MAM-AI: An Offline AI Assistant Boosts Midwifery Care in Zanzibar

Researchers have developed and evaluated MAM-AI, an innovative on-device medical question-answering system designed to support nurse-midwives in Zanzibar. This initiative directly addresses the critical issue of high maternal and newborn mortality rates in sub-Saharan Africa, where healthcare providers often lack adequate midwifery training and face challenges accessing up-to-date medical guidelines due to poor internet connectivity.

MAM-AI functions entirely offline on standard Android devices. It utilizes a compact AI model (EmbeddingGemma 300M for retrieval and Gemma 4 E4B for generation) to match user questions against a curated local corpus of 87 medical guideline documents. The system then provides cited answers, ensuring that critical medical information is available at the point of care without requiring any internet connection or sending queries off-device.

The evaluation of MAM-AI employed a rigorous, layered methodology, assessing the retriever, generator, and end-to-end system performance. While the on-device retrieval component proved highly effective, rivaling cloud-based systems, the smaller generative model presented a challenge in balancing helpfulness with safety. Researchers managed to overcome this by deploying a more faithful (less prone to dangerous errors) generator and significantly improving its helpfulness through refined prompt engineering.

This open-source research prototype demonstrates the potential of localized AI solutions to bridge critical knowledge gaps in resource-constrained settings. By providing instant, reliable medical guidance, MAM-AI could empower nurse-midwives in Zanzibar and potentially across sub-Saharan Africa, contributing significantly to improved maternal and newborn health outcomes in regions where access to information and specialized training remains a major hurdle.

Although MAM-AI is currently a research prototype and not a commercial product, the release of its system, knowledge base, benchmarks, and evaluation tools provides a valuable foundation for future development and deployment of similar AI-powered healthcare tools tailored for offline environments.