An Interactive Dialogue Agent to Assist African American and Hispanic/Latino Heart Failure Patients with Self-Care Needs

Introduction Heart failure (HF) is a public health problem, and self-care remains poor, especially among minorities. Access to personalized educational material via an interactive dialogue agent (DA), a system intended to converse with humans in natural language, has the potential to improve self-care. This abstract outlines the initial steps of developing an artificial intelligence-based DA to assist African American (AA) and Hispanic/Latino (H/L) HF patients with their self-care needs. Hypothesis Analysis of HF education sessions between health educators (HE) and patients can provide insight into the topics that AA and H/L patients value. Method In this IRB approved pilot study, we have recorded, transcribed, and verified 20 (18 AA and 2 H/L) HF education sessions between HE and HF patients. Latent Dirichlet Allocation (LDA), a topic modeling algorithm and Empath, a text analysis tool were used to identify common discussion topics in the transcripts. An initiative analysis was performed to identify conversation drivers where each turn (unit of speech by a single speaker, without interruption from the other speaker) was classified as either a question (ending with ‘?’), a prompt (having only filler words like umm, okay) or an assertion/command (others). Lastly, we compared the transcripts against a HF ontology published by the National Center for Biomedical Ontology and the Consumer Health Vocabulary (CHV) available in the Unified Medical Language System (UMLS) to identify the term overlap. Results On average, HE took 117 turns comprising 205 sentences and 2281 words per conversation, whereas patients took 108 turns comprising 131 sentences and 850 words. Per conversation, HE asked 26 questions and had 17 prompts as opposed to 3 questions and 39 prompts by the patients. LDA identified HF and heart function, effects of HF, low salt diet, and follow-up appointment and medication as the top 4 most common topics discussed by the HE. When looking at the entire transcripts, Empath identified eating, health, and cooking as the most common topics for both the HE and the patients. Patients frequently discussed children and family, whereas HE focused on providing HF information as indicated by LDA. Lastly, only 2.1% of HE terms overlapped with the HF ontology. Our analysis revealed that 25% and 22% of the terms (without stop words such as at, the) used by the patients and HE respectively match with the ‘preferred label’ in the CHV. For both, the high frequency terms included heart, heart failure, salt, water, and fluid. This also correlated with our topic analysis findings from Empath and LDA. Conclusion Our analysis helps triangulate the kind of information HE and HF patients value. Though mostly HE dominated the conversation, it is essential to incorporate the topics that patients brought up into the DA and use culturally sensitive vocabulary while communicating them. Next, we will collect more recordings, conduct focus groups, and evaluate the usability of our prototype DA.