Zambia faces a high cervical cancer burden, with incidence exceeding 65 per 100,000 women. Although preventable with early detection, most primary care facilities lack specialists, pathology services, or reliable HPV screening tests. Midwives using visual inspection with acetic acid (VIA), a low cost cervical cancer screening method, are left to make binary clinical decisions without specialist input, that is, whether to refer a patient or send them home to await confirmatory results that may take months. The consequence is long turnaround times and lost follow-ups, leading to late-stage diagnosis. With screening coverage under 33%, the DawaMom screening tool supports frontline midwives and clinical officers as well as patients at the point of care in low-connectivity settings.
Developed by Dawa Health, the DawaMom VIA module is a 5-class cervical pathology classifier integrated into a mobile app and powered by an EdgeAI kit, a Raspberry Pi 5 and Hailo-8L NPU–based system, for offline clinical inference. VIA images captured by midwives are transmitted over local WiFi to a three-step pipeline:
- ResNet50 Gatekeeper model validates image quality and relevance (filtering out poor-quality or non-cervical inputs)
- The classifier assigns one of the five pathology categories built with MedSigLIP
- A Retrieval-Augmented Generation (RAG) system with Gemini Flash generates structured clinical recommendations for the healthcare worker following World Health Organization (WHO) and Ministry of Health (MOH) Zambia protocols.
MedSigLIP's domain-specific pre-training enabled efficient training of DawaMom's classifier on over 5500 labeled cervicography images and eliminated the need for large-scale annotated training from scratch. It also reduced compute requirement to a single GPU fine-tuning session within 16GB VRAM, compared to the multi-GPU fine-tuning runs required by a previous ResNet50-based approach. DawaMom's full classifier is exported as a full-precision ONNX model and deployed entirely on-device, running inference in approximately 50 milliseconds in native mode and 12 seconds in ONNX fallback mode on the EdgeAI kit. With this low latency, DawaMom's system provides rapid and reliable point-of-care decision support in low-resource and off-grid settings.
The DawaMom VIA module on EdgeAI device has powered the screening of more than 3,500 patients since Dawa Health developed its first-place winning prototype at the Google-sponsored Data Science for Health Ideathon across Africa in late 2025. Additionally, Dawa Health has trained 10 frontline clinicians and 25 community health workers as the last-mile support pipeline into the screening clinics.
"I attended the Greater Horn Oncology Symposium and couldn't stop talking about how fast the DawaMom app's technology is guiding midwives and clinicians during cervical cancer screening"
— Symphorose, Nurse and Media Relations Officer, Zambia Cancer Society
United Nations Population Fund (UNFPA) Zambia, in collaboration with MOH Zambia, has partnered with Dawa Health to conduct a prospective validation of the DawaMom EdgeAI device. For the validation study and in real world deployments that will follow, a human-in-the-loop will be included throughout: the model surfaces a recommendation, providing the percentage probability of each pathology class, explaining the distribution, and outlining next steps according to the WHO and MOH treatment protocols. The trained midwife reviews every case and uses their clinical judgement to decide the final course of action. Any detected lesion requires an HPV PCR test result with follow-up biopsy, if needed.
The rollout, planned for Q3-Q4 2026, will be across 78 facilities in the Eastern Province and targeting over 48K women. Under the MOH Zambia regulatory sandbox, the real-world sensitivity and specificity of DawaMom will be tracked against confirmed outcomes, referral completion, and override rates by the midwife.
Dawa Health is also conducting generalizability testing and fine-tuning of a dataset of 33K cervicography images from Zambia, Zimbabwe, and Rwanda to power expansion across 12 mobile clinics in Zimbabwe. Technically, the description generation is migrating from the RAG system built on Gemini Flash to a quantized 4-bit MedGemma model to enable 100% on-device inference. This transition is aimed at reducing the system latency in low-connectivity environments while managing costs and keeping sensitive patient data entirely localized to the device. Finally, Dawa Health plans to use open-source Rockchip AI Boxes for cost-effective edge inference at scale.
DawaMom Focus Group Discussion with SRH users of the chatbot, Zambia (© 2026, Dawa Health)