Multi-Agent Clinical Diagnosis with Self-Learned Knowledge for LLM

View the PDF file for the sheet titled MACD: Self -learned Clinical Clinical Diagnosis with the self -learned knowledge of LLM, by Wenliang Li and 12 other authors

PDF view

a summary:LLMS models have shown noticeable possibilities in medical applications, however they face great challenges in dealing with complex clinical diagnoses in the real world using traditional methods. Current immediate engineering and multi -agent curricula usually improve isolated inferences, while neglecting the accumulation of reusable clinical experience. To address this, this study suggests a new framework for multi -agent clinical diagnosis (MACD), which allows LLMS to self -learning clinical knowledge via a multi -agent pipeline summarizes, improves it, and applies diagnostic visions. It reflects how doctors develop experience through experience, allowing a more focused and accurate diagnosis on the main signals of diseases. We extend it to the Macd-Human cooperative course, where many LLM diagnosis factors participate in repetitive consultations, with the support of the evaluator and human control agent for cases where the agreement is not reached. It was evaluated on 4,390 cases of the patient in the real world across seven diseases using various open source LLMS (Llama-3.1/70B, Deepseek-R1-Distill-Lama 70B), MACD works to significantly improve the initial accuracy, and outperforms the clinical guidelines in force with gains of up to 22.3 % (MACD). On the sub -group of data, the performance is equal to or bypassing human doctors (an improvement up to 16 % on the diagnosis only for doctors). In addition, in the Macd-Human workflow, it achieves 18.6 % improvement compared to the diagnosis of doctors only. Moreover, the self -learning knowledge shows strong stability of the model via the model, the ability to transfer, and the customization of the model, while the system can generate the justifications that can be tracked, which enhances the ability to explain. Consequently, this action provides a model for self -developmental self -learning with the help of LLM, and fill the gap between the fundamental knowledge of LLMS and clinical practice in the real world.