A Collaborative Agentic AI Framework for Complex Disease Diagnosis

Abstract

Diagnosing complex and rare diseases demands extensive medical expertise and comprehensive diagnostic processes. This remains a significant challenge in clinical practice, especially in primary care settings where physicians may lack specialized experience. We present a novel collaborative agentic AI framework to enhance the diagnosis of complex medical conditions. Our system comprises four modules: a Disease Proposer, a Related Disease Finder, a Skeptic Agent, and a Diagnosis Agent. The input to the diagnostic workflow is patient data, which may include the chief complaint, symptoms, available clinical data, and test results. Based on this input, the Disease Proposer invokes a Common Disease Agent and a Rare Disease Agent to generate a list of potential diagnoses. The generated list is then sent to the Related Disease Finder, which expands the pool by adding diseases that are commonly confused with those proposed. Subsequently, the Disease Background Agents are invoked to generate detailed information for each disease. The Skeptic Agent verifies all gathered information provided by the Disease Proposer and the Related Disease Finder. The verified information is then fed to the Diagnosis Agent for a final diagnosis. The proposed agentic AI approach shows excellent accuracy in diagnosing both rare and common diseases, compared with foundation models (DeepSeek-V3 and GPT-4o) and the state-of-the-art agentic diagnostic workflow MAI-DxO, demonstrating its potential as an automated diagnostic copilot for physicians, particularly in complex cases requiring simultaneous consideration of both common and rare conditions.

Ähnliche Arbeiten