Graph-Enhanced Multi-Task Learning for Type 2 Diabetes Comorbidity Risk Prediction.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Early prediction of Type 2 diabetes mellitus (T2DM) complications is vital for patient outcomes and reducing healthcare costs.
• Existing prediction methods for diabetic complications have significant limitations.

Purpose of the Study:

• To propose a novel Graph-Enhanced Multi-Task Learning (GEMTL) framework for simultaneous prediction of multiple diabetic complication risks.
• To improve the accuracy and effectiveness of predicting diabetic complications.

Main Methods:

• Constructed disease relation graphs using a hybrid strategy combining data-driven and knowledge-driven approaches.
• Employed graph neural networks (GNNs) for higher-order disease dependency analysis.
• Utilized cross-attention mechanisms for fusing patient features with disease graph embeddings and a multi-gating expert network for task-specific modeling.

Main Results:

• The GEMTL framework achieved a macro-averaged F1 score of 0.723, micro-averaged F1 score of 0.856, and mean Average Precision of 0.759 on the MIMIC-IV dataset.
• Demonstrated significant outperformance compared to traditional machine learning, deep multi-task learning, GNN, and multi-expert baseline methods across all metrics.

Conclusions:

• The proposed GEMTL framework offers an effective technical solution for complex medical multi-task prediction.
• This framework has broad application prospects in precision diabetes management and clinical decision support for Type 2 diabetes mellitus.