CoMed: An Intelligent Analysis Framework for Drug Co-medication Risks Based on Large Language Models

CoMed is an AI framework integrating Retrieval-Augmented Generation (RAG), Chain-of-Thought (CoT) reasoning, and multi-agent collaboration. It can automatically search medical literature, analyze drug interactions, and generate detailed risk assessment reports, providing intelligent support for clinical research and drug safety analysis. Original author/maintainer: studentiz; Source platform: GitHub; Original title: "CoMed: Comprehensive framework for analyzing drug co-medication risks using Chain-of-Thought reasoning and LLMs"; Release date: June 2026; DOI: 10.1021/acs.jmedchem.5c03511; Original link: https://github.com/studentiz/comed.

In modern medical practice, patients often need to take multiple drugs simultaneously to treat complex diseases, but polypharmacy carries potential risks of drug interactions (e.g., reduced efficacy, enhanced toxic side effects). Traditional analysis relies on doctors' experience and limited databases, making it difficult to track massive literature; the mechanisms of drug interactions are complex (pharmacokinetic and pharmacodynamic processes). AI (especially LLMs) has the potential to solve this problem, but there are risks such as hallucinations, outdated knowledge, and insufficient interpretability. The CoMed framework is designed to address these challenges.

CoMed innovatively integrates multiple AI technologies:

1. Retrieval-Augmented Generation (RAG): After users input a drug combination, it automatically constructs PubMed search queries, retrieves and filters relevant literature, and provides it as context to the analysis module. This solves the issues of LLM knowledge timeliness and hallucinations, and conclusions can be traced back to literature sources.
2. Chain-of-Thought (CoT) Reasoning: Analyzes problems step by step (e.g., pharmacological mechanisms, overlapping metabolic pathways, clinical cases, etc.) to improve accuracy and interpretability.
3. Multi-agent Collaboration: Includes risk analysis, safety assessment, and clinical recommendation agents, which collaborate through consensus or debate modes to generate comprehensive risk assessment reports.

CoMed adopts a modular architecture:

• RAG Module (rag.py): Responsible for medical literature retrieval, relevance scoring, and statistical analysis, supporting custom parameters.
• CoT Module (cot.py): Implements Chain-of-Thought reasoning, decomposing complex problems into sub-problems.
• Multi-agent Module (agents.py): Implements professional agents and collaboration protocols.
• Core Module (core.py): Component integration, configuration management, and result aggregation. Workflow: Drug combination input → Literature retrieval → Correlation analysis → Risk assessment → Report generation. Each stage has verification and quality control mechanisms.

CoMed provides a Python API:

• Configure environment variables for LLMs (e.g., GPT-4o), initialize drug lists, run full analysis or execute step-by-step (search, correlation analysis, risk assessment, report generation), supporting chain calls. Supported scenarios:
• Analysis of drug combinations for cardiovascular diseases (e.g., warfarin, aspirin, etc.).
• Evaluation of drug combinations for diabetes (e.g., metformin, insulin, etc.).
• Incremental analysis (dynamically add drugs to update results).
• Batch processing (analysis of multiple drug combinations).

CoMed is positioned as a research and educational tool, targeting clinical researchers, medical professionals, and medical students. It is not suitable for direct patient care, should not be the sole basis for clinical decisions, and cannot replace professional medical judgment. Using the software implies understanding its research nature; clinical decisions should consult qualified medical professionals.

As an AI tool, CoMed helps professionals process information efficiently, accelerate literature reviews, and assist in drug safety research and clinical decision support. In the future, it can integrate more data sources (electronic health records, genomic data, real-world evidence) to provide more personalized and precise assessments. It is necessary to clarify the tool's boundaries, maintain a human-machine collaboration model, and promote the healthy development of medical AI.