Sara: Screening Agents for Rheumatoid Arthritis

Umakanta Maharana, Sarthak Verma, Avarna Agarwal, Prakashini Mruthyunjaya, Sakir Ahmed, Murari Mandal

Keywords: LLM Agents, Rheumatoid Arthritis (RA)

TL;DR: Agentic framework for RA diagnosis

Abstract: Early diagnosis of Rheumatoid Arthritis (RA) remains a critical challenge in healthcare due to its nonspecific early symptoms and reliance on prolonged clinical evaluations, which can delay treatment and worsen patient outcomes. Although Large Language Models (LLMs) show promise in medical applications, their adaptation for specialized diagnostic tasks requires tailored knowledge integration and interpretability---a gap in current AI-driven solutions. In this work, we propose an LLM-based agentic framework SARA, for early screening and diagnosis of RA across diverse clinical stages. We introduce PreRAID (Prescreening Rheumatoid Arthritis Information Database), a real-world dataset comprising data from 160 patients. SARA employs a multi-stage reasoning approach that combines pattern recognition with clinical heuristics to analyze patient symptoms, medical history, and laboratory findings. The PreRAID dataset serves as a contextual knowledge base. The system not only identifies potential RA cases but also generates human-readable explanations for its conclusions, aligning with clinical demands for transparency and accountability in AI-assisted diagnosis. Through rigorous validation on both synthetic and retrospective patient datasets, our framework achieved diagnostic accuracies of up to 95\% and generated explanations deemed actionable in 92\% of cases by both rheumatologists and medical interns. Furthermore, several cross-validation results demonstrate robust performance across diverse patient demographics and clinical presentations, suggesting its potential for widespread implementation. This work demonstrates the viability of LLM agents as scalable, explainable tools for complex diagnostic tasks, especially in resource-constrained healthcare settings where specialist access may be limited.

Submission Number: 7