Personalized Medication for Chronic Diseases Using Multimodal Data‐Driven Chain‐of‐Decisions

The precise matching of medication regimens to individual patients, known as personalized medication, is critical for the effective management of chronic diseases. Traditional machine learning-based models for personalized medication regimens typically rely solely on either clinical macro-phenotypes or molecular-level drug characteristics. It remains challenging to capture the patient-medication relationship from a comprehensive perspective that integrates individual patient characteristics with macro- and micro-level properties of the medication. Determining patient-medication relationships constitutes a three-stage sequential decision process from a clinical decision-making perspective. Therefore, inspired by Chain-of-Thought prompting, which simulates the decision-making process of human experts, a Multimodal Data-Driven Chain-of-Decisions (MDD-CoD) framework is proposed, where three-stage deep learning tasks are sequentially organized to reflect upstream-downstream logical dependencies, thereby forming a coherent clinical decision-making process. The model incorporates multimodal clinical phenotype data, multi-attribute medication data, and insights from clinical experts. Performance evaluation of the model involved comprehensive experiments utilizing five datasets covering four chronic diseases sourced from three hospitals. The dataset comprises information from chronic kidney disease (CKD), membranous nephropathy (MN), rheumatoid arthritis (RA), colorectal cancer (CRC), and knee osteoarthritis (KOA), totaling 3173 unimodal, 502 multimodal, and 2187 medication records from 3675 patients. Experimental results demonstrate that the framework achieves enhanced predictive performance in personalized medication decision-making based on individual patient disease characteristics, surpassing the strongest baseline across all tasks. This framework serves as a foundational model for clinical mixed data, with improved generalization and interpretability in cross-disease personalized decision-making tasks. It offers a scalable solution for the implementation of personalized medication regimens for chronic diseases.