Abstract: In order to minimize the stochasticity of agents used in disease diagnosis within the dialogue system, and to enable them to interact with users based on the inherent connections between symptoms and diseases, while simultaneously addressing the issue of limited medical data, we propose the Hierarchical Reinforcement Learning with Multi-expert Feedback framework. The framework constructs a reward model in the lower-level networks of the hierarchical structure. Here, the discriminator leveraging the concept of adversarial networks generates rewards by evaluating the authenticity of symptom query sequences generated by the agent, and the large language model of human experts synthesizes various factors to assess the reasonableness of the agent's current symptom queries, thereby guiding the learning of the policy network. The algorithm addresses the deficiencies in data characteristics and improves the policy's capability to leverage feature information, thus making the process of disease diagnosis more aligned with clinical practice. Experimental results demonstrate that the proposed framework achieves diagnostic success rates of 61.5% on synthetic datasets and 84.4% on real-world datasets, while requiring fewer dialogue turns on average. Both metrics surpass those of conventional approaches, further indicating the framework's strong generalization ability.

Shi Li and Xueyao Sun, “Dialogue-Based Disease Diagnosis Using Hierarchical Reinforcement Learning with Multi-Expert Feedback” International Journal of Advanced Computer Science and Applications(IJACSA), 16(2), 2025. http://dx.doi.org/10.14569/IJACSA.2025.0160232

@article{Li2025,
title = {Dialogue-Based Disease Diagnosis Using Hierarchical Reinforcement Learning with Multi-Expert Feedback},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2025.0160232},
url = {http://dx.doi.org/10.14569/IJACSA.2025.0160232},
year = {2025},
publisher = {The Science and Information Organization},
volume = {16},
number = {2},
author = {Shi Li and Xueyao Sun}
}