MedGCN: Medication recommendation and lab test imputation via graph convolutional networks

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare

Background:

• Laboratory testing and medication prescription are crucial in clinical practice.
• Current methods may lead to redundant tests and suboptimal prescriptions.

Purpose of the Study:

• To develop an intelligent medical system (MedGCN) for automated medication recommendations and lab test imputations.
• To enhance clinical decision-making and reduce healthcare costs.

Main Methods:

• Integrated medical entities (patients, encounters, lab tests, medications) into a heterogeneous graph.
• Utilized graph convolutional networks (GCN) to learn entity representations.
• Implemented a cross-regularization strategy for multi-task learning.

Main Results:

• MedGCN demonstrated superior performance in both medication recommendation and lab test imputation tasks.
• The model was validated on two real-world datasets (NMEDW and MIMIC-III).

Conclusions:

• MedGCN offers a promising and reliable AI-driven solution for clinical practice.
• The system assists physicians in making informed medication prescriptions and optimizing lab test utilization.