Abstract
Medication recommendation systems are designed to deliver personalized drug suggestions that are closely aligned with individual patient needs. Previous studies have primarily concentrated on developing medication embeddings, achieving significant progress. Nonetheless, these approaches often fall short in accurately reflecting individual patient profiles, mainly due to challenges in distinguishing between various patient conditions and the inability to establish precise correlations between specific conditions and appropriate medications. In response to these issues, we introduce DisMed, a model that focuses on patient conditions to enhance personalization. DisMed employs causal inference to discern clear, quantifiable causal links. It then examines patient conditions in depth, recognizing and adapting to the evolving nuances of these conditions, and mapping them directly to corresponding medications. Additionally, DisMed leverages data from multiple patient visits to propose combinations of medications. Comprehensive testing on real-world datasets demonstrates that DisMed not only improves the customization of patient profiles but also surpasses leading models in both precision and safety.
Abstract (translated)
药物推荐系统旨在提供与个体患者需求高度相关的个性化药物建议。之前的研究主要集中在开发药物嵌入，取得了一定的进展。然而，这些方法往往难以准确反映个体患者的病历，主要原因是难以区分各种患者状况之间的差异以及无法建立特定状况与适当药物之间的精确关联。为了解决这些问题，我们引入了DisMed，一种关注患者状况的模型，以提高个性化。DisMed采用因果推断来确定清晰、可量化的因果关系。然后深入研究患者的状况，识别并适应这些状况的不断变化，并将它们直接映射到相应的药物。此外，DisMed利用多个患者就诊数据中的药物组合。在现实世界数据集上进行全面的测试表明，DisMed不仅提高了患者病历的定制性，而且超过了领先模型的精度和安全性。
URL
https://arxiv.org/abs/2404.12228
