Multi-Agent Reinforcement Learning for Sports Injury Prediction: A Comparative Evaluation with Deep Learning Baselines

Abstract

Sports injuries pose a pervasive challenge affecting athletes at every level of competition, from grassroots amateurs to elite professionals. The financial, physiological, and psychological costs are enormous. Teams lose key players at critical moments, athletes endure career-altering damage, and medical staff are often left struggling to act on incomplete information. Yet despite the growing availability of biomechanical data, GPS tracking, and electronic health records, the field still lacks a unified, adaptive framework capable of modeling the complex, interdependent nature of injury risk across an entire squad. This paper presents a systematic comparative evaluation of four machine learning approaches for sports injury prediction: Random Forest, Long Short-Term Memory networks (LSTM), Single-Agent Deep Deterministic Policy Gradient (DDPG), and a proposed Multi-Agent Reinforcement Learning (MARL) framework based on MADDPG with temporal attention. The MARL framework models each athlete as a cooperative agent in a shared physiological environment, with the hypothesis that capturing inter-player dependencies would improve prediction over single-player approaches. Each model is trained and evaluated under identical conditions on a real-world multimodal athlete-monitoring dataset. Experimental evaluations are conducted on a real multimodal athlete-monitoring dataset containing physiological signals, including heart rate variability, EMG amplitude, skin temperature, blood oxygen saturation, and workload metrics, from 5,430 athlete observations. Four models are evaluated: Random Forest, LSTM, Single-Agent DDPG, and the proposed MARL framework. Results show that Random Forest achieves the strongest discriminative performance with an AUC-ROC of 0.996, while the LSTM baseline reaches 0.932. The MARL framework faces convergence challenges on this single-player tabular dataset, revealing an important empirical finding: cooperative multi-agent dynamics are difficult to leverage without explicit inter-player sequential interaction structure in the data. This paper makes an honest empirical contribution: it establishes a rigorous benchmark for injury prediction on real physiological monitoring data, demonstrates that classical ensemble methods remain highly competitive baselines, and identifies the precise data conditions under which MARL frameworks are and are not effective. These findings provide a clear roadmap for future work.