An Adaptive Multi-Agent LLM-Based Clinical Decision Support System Integrating Biomedical RAG and Web Intelligence

Abstract

Increasing data complexity in clinical decision-making processes hinders physicians' ability to make rapid and accurate decisions. This study proposes an innovative solution to this problem by designing a multi-layered, adaptive Clinical Decision Support System (CDSS) comprising interacting large language model (LLM) agents. The proposed system performs semantic-level information retrieval using a BioBERT-based vector database, enhances information retrieval by accessing up-to-date medical resources via the web, and restructures outputs by activating an adaptive optimization loop in low-confidence situations. Through the structuring of clinical texts, cross-validation of symptom analyses with literature and internet sources, and collaborative data fusion among agents, the system integrates multi-source data and produces consistent decisions. In experiments conducted on the MedQA, PubMedQA, and MedBullets datasets, the system achieved accuracies of 94%, 88%, and 84%, respectively, representing substantial improvements over state-of-the-art methods and demonstrating the significance of the proposed architecture for clinical decision-making reliability. This framework is not merely an information retrieval engine; it is a clinical intelligence partner designed to learn, actively contribute to the decision process, and focus on reliability. In contrast to current CDSS protocols, which frequently depend on static modules or single-agent models, our architecture tackles some of the shortcomings in timeliness, multi-source evidence fusion, and confidence calibration. This originality enables the system to be a next-generation clinical intelligence partner by enabling an unprecedented level of transparency, customizability, and adaptability in real-world decision-making processes.