ReCLLaMA: A Reasoning-Centered LLM Agent for Medical Diagnosis

Abstract

Large Language Models (LLMs) have demonstrated impressive capabilities in natural language understanding, yet their application to clinical diagnosis remains constrained by hallucinations, limited interpretability, and the absence of explicit reasoning mechanisms. Recent AI agents extend LLMs beyond passive text generation by enabling multi-step planning, tool use, structured retrieval, and iterative decision making, making them promising for clinical decision support. However, existing agentic systems often still lack transparent and medically grounded reasoning processes.To address these limitations, we propose ReCLLaMA, a Reasoning-Centered LLM Agent for Medical Diagnosis, an agentic neuro-symbolic framework that integrates statistical language models with symbolic inference over structured biomedical knowledge. ReCLLaMA aligns free-text symptom descriptions with standardized medical ontologies using pretrained biomedical encoders and performs logical reasoning over heterogeneous knowledge graphs constructed from health records and pharmacological resources. To bridge representational mismatches across sources, we introduce a statistical entity alignment module based on random forest classification, enabling the construction of a unified knowledge space. Within this space, ReCLLaMA performs deductive and abductive reasoning to generate interpretable diagnostic pathways with calibrated confidence.Our framework advances the integration of subsymbolic and symbolic AI for healthcare, offering a principled approach to traceable and knowledge-grounded clinical decision support. Experiments on real-world datasets demonstrate superiority over black-box LLM baselines and conventional rule-based systems in both diagnostic accuracy and explainability.