Dual-Source Retrieval-Augmented Generation Chatbot for Women's Health (HerCare): Design and Multimethod Evaluation...

Conversational agents for women's health often fail to meet user needs, offering either clinically sterile advice or unreliable peer anecdotes. This limitation creates a tension between the need for factual safety and emotional resonance in sensitive health contexts.

We aimed to address this gap by developing and conducting a formative evaluation of HerCare, a conversational agent built on a novel dual-source retrieval-augmented generation architecture. The system integrates expert medical knowledge with peer narratives and makes the provenance of each response visible to users, enabling trust calibration through transparent source attribution.

We conducted a remote, web-based single-session field study (December 2024 to January 2025; North Dakota State University Institutional Review Board Protocol #IRB0005368) with 243 completers (from 335 eligible, consenting visitors) recruited via social media (Facebook [Meta], Reddit, and Instagram [Meta]) and university mailing lists. Eligible participants self-identified as women aged 18-45 years with English proficiency and internet access. We used a quantitative multimethod evaluation, combining standardized self-report metrics-the Chatbot Usability Questionnaire and net promoter score (NPS)-with computational linguistic analyses (VADER [Valence Aware Dictionary and Sentiment Reasoner] sentiment analysis and NRC [National Research Council] Emotion Lexicon) of 1191 conversational turns.

Among the 243 participants who completed the protocol, reported usability was high (Chatbot Usability Questionnaire median 78.1, IQR 65.2-87.5; mean 75.67, SD 15.50) and advocacy was strong (NPS 60.0; 171/243, 70.4% promoters, 25/243, 10.3% detractors), though this NPS reflects completers only. Postinteraction ratings were high (all facets median 4-5 on a 5-point scale; helpfulness, ease of use, and clarity median 5, IQR 4-5). Computational analysis revealed a consistent polarity shift from neutral to negative user queries (compound -0.18 to +0.15) to strongly positive agent responses (compound +0.55 to +0.83), with a recurring validate-then-redirect empathy pattern in which the agent acknowledges user distress before pivoting to constructive guidance.

Among completers, the dual-source architecture was associated with high perceived empathy and trust, suggesting it can combine clinical accuracy with emotional support. These formative findings indicate the feasibility of weaving clinical sources with lived experiences toward safer, more resonant health AI and surface a candidate design pattern for future empathy-attuned systems that warrants controlled evaluation.