Safe medicine recommendation via star interactive enhanced-based transformer model.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Computational Pharmacology

Background:

• Electronic Medical Records (EMRs) are increasingly used for medicine recommendation systems.
• Existing systems often lack explicit data on drug interactions and side effects.
• Transformer-based models show promise but suffer from computational demands and information loss.

Purpose of the Study:

• To develop an advanced medicine recommendation model addressing limitations of current systems.
• To enhance the accuracy and comprehensiveness of drug recommendations by incorporating richer medical knowledge.
• To improve the handling of inductive medicine recommendation tasks.

Main Methods:

• Proposed the Star Interactive Enhanced-based Transformer (SIET) model.
• Constructed a heterogeneous graph integrating EMR (MIMIC-III) with medical knowledge graphs (ICD-9, DrugBank).
• Extracted homogeneous graphs for diseases, medicines, and negative factors to capture enhanced neighbor information, feeding these into the SIET model.

Main Results:

• The SIET model demonstrated outstanding performance on MIMIC-III, DrugBank, and ICD-9 ontology datasets.
• Achieved superior results compared to strong baseline models in inductive medicine recommendation tasks.
• Effectively integrated diverse medical data sources for improved recommendation accuracy.

Conclusions:

• The SIET model offers a significant advancement in medicine recommendation systems.
• Integrating heterogeneous medical data and leveraging graph neural networks enhances drug prescription support.
• The proposed approach effectively overcomes drawbacks of traditional Transformer-based methods, including information loss and lack of side effect consideration.