Medha Thapliyal

Sheet 03 Writing Revised October 2026

Sheet 03/09. 14 July 2026. 3 minute read.

Novices with a probe

One day of training produced sonographer-level gestational age estimates in Lusaka. The hard question is what happens to the referral chain after the machine says high risk.

03/09 · Essay

The strongest evidence for AI in maternal health in low-resource settings is not a chatbot. It is an ultrasound probe. The FAMLI collaboration between a university in North Carolina and sites in Zambia trained a model on blind-sweep ultrasound from several thousand pregnancies and showed it could estimate gestational age with a mean error of under four days, slightly better than standard biometry done by trained sonographers. A follow-up diagnostic accuracy study put a handheld probe in the hands of novices with one day of training and found their AI-assisted estimates matched credentialed sonographers between fourteen and twenty-seven weeks.

This matters because gestational age is the quiet foundation of almost everything in antenatal care. Without it, you cannot time a referral, interpret a growth measurement, or decide whether a labour is preterm. In many of the districts I have worked in, the dating is a guess based on last menstrual period and the memory of a festival.

Accuracy is the easy part

I believe the accuracy results. What I do not yet believe is that accuracy translates into outcomes, because the evidence stops at the probe. There is still no randomised trial linking an AI ultrasound tool to a change in maternal or neonatal mortality. The studies measure agreement with experts, not what happens next.

And what happens next is the whole problem. In the WINGS work on growth monitoring, we treated a weighing as an implementation outcome rather than a clinical one. The question was never whether the child was weighed. It was whether the weight was recorded, whether the record triggered counselling, whether the counselling led to a referral, and whether the referral was completed. The chain broke in different places in different facilities, and the places it broke had more to do with who was accountable for the handoff than with the accuracy of the scale.

An AI probe is a very accurate scale. It makes the first link in the chain stronger. If the rest of the chain is weak, a stronger first link mostly produces more precisely dated women who still do not reach a facility that can help them.

Designing for the chain

If I were advising a ministry or an NGO considering these tools, I would ask for an implementation study designed around the referral, not the device. Measure how many high-risk flags lead to a completed referral within the window that matters. Measure what the receiving facility does with a referral it did not expect. Measure the workload the tool adds to the nurse who now has a reason to refer and no additional time to do it.

I would also pay close attention to who the novices are. One day of training is a remarkable result, and it is also a recipe for task-shifting without pay-shifting. The person holding the probe will be asked to carry a new clinical responsibility, and that responsibility has to come with authority, supervision, and recognition, or the tool will be quietly abandoned the first time a referral goes wrong.

A better kind of evidence

The ultrasound work shows what good AI evidence in maternal health can look like: prospective, multi-site, with external validation and a willingness to publish the error. It deserves the next step, which is the harder and less photogenic study of what the chain does with the number. Until that study exists, I will keep describing these tools as promising rather than proven, and keep asking the question that implementation science is for.

Sources

03 · Adjoining

Keep reading.

  1. Sept 2026 The 0.9 percent Rapid qualitative analysis appears in less than one percent of implementation science studies. AI may finally make it common. The question is what we trade for the speed.
  2. Aug 2026 The red alert and the ASHA A clinical copilot in Nairobi cut diagnostic errors most when it interrupted. What would an interruption mean for a community health worker with no clinical backstop?
  3. Jun 2026 Silicon mothers Language models can produce plausible answers about pregnancy in Assam. Plausibility is not validity, and formative research cannot be outsourced to a model that has never missed an antenatal visit.