KAMPNet: multi-source medical knowledge augmented medication prediction network with multi-level graph contrastive...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Data Mining

Background:

• Medication prediction is vital for intelligent healthcare systems using electronic medical records (EMR).
• Existing methods struggle with complex, heterogeneous medical data and overlook crucial relations like synergistic, concurrent, and therapeutic links.
• This limits prediction accuracy and real-world applicability.

Purpose of the Study:

• To develop KAMPNet, a novel network for medication prediction that leverages multi-sourced medical knowledge.
• To effectively capture diverse relationships within medical data, including implicit, correlative, and temporal connections.
• To improve medication prediction performance in intelligent healthcare.

Main Methods:

• KAMPNet employs a multi-level graph contrastive learning framework to capture diverse medical code relations.
• It utilizes unsupervised graph contrastive learning with graph attention networks and weighted graph convolutional networks for knowledge and relation augmentation.
• Augmented embeddings are integrated with supervised embeddings in a sequential learning network for temporal analysis and prediction.

Main Results:

• KAMPNet demonstrated superior performance compared to baseline models on the MIMIC-III dataset.
• Key metrics including Jaccard index, F1 score, and PR-AUC validated the model's effectiveness in medication prediction.
• The model successfully captured complex medical code relationships.

Conclusions:

• KAMPNet effectively integrates multi-sourced medical knowledge and diverse code relations via multi-level graph contrastive learning.
• The multi-channel sequence learning network enhances the capture of temporal dynamics for comprehensive patient representations.
• This facilitates improved downstream tasks, particularly medication prediction in clinical settings.