EDRMM: enhancing drug recommendation via multi-granularity and multi-attribute representation.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Decision Support Systems

Background:

• Current AI drug recommendation models often overlook fine-grained correlations in patient history and underutilize Electronic Health Records (EHRs).
• Existing methods may not effectively capture the nuanced relationship between past and present patient information, limiting recommendation accuracy.

Purpose of the Study:

• To propose a novel drug recommendation model, EDRMM, that addresses limitations in existing approaches by incorporating multi-granularity and multi-attribute information.
• To enhance patient representation learning by effectively identifying relevant historical information and integrating multi-attribute EHR data.

Main Methods:

• Developed a longitudinal attribute-level history selection mechanism to pinpoint fine-grained historical data relevant to current clinical conditions.
• Integrated multi-attribute EHR data with attribute-specific encoding strategies for comprehensive patient representations.
• Designed an adaptive global Drug-Drug Interaction (DDI) risk regularization term to balance recommendation accuracy and patient safety.

Main Results:

• The proposed EDRMM model achieved state-of-the-art performance on the MIMIC-III dataset.
• Experimental results demonstrated the effectiveness of incorporating key EHR attributes into patient representations.
• The model showed superior performance compared to existing methods.

Conclusions:

• EDRMM overcomes key limitations in drug recommendation by employing dynamic attribute-level history selection and integrating multi-attribute EHR data.
• The model's hybrid optimization strategy, using adaptive DDI regularization, effectively balances accuracy and safety.
• EDRMM achieves optimal drug recommendation performance by leveraging comprehensive patient representations and relevant historical data.