DKINet: Medication Recommendation via Domain Knowledge Informed Deep Learning

TL;DR

DKINet tackles medication recommendation for patients with complex conditions by integrating external domain knowledge from UMLS with observable EHR signals. It introduces a graph-aggregation module over UMLS, a knowledge-injected visit representation, and a historical medication-aware branch, all guided by a mutual-information-based regularization and an external-knowledge loss term. Across MIMIC-III, MIMIC-IV, and eICU, DKINet achieves state-of-the-art results on multi-label medication prediction, and ablations confirm the essential role of each component. The work demonstrates the practical value of combining domain knowledge with longitudinal patient data, and provides code to support reproducibility and further research.

Abstract

Medication recommendation is a fundamental yet crucial branch of healthcare that presents opportunities to assist physicians in making more accurate medication prescriptions for patients with complex health conditions. Previous studies have primarily focused on learning patient representation from electronic health records (EHR). While considering the clinical manifestations of the patient is important, incorporating domain-specific prior knowledge is equally significant in diagnosing the patient's health conditions. However, effectively integrating domain knowledge with the patient's clinical manifestations can be challenging, particularly when dealing with complex clinical manifestations. Therefore, in this paper, we first identify comprehensive domain-specific prior knowledge, namely the Unified Medical Language System (UMLS), which is a comprehensive repository of biomedical vocabularies and standards, for knowledge extraction. Subsequently, we propose a knowledge injection module that addresses the effective integration of domain knowledge with complex clinical manifestations, enabling an effective characterization of the health conditions of the patient. Furthermore, considering the significant impact of a patient's medication history on their current medication, we introduce a historical medication-aware patient representation module to capture the longitudinal influence of historical medication information on the representation of current patients. Extensive experiments on three publicly benchmark datasets verify the superiority of our proposed method, which outperformed other methods by a significant margin. The code is available at: https://github.com/sherry6247/DKINet.

Figures (5)

• Figure 1: An example mapping of diagnoses and UMLS relation graph. The diagnosis code 401.9 is mapped to the corresponding CUI code C0085580 through the UMLS vocabulary. On the right of the figure, UMLS concepts connected to Essential hypertension are displayed along with their corresponding relationship pathways.
• Figure 2: An overview of the proposed DKINet framework. DKINet has four main components: a) EHR-based visit-level representation module takes diagnosis $d^t$, produces $p^t$, and previous visit medication record $m^{t-1}$ as input to get the EHR-based representation of the patient; b) graph aggregation module extracts the domain knowledge information and combine the current EHR input to obtain the knowledge-based embedding vectors $e_{d_k}^t,e_{p_k}^t,e_{m_k}^{t-1}$; c) knowledge-injected visit-level representation module incorporates the domain knowledge with the clinical manifestation and generates the knowledge-enhanced representation; d) historical medication-aware patient representation module is designed to capture the longitudinal influence of historical medication information. Finally, an output layer is designed to make recommendation.
• Figure 3: Performance comparison of DKINet, DKINet$_{w/o \mathcal{L}_{mi}}$ and DKINet$_{w/o \mathcal{L}_{ekg}}$ on three metrics (i.e., Jaccard, F1, PRAUC), across the MIMIC-III, MIMIC-IV, and eICU datasets.
• Figure 4: Explanations of filter-base path aggregation and real example in MIMIC-III dataset. Best viewed in color.
• Figure 5: A visualization of the distribution between knowledge-enhanced and EHR-based visit representation in MIMIC-III dataset at different training epochs employing Umap.