Sports injury risk prediction based on temporal graph encoding and graph neural networks: A cross-sport transfer learning framework

Sports injuries significantly impact athletes' health and performance, yet existing prediction methods struggle to capture complex athlete interactions and to generalize across sports with limited data. This study proposes a novel injury risk prediction framework integrating temporal graph encoding with graph neural networks and cross-sport transfer learning. The approach transforms multivariate training data into graph structures using Gramian Angular Fields and Markov Transition Fields, enabling spatiotemporal feature extraction through parallel graph convolution and temporal convolution pathways. A domain adaptation mechanism facilitates knowledge transfer from data-rich source sports to data-scarce target sports. Evaluated on a dataset of 312 athletes across five sports, the framework achieved an AUC of 0.826 ± 0.025, outperforming state-of-the-art baselines, including GAT (0.781) by 5.8%. Transfer learning experiments demonstrated remarkable small-sample performance, maintaining 0.80 AUC with only 50 target domain samples. Feature importance analysis revealed training load (0.182), load variability (0.156), and recovery time (0.138) as primary risk factors. The attention mechanism identified notable athlete interactions (weights > 0.7), providing interpretable insights for coaches. This framework demonstrates a promising approach for injury prevention in the evaluated sports contexts, particularly benefiting sports with limited historical data, and contributes to advancing sports health management systems.