Dual-Level Interaction Aware Heterogeneous Graph Neural Network for Medicine Package Recommendation

Computer Science Information Science Artificial Intelligence Pharmacy Medicine Environmental and Public Health
medicine package recommendation dual-level interaction heterogeneous graph neural network electronic health records clinical decision-making

Research on Medical Package Recommendation Systems: Heterogeneous Graph Neural Network Based on Dual-Level Interaction Awareness

With the widespread application of electronic health records (EHRs) in the medical field, how to mine potential and valuable medical knowledge to support clinical decision-making has become an important research direction in deep learning technology. Personalized medical package recommendation is a crucial task in this field. Its goal is to help doctors select safer and more effective medication packages for each patient using extensive medical records. However, existing medical package recommendation methods primarily model the task as a multi-label classification or sequence generation problem, focusing mainly on the relationship between individual drugs and other medical entities while generally neglecting the interaction between medical packages and other medical entities. This could lead to incomplete package recommendations. Additionally, the medical common knowledge considered by existing methods is relatively limited, making it very difficult to deeply understand the decision-making process of doctors.

Overview

This paper was written by researchers Fanglin Zhu, Xu Zhang, Batuo Zhang, Yonghui Xu, and Lizhen Cui (Senior Member, IEEE) from the School of Software at Shandong University, published in the April 2024 issue of the IEEE Journal of Biomedical and Health Informatics. The paper aims to address the shortcomings of existing methods in medical package recommendations and proposes a dual-level interaction awareness heterogeneous graph neural network (DIAGNN) to improve the completeness and accuracy of medical package recommendations.

Research Process

The proposed DIAGNN introduces a heterogeneous graph to explicitly model the interactions among medical entities in EHRs and captures semantic information in the medical heterogeneous graph through a dual-level graph convolutional network. Furthermore, the study incorporates drug indications as medical common knowledge into the heterogeneous graph. The specific process is as follows:

  1. Dual-Level Interaction Modeling: Models the interaction relationships between individual drugs and drug packages in EHRs with other medical entities (such as patients and diseases), constructing a medical heterogeneous graph (including node and edge categories).
  2. Dual-Level Information Propagation: Uses a dual-level graph convolutional network for feature learning on the heterogeneous graph, encoding, propagating, and aggregating information for different levels of nodes to enhance the understanding of various medical entity relationships.
  3. Medical Package Prediction: Based on the learned medical entity features, predicts and generates recommended drug packages, calculates their relevance to the patient’s health status, and selects the most relevant drug packages as the final recommendation result.

Main Results

The key contributions of the study are as follows:

  • Innovative Introduction of Dual-Level Interaction Awareness Mechanism: By considering the interactions at both the individual drug and drug package levels, the completeness of the recommended drug packages and the precision of the recommendation system are improved.
  • Integration of Drug Indication Knowledge: By introducing new medical common knowledge, particularly drug indications, into the model, the understanding of drug and disease relationships is enhanced, preventing inaccuracies in recommendations due to knowledge gaps.
  • Extensive Experimental Validation: Experimental results on real-world datasets validate the effectiveness and superiority of the proposed method, demonstrating its ability to recommend safer and more effective drug packages.

Conclusion and Significance

The proposed DIAGNN significantly enhances the completeness and accuracy of medical package recommendations through a dual-level interaction awareness mechanism and a heterogeneous graph neural network. The study not only holds significant value scientifically, improving the clinical application level of medical package recommendations, but also has broad application potential in actual medical services, contributing to the improvement of medical service quality and efficiency.

Highlights and Innovations

  1. Dual-Level Interaction Modeling: For the first time in medical package recommendations, the interaction between drug packages and other medical entities is considered, significantly improving the completeness of the recommended drug packages.
  2. Integration of Drug Common Knowledge: The introduction of medical common knowledge, such as drug indications, enhances the scientificity and practicality of medical package recommendations.
  3. Extensive Experimental Validation: Experimental results on multiple real-world datasets show the superior performance of the proposed model, laying a solid foundation for further research and application.