A Transformer and Graph-based Multi-Task Learning for Optimized Decision-Making in Chronic Kidney Disease Management

Chronic Kidney Disease (CKD) is a endemic and severe health condition requiring early diagnosis, timely intervention, and personalized management. This paper proposes a novel multi-task learning framework for CKD management that leverages a Transformer-based encoder and Graph Neural Network (GNN) to predict CKD progression, estimate the time-to-event for disease onset, and recommend personalized interventions. The shared encoder effectively captures complex inter-feature dependencies and patient similarity relationships, while task-specific attention mechanisms ensure precision in each task. The proposed model was evaluated on a real-world CKD dataset, comprising 491 patients, and demonstrated state-of-the-art performance across all tasks. For CKD progression prediction, the model achieved an accuracy of 91%, an F1-score of 90%, and an ROC-AUC of 95%, outperforming traditional methods such as Random Forests and XGBoost. In the regression task, the model predicted time-to-event with a Mean Absolute Error (MAE) of 7.8 months and an R-squared value of 0.92, improving significantly over baseline approaches. For personalized intervention recommendations, the proposed framework attained a Precision@5 of 81%, a Recall@5 of 69%, and an MRR of 73%. Additionally, the model achieved a Weighted Multi-Task Score (WMTS) of 0.85, highlighting its ability to balance performance across multiple objectives. The integration of a Transformer-based encoder and GNN enables robust, interpretable, and actionable insights, making the framework a promising tool for advancing CKD management.