Dual-Granularity Medication Recommendation Based on Causal Inference

Project completion date: 2024-3
Final publication date: unknown

Medication recommendation aims to integrate patients’ long-term health records, recommending accuracy and safe medication combinations for specific health status. Although existing research makes significant strides from the perspectives of relationships between medical entities level or their molecular structures level, it does not consider combining these two different granularities to express the complex relationships between medications and diseases. To address this challenge, we propose a novel medication recommendation framework, CIDGMed. It utilizes causal inference methods to merge information from both entity and molecular levels, facilitating a complementary interplay of dual-granularity information. Specifically, we first use causal inference to learn and quantify the relationships between medications and diseases. Then, we align medication and disease embeddings in the molecular space to address heterogeneity issues. Finally, we integrate the patient’s longitudinal records to generate personalized representations and recommend medication combinations based on the current health status and causal effects between entities. Through extensive experimentation, CIDGMed outperforms the current state-of-the-art models on multiple metrics, with an increase in accuracy by 4.40%, a decrease in side effects by 6.14%, and a 47.15% improvement in time efficiency. Additionally, we demonstrate the interpretability of CIDGMed through a case study.

Recommended citation: Liang S, Li X, Li C, et al. Dual-Granularity Medication Recommendation Based on Causal Inference[J]. arXiv preprint arXiv:2403.00880, 2024. https://arxiv.org/abs/2403.00880