MMAN: A Multi-Metric Attention Network for the intelligent physiological fatigue risk classification

Accurate physiological monitoring of athlete fatigue in competitive sports is essential for performance optimization and injury prevention, yet existing methods relying on subjective reporting or single-metric signals lack robustness and sensitivity. To address this limitation, we propose a Multi-Metric Attention Network (MMAN) for fatigue risk classification using heterogeneous physiological and training data. The MMAN model is evaluated on the AFR-1000 synthetic dataset, which is generated based on exercise physiology principles to simulate athletic monitoring scenarios. MMAN incorporates domain specific processing branches and a cross metric attention mechanism that captures complementary interactions across demographics, training load, physiological markers, and recovery indicators. A feature group gating module further provides interpretable importance weighting. Evaluated on the AFR-1000 dataset, MMAN achieves 94.5% F1-score and substantially outperforms traditional machine learning(ML) and early fusion baselines. Ablation results confirm the benefits of cross metric attention and feature group gating, while analysis of learned weights highlights the dominant contribution of physiological and recovery features. The system also demonstrates efficient inference suitable for practical deployment. Overall, MMAN advances fatigue monitoring by integrating domain-aware representation learning with interpretable multi-metric fusion, offering a robust foundation for future validation on real athlete populations.