Athlete-AI: a multi-modal artificial intelligence framework for personalized athlete monitoring and injury risk forecasting
R. Kanthavel, R. Dhaya, Satish Raosaheb Jondhale, Amruta S. Jondhale, Geetanjali Sharma · Scientific Reports · 2026
Artificial intelligence (AI) provides opportunities for computational athlete monitoring, adaptive training analysis, and injury-risk modelling. This study presents Athlete-AI, a multimodal AI framework that integrates wearable sensor data, human motion information, pose features, and literature-derived athlete variables into a unified computational architecture. The framework combines multimodal feature fusion, reinforcement learning for adaptive training simulation, and machine-learning-based injury-risk modelling. The evaluation was conducted using publicly available benchmark datasets, including PAMAP2, Human3.6 M, and Microsoft COCO, together with literature-derived parameters used to construct simulated athlete profiles, workload, fatigue, recovery, and injury-risk scenarios. The framework was evaluated through computational experiments and simulation-based athlete-monitoring scenarios to examine multimodal integration, injury-risk classification, and adaptive training behaviour. Repeated-run evaluation across five random seeds, probability-calibration analysis, component ablation, comparison with multimodal fusion baselines, and a leakage-controlled analysis of the dependence on label-defining variables were also performed. When the workload, fatigue, recovery, and biomechanical-risk variables used to define the simulated labels were removed from the predictors, the F1-score decreased from 86.6 to 72.8%. Negative-control analyses showed that the benchmark-derived features-only model produced performance comparable to the shuffled-label control. The
Results: demonstrate the computational feasibility of integrating heterogeneous athlete data and evaluating adaptive training and injury-risk modelling within a unified simulation environment. The reported performance values represent preliminary computational
results and do not constitute real-world, clinical, or prospective athlete validation. Therefore, the proposed framework should be regarded as a computational research framework that requires validation with longitudinal data from real athletes before practical or clinical deployment.