Knowledge-Induced Medicine Prescribing Network for Medication Recommendation

TL;DR

KindMed addresses medication recommendation by unifying EHR data with external medical knowledge through patient-specific medical KGs. It introduces relational KG learning via R-GCNs, hierarchical sequence modeling to fuse clinical and medication dynamics, and an attention-based prescribing module that accounts for current state and historical trends. The model jointly optimizes BCE, multi-label, and DDI losses, with a controllable DDI threshold to balance safety and accuracy, and demonstrates superior performance and safety across MIMIC-III and MIMIC-IV compared to strong baselines. The work highlights the value of integrating ontology, semantics, and DDIs in a cohesive KG framework to advance personalized, safe prescribing in real-world clinical settings.

Abstract

Extensive adoption of electronic health records (EHRs) offers opportunities for their use in various downstream clinical analyses. To accomplish this purpose, enriching an EHR cohort with external knowledge (e.g., standardized medical ontology and wealthy semantics) could help us reveal more comprehensive insights via a spectrum of informative relations among medical codes. Nevertheless, harnessing those beneficial interconnections was scarcely exercised, especially in the medication recommendation task. This study proposes a novel Knowledge-Induced Medicine Prescribing Network (KindMed) to recommend medicines by inducing knowledge from myriad medical-related external sources upon the EHR cohort and rendering interconnected medical codes as medical knowledge graphs (KGs). On top of relation-aware graph representation learning to obtain an adequate embedding over such KGs, we leverage hierarchical sequence learning to discover and fuse temporal dynamics of clinical (i.e., diagnosis and procedures) and medicine streams across patients' historical admissions to foster personalized recommendations. Eventually, we employ attentive prescribing that accounts for three essential patient representations, i.e., a summary of joint historical medical records, clinical progression, and the current clinical state of patients. We validated the effectiveness of our KindMed on the augmented real-world EHR cohorts, achieving improved recommendation performances against a handful of graph-driven baselines.

Figures (5)

• Figure 1: Overview of KindMed and its internal modules, including relation-aware GNNs for learning and enriching the node embedding in medical KGs, a fusion module for integrating temporal features from clinical and medicine streams, and an attentive prescribing module for recommending medications.
• Figure 2: Summary of demographic bins in MIMIC-III cohort.
• Figure 3: Summary of demographic bins in MIMIC-IV cohort.
• Figure 4: Performance comparison of KindMed and SafeDrug by varying DDI threshold hyperparameter.
• Figure 5: The t-SNE plots over KindMed variants on the MIMIC-III cohort. Ontology and semantic relations are shown in green and yellow, respectively. We deliberately showed semantic inter-relations across two entities to make them more prominent.