Pretraining-Based Relevance-Aware Visit Similarity Network for Drug Recommendation.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Electronic Health Records (EHR) enable drug recommendation but face challenges due to sparse data and ambiguous patient similarity.
• Precise patient modeling is crucial, requiring analysis of disease progression and similar patients' medication histories.

Purpose of the Study:

• To develop an advanced drug recommendation system that overcomes limitations in EHR data.
• To improve the accuracy and relevance of drug suggestions by integrating visit similarity and medical knowledge.

Main Methods:

• Proposed RaVSNet (Relevance aware Visit Similarity Network) leveraging longitudinal and transversal visit similarity.
• Integrated medical relevance knowledge and employed a relevance-aware network for condition-medication matching.
• Introduced a pretraining framework with Medication Sequence Reconstruction (MSR) and Causal Effect Inference (CEI) tasks.

Main Results:

• RaVSNet demonstrated superior performance over state-of-the-art methods on MIMIC-III and MIMIC-IV datasets.
• Achieved more accurate drug recommendation combinations.
• The pretraining framework improved performance when integrated with other drug recommendation methods.

Conclusions:

• RaVSNet effectively addresses challenges in EHR-based drug recommendation through enhanced patient modeling.
• The proposed pretraining framework offers a generalizable method to boost drug recommendation performance.
• This work advances the application of AI in personalized medicine and clinical decision support.