Relation-aware pre-trained network with hierarchical aggregation mechanism for cold-start drug recommendation

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine

Background:

• Existing drug recommendation systems struggle with data sparsity for single-visit patients.
• Current models often overlook the challenge of limited patient data in cold-start scenarios.

Purpose of the Study:

• To propose a novel Relation-aware Pre-trained Network (RPNet) for enhanced drug recommendation.
• To address data sparsity and cold-start issues in patient-specific drug prescription.

Main Methods:

• Developed a pre-training and fine-tuning framework (RPNet).
• Introduced a code matching discrimination task using mask-replace contrastive learning.
• Implemented a hierarchical aggregation mechanism for visit selection and patient similarity matching.

Main Results:

• RPNet demonstrated superior performance on two real-world datasets.
• Achieved notable improvements in the F1 metric by 1.32% and 1.19% compared to existing methods.
• Effectively captured robust feature representations and integrated drug information.

Conclusions:

• RPNet offers a significant advancement in drug recommendation systems, especially for data-scarce patients.
• The proposed methods effectively handle complex relationships between medical entities and patient data.
• RPNet provides a promising solution for personalized and precise drug prescriptions.