JMIR Publications

ABSTRACT

Background:

Heart failure (HF) is a chronic condition managed largely between clinical visits, and consistent self-management is a major determinant of hospitalization risk, functional status, and quality of life. Digital health tools—including wearables, apps, and remote monitoring—hold promise for supporting consistent self-management in the contexts of patients’ everyday lives, and large language model (LLM)-based chatbots extend this promise further by generating real-time, individualized content tailored to a patient's questions, health status, and needs. However, the design and evaluation of chatbots for cardiovascular disease remain nascent, and existing chatbots' design and architecture are often reported in insufficient depth to support consensus on future design directions.

Objective:

This pilot study aimed to evaluate feasibility, engagement patterns, and patient experience of a multi-agent, psychosocially personalized AI chatbot designed to support HF self-management.

Methods:

Fifteen participants with HF, recruited from a parent digital health study, were assigned one of three psychosocial phenotypes derived from prior work (Engaged Self-Regulators, Activated Learners, Challenged Survivors) and interacted with the chatbot over a two-week period. The chatbot, a modular multi-agent system routing patient messages to six specialist sub-agents, personalized tone, content, and complexity to each participant. Engagement was assessed via backend logs and usability was assessed via a post-intervention survey. Overall perceptions and experiences were assessed qualitatively through semi-structured interviews and analyzed using reflexive thematic analysis.

Results:

Thirteen of 15 participants interacted with the chatbot and completed the study period. Participants engaged with the chatbot for a mean of 3.2 (SD 1.8) days, 22.1 (SD 14.3) messages, and 33.0 (SD 26.2) total minutes. Activated Learners showed the highest engagement (4.0 days, 31.5 messages), followed by Engaged Self-Regulators (3.3 days, 19.7 messages), with Challenged Survivors lowest (1.5 days, 11.5 messages). On post-intervention surveys, 11/13 (84.6%) of participants rated the chatbot's self-care suggestions a 4 or 5 out of 5 (with 5 indicating “excellent”), with a mean rating of 4.08 (SD 0.64), and 9/13 (69.2%) rated its support for decision-making around healthy self-care behaviors a 4 or 5, with a mean rating of 3.85 (SD 0.90). Qualitative interviews (n=14) found the chatbot to be approachable, accurate, and supportive of behavior change for several participants, particularly Activated Learners and Challenged Survivors, while Engaged Self-Regulators described a ceiling effect in which the chatbot added little beyond their existing self-management routines. Participants also identified limitations including passivity, limited cross-session memory, and gaps in holistic, multi-condition guidance.

Conclusions:

A multi-agent, phenotype-personalized chatbot was feasible, usable, and well-received for HF self-management, with the clearest benefit among participants with greater unmet self-management needs. Findings point to specific, addressable design gaps—including memory, supporting sustained engagement, and context-aware reasoning across a user’s full health picture—and support further testing of psychosocial personalization in larger, more diverse samples.