RDguru: A Conversational Intelligent Agent for Rare Diseases

Aim:

              To design an advanced conversational diagnostic system (RDguru++) that improves rare disease diagnosis accuracy, interpretability, and response speed by integrating Groq LLM, Double Deep Q-Network (DDQN), Adaptive Knowledge Retrieval (AKR), and an Explainable Diagnosis Engine (XDE).

Abstract:

        The proposed RDguru++ is an enhanced conversational agent for rare disease diagnosis that advances beyond the original RDguru system. It integrates Groq’s low-latency large language model (LLM) with a Double Deep Q-Network (DDQN) to achieve faster, more stable, and explainable diagnostic reasoning. Unlike the earlier GPT-3.5-based model with a single DQN, RDguru++ introduces two major innovations: Adaptive Knowledge Retrieval (AKR) and the Explainable Diagnosis Engine (XDE). AKR dynamically selects the most relevant data source—such as Orphanet, OMIM, or GARD—based on the query type, ensuring precision and efficiency in information access. The DDQN-based fusion model integrates outputs from PheLR, Groq-LLM reasoning, and phenotype matching to reduce overestimation bias and enhance diagnostic accuracy. Meanwhile, XDE produces interpretable reasoning summaries with evidence-traceable citations. Together, these modules create a real-time, transparent diagnostic workflow. RDguru++ demonstrates higher reliability, improved interpretability, and a projected top-5 recall above 70%, establishing a faster, more explainable, and clinically trustworthy AI assistant for rare disease consultation

Proposed System:        
The proposed RDguru++ is an upgraded conversational diagnostic system that enhances the original RDguru through integration of Groq’s low-latency LLM and a Double Deep Q-Network (DDQN) for robust and stable learning. It introduces two new modules: Adaptive Knowledge Retrieval (AKR), which dynamically selects the most relevant data source (Orphanet, OMIM, GARD, or Orphadata) based on query type, and the Explainable Diagnosis Engine (XDE), which generates clear reasoning summaries with supporting citations. Clinical text is converted into HPO terms and analyzed by three reasoning components—PheLR, Groq-LLM reasoning, and phenotype matching—whose outputs are fused by the DDQN model to reduce bias and improve diagnostic accuracy. These enhancements enable real-time, evidence-based, and interpretable rare disease diagnosis. The proposed architecture achieves faster inference, adaptive knowledge selection, and higher diagnostic reliability, providing clinicians with transparent and traceable decision support that significantly improves performance compared to the original RDguru system.

Advantage:

• Faster Inference: Groq LLM provides real-time response generation with minimal latency.
• Improved Stability: DDQN minimizes Q-value overestimation, ensuring reliable learning.
• Adaptive Retrieval: AKR dynamically selects the most relevant medical database per query.
• Higher Accuracy: Multi-source fusion improves diagnostic precision and recall rate.
• Explainable Output: XDE provides transparent, evidence-supported diagnostic reasoning.
• Clinical Reliability: The system delivers consistent, traceable, and trustworthy results for practitioners.