HetMS-AMRGNN: Heterogeneous multi-scale graph neural network for antimicrobial drug recommendation in electronic health records

Objective:

To develop a novel heterogeneous graph representation learning approach for antimicrobial drug recommendation in intensive care units (ICUs) that effectively addresses the complexities of combination therapy and heterogeneous electronic health records (EHRs) data.

Methods:

We propose HetMS-AMRGNN, which represents EHR data as a heterogeneous graph with multiple types of nodes and edges capturing clinical relationships. The model employs multi-view feature extraction and multi-scale graph convolution to capture structural information at different scales, while using metapath-based aggregation to integrate diverse semantic relationships. A hierarchical contrastive learning mechanism is introduced to handle intra-node heterogeneity, and the node representations are enhanced with historical diagnosis and drug–drug interaction knowledge for accurate prediction.

Results:

Experimental validation on real-world ICU EHR data demonstrates that HetMS-AMRGNN significantly outperforms existing approaches in antimicrobial drug recommendation tasks. The model shows particular strength in recommending combination therapies, effectively capturing complex patient characteristics and drug interaction patterns.

Conclusion:

HetMS-AMRGNN provides an effective solution for antimicrobial drug recommendation in ICU settings, successfully addressing the challenges of combination therapy and heterogeneous data integration. The model’s superior performance, particularly in complex cases requiring combination therapy, suggests its potential for improving antimicrobial prescribing practices in critical care.