competitive season, using rolling 30-day predictor windows and 28-day outcome windows (840 athlete-window observations nested within 150 athletes), encompassing physiological, biomechanical, neuromuscular, and cognitive assessments alongside continuous training load monitoring. An ensemble combining a Gradient Boosting Machine (GBM) and a Long Short-Term Memory (LSTM) network estimated dimensional performance change and time-loss injury occurrence. Using athlete-grouped nested cross-validation, the ensemble achieved R² = 0.92 (95% CI 0.90 to 0.94) and RMSE = 5.34 (95% CI 4.93 to 5.75) under the primary class-weighted configuration, exceeding both individual model components and a regularized linear baseline. For injury classification, the primary athlete-level AUC-ROC was 0.85 (95% CI 0.77 to 0.91); this estimate rests on only 34 unique injured athletes and should be interpreted as purely exploratory and explicitly underpowered. Dimension-specific R² ranged from 0.89 for cognitive function to 0.93 for endurance. VO₂ max, fatigue index, and acute-to-chronic workload ratio (ACWR) emerged as the strongest cross-dimensional predictors in global SHAP analysis. Athletes with greater dimensional profile imbalance showed higher observed injury incidence ( r = 0.48, 95% CI 0.36 to 0.58, p < 0.001, FDR-corrected), and post-fatigue cognitive performance was associated with injury occurrence ( r = 0.44, 95% CI 0.31 to 0.55, p = 0.002, FDR-corrected). As the model has not undergone external validation, is based on a single-city, single-cohort, single-season design, and the injury classification task is constrained by a small number of independent injured athletes ( n = 34), these findings should be interpreted strictly as hypothesis-generating, and no clinical, operational, or training application should be inferred from this work at this stage. A rule-based training-prescription overlay is described as an illustrative, non-validated exploratory concept only and has not been prospectively tested.