Knowledge-Driven Hierarchical Reinforcement Learning Framework for Automated Disease Diagnosis in Dialogue Systems

This paper presents a novel Knowledge-Driven Hierarchical Reinforcement Learning (KHRL) framework for automated disease diagnosis in dialogue systems. The framework integrates Hierarchical Deep Q-Networks (h-DQN) with structured medical knowledge from ontologies such as SNOMED CT and ICD, enabling more interpretable and medically sound diagnostic interactions. Unlike conventional models, KHRL employs a two-tiered policy structure-highlevel policies orchestrate diagnostic tasks, while low-level policies manage symptom-specific dialogue actions. A hybrid reward function is designed to simultaneously optimize diagnostic accuracy, dialogue efficiency, and user satisfaction. Experimental evaluations using synthetic patient datasets modeled after real-world clinical scenarios show that the KHRL framework outperforms both rule-based systems and flat reinforcement learning models. Specifically, it achieved a diagnostic accuracy of 91.6 %, reduced average dialogue turns to 8.7, and recorded a user satisfaction score of 4.8 out of 5reflecting an improvement of over 9 % in diagnostic accuracy compared to baseline methods. This work contributes to the body of knowledge by demonstrating how hierarchical structures combined with domain-specific knowledge can significantly enhance decision-making in healthcare dialogue systems. The proposed framework is scalable, interpretable, and adaptable, laying the groundwork for future real-world deployment in intelligent, patient-centric telemedicine platforms.