Demo: Designing a Patient-Centered, Safety-Constrained RAG Chatbot for After-Hours Heart Failure Patient Support

Heart failure (HF) is a growing global health challenge that requires substantial patient engagement for self-management and knowing when to seek care. Limited access to healthcare professionals is associated with higher rates of hospitalization and readmission. Large language models (LLMs) can help close this gap by making guidance more available to patients 24/7, but their use in healthcare raises concerns about hallucination, safety, and practical value to patients. In this project, we introduce a patient-centered, safety-constrained Retrieval-Augmented Generation (RAG) chatbot for HF patients (HF Chatbot). Early evaluation of HF Chatbot with patients and clinicians shows encouraging results: Clarity (M = 2.96/3, AC2 = 0.96), Usefulness (M = 2.84/3, AC2 = 0.83), Accuracy (M = 2.79/3, AC2 = 0.77), and Harm Avoidance (M = 2.79/3, AC2 = 0.79), all with substantial inter-rater agreement. These findings underscore the value of pairing technical safety mechanisms with patient- and clinician-informed design in building trustworthy, patient-facing healthcare AI.