CausalMed: Causality-Based Personalized Medication Recommendation Centered on Patient Health State

Abstract

Medication recommendation systems are developed to recommend suitable\nmedications tailored to specific patient. Previous researches primarily focus\non learning medication representations, which have yielded notable advances.\nHowever, these methods are limited to capturing personalized patient\nrepresentations due to the following primary limitations: (i) unable to capture\nthe differences in the impact of diseases/procedures on patients across various\npatient health states; (ii) fail to model the direct causal relationships\nbetween medications and specific health state of patients, resulting in an\ninability to determine which specific disease each medication is treating. To\naddress these limitations, we propose CausalMed, a patient health state-centric\nmodel capable of enhancing the personalization of patient representations.\nSpecifically, CausalMed first captures the causal relationship between\ndiseases/procedures and medications through causal discovery and evaluates\ntheir causal effects. Building upon this, CausalMed focuses on analyzing the\nhealth state of patients, capturing the dynamic differences of\ndiseases/procedures in different health states of patients, and transforming\ndiseases/procedures into medications on direct causal relationships.\nUltimately, CausalMed integrates information from longitudinal visits to\nrecommend medication combinations. Extensive experiments on real-world datasets\nshow that our method learns more personalized patient representation and\noutperforms state-of-the-art models in accuracy and safety.\n