MKMed: Multi-Knowledge alignment framework for medication recommendation

Haixu Ma

doi:10.1007/s10489-026-07337-4

MKMed: Multi-Knowledge alignment framework for medication recommendation

2026-09-26

0

PRE

AI

DOI：10.1007/s10489-026-07337-4

Original

Original request for help

Save

Abstract

En 中文

Medication recommendation plays a critical role in healthcare by providing effective treatments based on patient’s electronic health records (EHR). Prior studies have shown that incorporating more medication-related knowledge significantly improves the quality of medication representations. However, not all medications are associated with multiple types of knowledge simultaneously. For example, some may have only textual descriptions but lack structured or molecular data. This imbalance in knowledge availability hampers the performance of existing models—a limitation we refer to as the “bucket effect” in medication recommendation. We further quantify its severity through a comprehensive statistical analysis on the distribution of modality coverage across medications. To address this issue, we propose a novel framework named Multi-Knowledge Medication recommendation (MKMed), which introduces a cross-modal medication encoder capable of aligning heterogeneous modalities into a unified representation space. Specifically, we pre-train the encoder using contrastive learning on five complementary knowledge modalities (text, image, structure, chemical properties, and knowledge graph), and integrate the resulting representations with patient EHR data to generate personalized medication recommendations. By effectively integrating incomplete and unevenly distributed knowledge sources, MKMed explicitly mitigates the bucket effect, enabling robust recommendation performance even in sparse knowledge settings. Extensive experiments on MIMIC-III and MIMIC-IV demonstrate that MKMed consistently outperforms state-of-the-art baselines across multiple evaluation metrics. Specifically, on MIMIC-III, MKMed achieves improvements of 1.9% in Jaccard similarity and 1.3% in PRAUC compared to the best-performing baseline. These results highlight the effectiveness of cross-modal knowledge alignment for medication representation learning, and suggest its potential to support more reliable and safer clinical decision-making in real-world healthcare scenarios.

Keywords:

Medication recommendation

Molecular representation

Knowledge fusion

MKMed: Multi-Knowledge alignment framework for medication recommendation

Abstract

Journal

Organization

Cited Papers