Interpretable prediction of knee joint loading during tennis serves based on GNN-GRU model and layer-wise relevance propagation.

The knee resultant joint moment is a critical indicator for assessing risk during the tennis serve. Traditional methods for obtaining this metric rely on laboratory-based equipment, limiting practical application. To address this limitation, this study proposes and validates a novel method for predicting the knee resultant joint moment method using a Graph Neural Network and Gated Recurrent Unit (GNN-GRU) model. An independent GRU model was used as a baseline for comparison. Biomechanical data were collected from 30 male tennis players (age: 20.30 ± 1.66 years, height: 176.60 ± 2.74 cm, weight: 70.80 ± 3.89 kg, BMI: 22.71 ± 1.38 kg/m², training experience: 9.20 ± 2.81 years) during the performance of the tennis serve. Sagittal plane joint angles of both lower limbs were used as model inputs to predict the resultant joint moment of the supporting leg. A paired-sample t-test compared predicted and actual values. Layer-wise Relevance Propagation (LRP) was applied to quantify the contribution of individual joint angles. The GNN-GRU model demonstrated significantly better prediction performance than the standalone GRU model (p < 0.05). No significant differences were observed between predicted and actual values (p > 0.05). LRP results showed knee contribution close to 1 during the Preparation Phase (PP). In the Flight Phase (FP), ankle and hip contributions increased significantly, both approaching 1. During the Landing Phase (LP), the knee joint maintained a contribution above 0.4. This study supports the identification of potentially high-risk movements in real-world tennis training and competition and provides a reference for the early detection of knee joint injuries.