Hi all, Nate here, back from a vacation traveling in the highlands of Ecuador (Incredible nature, strong indigenous culture, and the crime problems you hear about in the news are really only in coastal areas).
I’ve just done an update to the mapping of AI in CHW programs in LMICs that I first shared in December last year. This version expands the map from 38 to 44 programs. Here’s a summary of what I found.
The programs are concentrated in Sub-Saharan Africa and South Asia: Sub-Saharan Africa has 25 programs and South Asia has 17. Latin America has 5 (Brazil and Guatemala) and Southeast Asia has 2 (both in Indonesia). The countries with the most programs are India, with 11 programs, Kenya with 8, and Rwanda with 6, with Bangladesh, Nigeria, Tanzania, and Uganda at 4 each. I found nothing in the Middle East or North Africa, Central Asia, the Pacific Islands, and almost nothing in Francophone or Central Africa (Chad is the only Central-African entry).
Generative AI the most common technology: 23 of the 44 programs use an LLM, usually a chatbot that answers a health worker's questions from a set of guidelines. That is ahead of traditional machine learning (13 programs) and computer vision (11, used for tasks like reading ultrasound, chest X-rays, or cervical-screening images). Speech and voice interfaces appear in 8. Nineteen programs (43%) combine more than one technique.
Most tools are built to help the worker assess a patient, or to answer the worker's questions: The most common function is a worker-facing question-and-answer assistant. Close behind are clinical decision support and triage, and diagnostic detection or screening from an image, test, or sound. Smaller groups do patient risk prediction, data digitization, worker training, visit prioritization, disease surveillance, program-operations prediction such as worker dropout or supply needs, and language translation.
It is overwhelmingly applied in MNCH programs: Roughly two-thirds of programs address maternal, newborn, or child health. Every other health area is small: non-communicable diseases, TB, malaria, HIV, chronic respiratory, and antimicrobial resistance. Family planning appears in only a couple of programs, and mental health in none.
This is still mostly pilots: Only 11 of 44 programs claim national scale. Counting pilots, research studies, pre-launch projects, and early prototypes together, about 18 programs (41%) are at pilot stage or earlier. "National scaling" is self-reported, and on a close read several of those claims describe a single district or a funded plan to expand, not an actual national footprint.
Few programs are built into a government health system: Around 20 of the 44 programs run entirely outside government, developed by NGOs, startups, or research teams with no state role. Only a handful are genuinely embedded in a national system, and even the strongest cases tend to have government as a partner rather than the owner: Rwanda delivers its community-health platform through the Rwanda Biomedical Centre, Zanzibar's Jamii ni Afya program is government-owned, and Uganda's mosquito-surveillance tool runs with the Ministry of Health. The rest sit in between, piloting with a Ministry of Health or using government health workers without being built into the system.
The evidence base is thin: Only a handful of programs have been rigorously evaluated, and the results are mixed. A World Bank trial of an LLM assistant at EHA Clinics in Nigeria found no significant improvement in care and concluded the tool "is not yet a public health priority in low- and middle-income countries." PATH's real-world trial in Rwanda found that health workers' own referral decisions were already right 97.9% of the time; the best language model came close but none beat them. Two non-chatbot tools did better: safe+natal in Guatemala, which uses computer vision to read blood-pressure monitors and Doppler ultrasound, raised detection of hypertensive disorders in pregnancy in a feasibility trial; and ThinkMD, a decision-support tool, improved CHW adherence to treatment guidelines in a quasi-experimental study. Beyond these, fewer than half of the programs have any peer-reviewed publication, and most of those test accuracy, not whether care improves. About half rest on organizational reports or a website. For now, adoption is running well ahead of the evidence. And as I wrote about in a previous post, the LLM models these trials tested are already obsolete.
The two leading technologies sit at different stages of proof: Computer vision is the more established: the tools that read a test strip, an X-ray, or an ultrasound tend to carry peer-reviewed accuracy validation. That work shows the model reads the image correctly, not yet that it changes outcomes, but it is published and consistent. Language models are newer and less proven. They are the most-adopted category, with 23 of the 44 programs, but almost all are pilots or pre-launch, and their only two rigorous trials disappointed.
Basic information is often hard to pin down: For many of these programs, the available information is thin. Some describe their tool only as "AI," with no named model, so you cannot tell what is running. Others report scale that does not hold up: a program that calls itself national turns out to be one district, or counts workers trained rather than workers using the tool. A few are visible only through a press release or a social-media post, making it hard to confirm the program is even live.
Geographic and health area clustering impacts the evidence base: The clustering in India, Kenya, and Rwanda, and in maternal and child health, reflects where the funding, the research partnerships, and the community-health workload already sit. It also means the evidence base is being built in a few places, on a few problems. Family planning, mental health, and most of the world outside East Africa and South Asia are close to blank. (By the way, I’ve been awarded a Fulbright research grant to study AI in CHW programs in India, Kenya, and Rwanda, so stay tuned for more on that.)
You can download the mapping spreadsheet below.