AI portable ultrasound lets midwives and other non-specialists run simple pregnancy scans, and in studies AI has dated pregnancies from novice “blind sweeps” about as accurately as trained sonographers. But a large trial of routine antenatal ultrasound in five low- and middle-income countries showed that scanning alone saves no lives without working referral.
A midwife in a rural clinic places a handheld ultrasound probe on a woman’s abdomen and moves it in a few simple, standardised sweeps, up and down, side to side, without trying to find any particular view. She could not interpret the images herself; she has never been trained as a sonographer. Seconds later, an app on her phone tells her the estimated gestational age, whether the baby is head-down, and whether the woman is carrying twins.
That scene is no longer hypothetical in research settings, and it is one of the more genuinely exciting uses of AI in maternal health. Skilled obstetric ultrasound is one of the scarcest resources in many health systems. Dating a pregnancy accurately, finding twins and malpresentation, and spotting a low-lying placenta all shape where and how a woman should give birth. If a midwife with a cheap probe and an algorithm can do much of that, access expands dramatically.
The argument of this piece is that the technology is closer to ready than the systems around it. The evidence that AI can read novice scans is increasingly good. The evidence that scanning improves outcomes without a referral path that actually works is not just weak; there is a large trial showing it does not.
How accurate is AI ultrasound in the hands of non-specialists?
The most cited evidence comes from blind-sweep studies, where operators follow a fixed sweep protocol rather than acquiring standard diagnostic views.
In 2022, Pokaprakarn and colleagues reported in NEJM Evidence on an AI model trained on blind-sweep videos from 4,695 pregnant volunteers in North Carolina and Zambia. On the main test set, the model’s mean absolute error in estimating gestational age was 3.9 days, against 4.7 days for standard fetal biometry by trained sonographers. Crucially, in a Zambian cohort where sweeps were collected by untrained users on low-cost devices, the AI’s error was 4.9 days against 5.4 days for biometry.
The same year, a Google Health team led by Gomes described in Communications Medicine a mobile-optimised system for gestational age and fetal malpresentation, tested with sonographers and novices in the United States and Zambia. It performed comparably to standard biometry for dating and detected malpresentation with very high discrimination, with the model running on a phone in under three seconds.
Other groups have extended the approach to abdominal circumference, placental location and other measurements, using data from countries including Sierra Leone and Tanzania. A 2025 meta-analysis in Frontiers in Global Women’s Health pooled AI gestational age studies and found blind-sweep video approaches performing well against conventional imaging.
These are strong results for a narrow set of well-defined tasks. They do not show that AI can replace a detailed anomaly scan, assess fetal growth restriction reliably in every setting, or diagnose complex placental problems. And most of the data come from carefully run research studies, not routine clinics with patchy connectivity, battery problems and staff turnover.
Does routine ultrasound in pregnancy improve outcomes in low-resource settings?
This is the question that matters, and we have an unusually direct answer from before AI entered the picture.
The First Look trial, published in BJOG in 2018, was a cluster randomised trial across the Democratic Republic of Congo, Guatemala, Kenya, Pakistan and Zambia. In 28 intervention clusters, with 24,263 births, antenatal care was supplemented with two ultrasound scans and referral for complications. The 28 control clusters, with 23,160 births, received standard care. Sonographers were trained for the study.
The result was sobering. The composite of maternal death, maternal near-miss, stillbirth and neonatal death was not reduced, with a risk ratio of 1.09. Antenatal care attendance did not increase. Most tellingly, hospital delivery for women with complicated pregnancies did not increase either. The authors concluded that despite ultrasound being available, neither antenatal care use nor hospital delivery for complicated pregnancies went up, and outcomes did not improve.
In other words, the scans found problems, and the system did not convert findings into different births. A woman told her baby is breech still needs transport to a facility that can perform a caesarean, the money to get there, family agreement to go, and a facility with a surgeon, anaesthesia and blood on the day she arrives.
AI does not change any of that. It changes who can do the scan. That matters, but it attacks the part of the chain First Look suggests was not the binding constraint.
Why could AI ultrasound still help where First Look did not?
It would be too pessimistic to stop there, and the strongest counterargument is worth taking seriously.
First, cost and scale. First Look required training sonographers, a scarce and expensive investment. AI-assisted blind sweeps can be learned by midwives in a short training session, and handheld probes are far cheaper than cart-based machines. Even if outcomes do not improve automatically, lowering the cost of the information makes it more feasible to pair scanning with investment in referral.
Second, accurate dating has value beyond acute referral. Many women in low-resource settings do not know their last menstrual period reliably. Accurate gestational age affects decisions on preterm labour management, the timing of induction for post-term pregnancy, and interpretation of fetal growth. It also underpins data on preterm birth rates that health systems need for planning.
Third, the WHO’s 2016 antenatal care recommendations already call for one ultrasound scan before 24 weeks of gestation to estimate gestational age, detect fetal anomalies and multiple pregnancies, and improve the pregnancy experience. Many countries cannot meet that standard with sonographers alone. AI-enabled scanning by midwives may be the only realistic way to do so.
These points are fair. They support deploying AI ultrasound, but they reinforce the conditional: it helps if the system acts on what it finds. That is the difference between a better test and a better outcome.
A worked example: a district that adds AI scanning
Hypothetical, with round numbers for illustration. A rural district has 20,000 births a year and one sonographer at the district hospital. It equips 40 health centres with handheld AI ultrasound and trains midwives in blind sweeps.
In the first year, most women who attend antenatal care before 24 weeks get a dating scan for the first time. The AI flags roughly 3–4% of pregnancies as breech in late pregnancy and around 1–2% as twins, which in a district this size means several hundred women a year identified for planned facility birth.
Now follow those women. If the district has also set up a maternity waiting home at the hospital, a transport voucher and a phone line to the obstetric team, a large share can be admitted before labour. If none of that exists, many will labour at home or at a health centre with the same risks as before, now with the knowledge that their pregnancy was high risk.
The AI performed the same way in both versions of the district. The outcome depends on the waiting home, the transport and the surgical capacity. A health ministry budgeting for the probes should budget for those first, or at the same time.
What does it take to deploy AI ultrasound safely?
The practical requirements are unglamorous and mostly human.
Training has to cover the sweep protocol, device care, when to repeat a scan, and what to do with each result. Midwives need clear pathways: this finding means refer now, this one means refer before 36 weeks, this one means a repeat scan. Supervision and quality assurance matter, with a sample of scans reviewed remotely by a sonographer where connectivity allows. As we argued in our piece on community health workers, adding tasks to front-line staff without support and pay burns them out.
The devices need power, maintenance and replacement budgets, and the software needs to work offline. Our earlier piece on digital health in low-bandwidth settings made the same point: the connection is rarely the missing piece; the person and the process usually are.
Regulation matters too. AI-enabled ultrasound devices should be approved for the specific tasks and user groups they are marketed for. A model validated for gestational age in the second trimester should not be quietly used to screen for anomalies.
What this means for ministries, funders and clinicians
For ministries of health, treat AI ultrasound as part of a referral programme, not a device purchase. Map every finding the tool can produce to a named action and a named facility. Measure success by facility births among women with flagged complications, not by the number of scans performed.
For funders and implementers, the First Look result is the key design constraint. Fund referral transport, maternity waiting homes and surgical capacity alongside probes, and evaluate with outcomes that matter, not scan counts. Randomised or well-controlled evaluations of AI ultrasound plus referral would fill a real evidence gap.
For midwives and clinicians, the tool extends what you can see; it does not change your clinical judgment about a woman in front of you. Know what the device is and is not validated for, and escalate uncertain results.
For developers, design for offline use, low power and simple outputs linked to actions. Publish performance by operator type, device and population, and support independent evaluation.
For pregnant women, a scan is one part of antenatal care. If a scan shows a finding such as breech or twins, discuss with your midwife or doctor where and how it is safest to give birth.
Frequently asked questions
Can AI ultrasound be used by midwives without sonography training?
Research suggests yes for specific tasks. In a 2022 NEJM Evidence study, an AI model estimated gestational age from blind sweeps by untrained users on low-cost devices in Zambia with a mean error of 4.9 days, similar to 5.4 days for standard measurement by sonographers. Training in the sweep protocol and clear referral pathways are still required.
What can AI portable ultrasound detect in pregnancy?
Validated tasks in research include estimating gestational age, detecting whether the baby is head-down or breech, identifying twins and, in some studies, measuring fetal size and locating the placenta. AI systems are not a substitute for a detailed anomaly scan by a trained specialist, and each tool should be used only for the tasks it is approved for.
Does routine ultrasound reduce maternal and newborn deaths?
Not on its own. The First Look cluster randomised trial in five low- and middle-income countries, published in 2018, added two antenatal scans and referral for complications, but found no reduction in maternal or newborn deaths, and hospital delivery for complicated pregnancies did not increase. Benefits depend on acting on scan findings.
Why does the WHO recommend an early ultrasound in pregnancy?
WHO’s 2016 antenatal care recommendations include one ultrasound before 24 weeks to estimate gestational age, detect anomalies and multiple pregnancies, and improve women’s experience of care. Accurate dating helps decisions about preterm and post-term pregnancy. AI-enabled scanning by midwives could help more countries meet this standard.
The most important finding in AI obstetric ultrasound is not a mean error in days. It is the trial that came before it, showing that seeing a problem is not the same as reaching care. The algorithm can now see. The work is building the road.
If you work in maternity care in a low-resource setting, what happens today after a scan finds a breech baby or twins, and what would need to change for that finding to save a life?
