MAM-AI: Offline AI Assistant Empowers Zanzibar's Nurse-Midwives with Critical Medical Guidance

Maternal and newborn mortality rates remain alarmingly high across sub-Saharan Africa, often exacerbated by a shortage of adequately trained midwifery staff and difficulties accessing up-to-date medical guidelines, particularly in areas with unreliable internet connectivity. To address this critical challenge, researchers have developed MAM-AI, an innovative on-device medical question-answering assistant specifically tailored for nurse-midwives in Zanzibar. This system aims to provide immediate, authoritative guidance at the point of care, even in offline environments.

MAM-AI functions as a retrieval-augmented generation (RAG) system, operating entirely on standard Android devices. It processes user questions by embedding them using a compact 300M parameter model (EmbeddingGemma) and matching them against a curated local corpus of 87 medical guideline documents. The system then generates answers with citations using a 4B parameter generator model (Gemma 4 E4B), ensuring that no query data leaves the device and full functionality is maintained without internet access.

A rigorous evaluation of MAM-AI's deployed configuration revealed key insights. The on-device retrieval component proved highly effective, rivaling cloud-based systems in its ability to locate relevant passages. However, the smaller 4B parameter generator presented a challenge, struggling to simultaneously be helpful and safe. Researchers addressed this by selecting a more faithful, albeit less initially helpful, model and optimizing its performance through prompt redesign, significantly reducing deflection. The study underscored that the quality and comprehensiveness of the local guideline corpus are paramount for generating specific and actionable advice.

While currently a thoroughly evaluated, open-source research prototype rather than a commercial product, MAM-AI holds significant promise for improving healthcare outcomes in African contexts like Zanzibar. By providing essential medical knowledge directly to healthcare providers, it can bridge gaps in training and access to information, particularly in remote or underserved areas. The open-source release of the system, its knowledge base, benchmarks, and evaluation tools encourages further development and adaptation, fostering local innovation in AI-powered healthcare solutions across the continent.