GNN-driven modeling and prediction of multi-dimensional correlation of international medical education quality based on competence

This study addresses the challenge of modeling competency-based international medical education (IME) quality, where traditional methods struggle with heterogeneous multi-dimensional indicators and temporal dynamics. We propose RT-GAT (Relational Temporal-Graph Attention Network), a GNN-based model integrating relational attention and temporal encoding to improve competency correlation modeling and prediction accuracy. The model constructs a multi-relation heterogeneous graph with nodes representing knowledge/skills, clinical practice, communication, and teamwork, connected by driving, collaborative, and feedback edges. A relation-aware attention mechanism adapts GAT by applying edge-specific transformations and attention vectors to distinguish relationship influences. Temporal features are captured via sine-cosine positional encoding and Bi-LSTM, enabling dynamic competency evolution tracking across teaching stages. A multi-head attention architecture learns multi-subspace features, while a multi-task regressor jointly predicts four-dimensional competencies. Evaluated on five-stage data from 900 international medical students, RT-GAT achieves a cosine similarity of 0.91, adjacency reconstruction error of 0.83, and spectral distance of 0.36 in competency correlation modeling. Incorporating temporal relationships boosts the Pearson correlation coefficient by 0.168 on average. Prediction performance shows RMSE (∼4.8), MAE (∼3.7), and MAPE (∼6.1%). Cross-cultural validation with students from the US, India, UK, Russia, and Nigeria demonstrates stable accuracy (RMSE: 4.8–6.4; MAE: 3.7–5.1), confirming adaptability to heterogeneous backgrounds. RT-GAT excels in structural representation and temporal evolution capture, offering a robust technical solution for competency-based IME evaluation and prediction.