CausalMed: Causality-Based Personalized Medication Recommendation Centered on Patient health state

TL;DR

CausalMed tackles insufficient personalization in medication recommendations by shifting from co-occurrence-based mappings to direct, causal disease/procedure–medication links learned via causal discovery. It introduces Dynamic Self-Adaptive Attention to model how health state modulates the role of diseases and procedures, and integrates these signals with a longitudinal EHR framework through RGCN-based relation learning. The approach yields point-to-point causal relationships, personalized patient representations, and safer medication recommendations, validated by extensive experiments on real-world MIMIC datasets that exceed state-of-the-art baselines. The work provides a transparent, interpretable pipeline with strong performance and safety gains, and releases its code publicly.

Abstract

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

Figures (6)

• Figure 1: The problem of insufficient personalization of medication recommendation.
• Figure 2: CausalMed framework: The upper section illustrates the process of learning the patient's representation from a single clinical visit. The lower section represents the integration of information from multiple visits and the prediction of medication combinations.
• Figure 3: An example of causal discovery, discovering the true relationship between $d_1$ and $d_2$ and eliminating the false relationship between $d_2$ and $m_1$, ultimately transforming set-to-set relationships into point-to-point relationships.
• Figure 4: Using the classifier within DSA, we analyze the causal position of each node in the causal graph $G^d_{v_t}$ to gauge its influence on the patient's current state. For instance, if $d_3$ is identified as a causal disease, it is classified as ${D}^1_t$.
• Figure 5: The performance of co-occurrence-based and causality-based methods on Jaccard.