ºÃÉ«ÏÈÉúTV

Building a Maternal Health Chatbot

Onboarding pregnant women to the askNivi platform at a Marwari hospitalAI-SDM is helping to transform maternal healthcare delivery in underserved regions through the integration of large language models (LLMs) and advanced decision-making strategies. Partnering with askNivi—which has rolled out a WhatsApp-based interface serving pregnant women and their families in India—researchers are addressing a critical gap in accessible, personalized medical information. The core of this research involves transitioning from traditional, scripted interactions to a sophisticated generative framework capable of interpreting diverse, open-ended maternal health queries. This advancement is not merely a technological upgrade, but a strategic effort to apply human-centric AI to high-stakes societal health decisions—ensuring that every user receives contextually relevant and medically sound guidance, while also deepening our understanding of how AI systems shape real-world health decision-making.

Stage-aware RAG system architecture for maternal health support

Stage-dependent escalation logic in maternal health support with clinically distinct risk thresholdsTo safely deploy generative AI in a high-stakes medical context, the team developed a stage-aware Retrieval-Augmented Generation (RAG) architecture tailored specifically to maternal and newborn care. Rather than relying solely on the model’s internal training data, the system grounds every response in curated public health guidelines and structured symptom knowledge. A pre-generation safety triage layer evaluates maternal stage and reported symptoms before producing a response, conservatively routing high-risk queries to expert-written escalation templates and allowing lower-risk informational queries to proceed through evidence-grounded generation. This design ensures that identical symptoms may trigger different responses depending on pregnancy, postpartum, or newborn stage—reflecting clinically meaningful risk distinctions.

To better understand the contribution of retrieval and safety triage, we evaluated three system configurations: a baseline off-the-shelf LLM, a standard retrieval-augmented generation (RAG) model, and a stage-aware RAG variant incorporating safety triage. Each system was assessed using structured quality and safety metrics, including factual correctness, emergency detection, and guardrail adherence. The comparison illustrates how retrieval improves grounding, and how stage-aware triage further strengthens safety-related performance.

Performance of baseline LLM, standard RAG, and stage-aware RAG across selected quality and safety metrics (lower is better)

Beyond the computational linguistic challenges, the AI-SDM team is focusing on the behavioral dimensions of technology adoption by investigating patterns of user attrition, ensuring the platform remains a reliable and effective tool for maternal health. By analyzing historical data and developing quantitative metrics to track how and why users interact with the platform, the project seeks to understand how conversational design influences trust and sustained use. In addition, the team is seeking to determine measurable gains in health knowledge and the adoption of positive health behaviors throughout the user journey. The current research phase is moving toward a rigorous A/B testing deployment, where users are randomized between the original and enhanced platforms to evaluate differences in engagement patterns and self-reported health behaviors. The team is also developing clarification and personalization strategies to reduce ambiguity and improve response relevance across repeated interactions. This holistic approach integrates NLP capabilities with decision science to bridge health equity gaps, providing a scalable model for localized, culturally specific AI interventions in global public health.