Last Tuesday, I was standing in a cramped exam room with a pregnant patient whose prenatal care had already been interrupted twice by transportation problems and once by a clinic closure. The ultrasound image on the screen was fine in the way that hospital images are fine when everything is easy, which is to say it looked reassuring until I remembered how often the people who most need certainty are the ones least able to get back for repeat scans. She looked at me and said, quietly, “I just need to know the baby is okay.”
I have heard versions of that sentence for years. It lands differently when the room is hot, the machine is temperamental, and everyone in it knows that the next appointment may never happen.
That is where my thinking about AI and the global health mission changed. I used to think global health technology was mostly about scaling expertise outward, exporting what already worked in wealthy systems and trimming the edges for low-resource settings. Then I spent enough time in clinics where the bottleneck was not brilliance but access, and I began to see the problem more clearly. The real question is whether a tool can survive contact with scarcity. That is the test.
What Global Health Actually Demands
Global health has always asked for more than good intentions. It asks whether a diagnostic idea can function when the internet fails, when the power cuts out, when one radiologist is covering a district, when the patient has one bus fare left and six children at home. AI enters that world carrying a lot of hype and a small amount of humility. Only the humility matters.
I think of this as the scarcity test. A tool passes only if it still improves care after you subtract bandwidth, staffing, language concordance, maintenance, and the luxury of second chances. I have seen many promising technologies die here. The ones that survive usually do something boring and essential: they make a human clinician more effective, sooner, with less friction.
That is why the Gates Foundation’s fetal ultrasound AI work matters to me. It points toward a future where a frontline clinician, midwife, or community health worker can obtain a usable answer from a device that does not require a subspecialist in the room. For maternal health, that matters in a way a conference keynote never can. It matters because a single missed estimate, a malpositioned probe, or a delayed referral can change the shape of a pregnancy, and sometimes the shape of a family.
The Promise, and the Part People Miss
The obvious reading is that AI will democratize expertise. The clinic teaches otherwise. AI only helps if it is built for the workflow that already exists, not the one vendors imagine on a polished demo day. In low- and middle-income countries, the workflow is often improvisational, under-resourced, and deeply local. A system that needs perfect inputs and constant cloud connectivity is not equity. It is ornament.
The 2026 review in Health Informatics Journal on deploying AI for global health equity emphasizes readiness, not novelty, and that emphasis is exactly right. Readiness is the difference between a model that looks good in a slide deck and one that can actually be used where maternal mortality is not an abstraction. The review’s framework is persuasive because it asks the unglamorous questions: infrastructure, governance, local adaptation, training, and sustainability. Those are the gates. Most tools never make it through the first one.
Another useful anchor comes from Localized AI for stroke care in LMICs, a 2026 Frontiers in Public Health paper that frames local deployment as a structural response to diagnostic barriers. That paper belongs in every global health conversation because it makes the point clinicians know instinctively: speed matters, but speed without local fit just moves the problem around. Stroke care, maternal ultrasound, hypertensive triage, neonatal risk detection, these are not abstract use cases. They are time-sensitive decisions in places where the system can already be brittle.
In my own practice, I have become much less interested in whether an AI system is impressive and much more interested in whether it can be boringly dependable. I want to know what happens when the patient is late, when the image quality is poor, when the label set is incomplete, when the village clinic has one intermittent nurse and no specialist backup. I want the failure modes named in advance. I want the off-ramps.
Why I Stopped Romanticizing Scale
For a long time, I believed scale itself was the moral good. If a tool could reach more people, I assumed that was enough. Then I saw how often scale rewards the systems that already have data, money, and governance, while the people with the least access become an afterthought in validation studies. Now I think equity requires something harder than scale. It requires design discipline.
The article by Designing AI tools to advance health equity in resource-constrained low- and middle-income countries makes this concrete. It places special attention on local constraints, which is where the work really begins. A model that performs beautifully in a tertiary center can fail in the field for reasons no one charted. Population shift. Device mismatch. Staff turnover. Language differences. Those are not edge cases. They are the center of the problem.
I once watched a supposedly helpful clinical decision aid get abandoned after three weeks because it added one more login screen to an already overloaded workflow. That is the kind of failure clinicians recognize immediately. The software did not break. The human system did. If AI is serious about global health, it has to earn its place in that system.
What the Evidence Is Telling Us
The strongest literature is moving in a consistent direction. In 2026, a review in Nano-Micro Letters on AI-enhanced wearable blood pressure monitoring in resource-limited settings argued for co-design across sensors, models, deployment, and assessment. That matters because blood pressure is a maternal health problem, a cardiovascular problem, and a systems problem all at once. The review notes that current devices are still often judged against AAMI accuracy thresholds, where mean error should be under 5 mmHg and standard deviation under 8 mmHg, yet many systems struggle outside controlled environments. That gap is the story. Not the algorithm. The gap.
Maternal and neonatal care is where the stakes become visceral. The 2026 review in The Lancet Infectious Diseases on tuberculosis burden by HIV status from 1990 to 2023 is not an AI paper, but it reminds us why AI cannot drift away from epidemiology. Global disease burden is still shaped by geography, co-morbidity, and poverty. Any AI model that ignores those layers will simply produce cleaner versions of old inequity. It will be mathematically neat and clinically incomplete.
I also pay attention to the economics. The 2026 Cost Effectiveness and Resource Allocation review of AI in maternal and neonatal health argues that evaluation frameworks must account for benefits, limits, and implementation costs, not just diagnostic performance. That is the kind of realism global health needs. In a constrained setting, a tool that saves money but adds complexity can still fail. In a fragile system, complexity has a price.
One more number matters to me. The 2026 readiness review reports that implementation evidence is strongest when AI is paired with human support and integrated into existing services, rather than dropped in as a standalone product. That matches what I have seen. A model can point, but a person has to act. A device can suggest, but a clinician has to carry the burden of judgment.
What I Would Not Do
I would not deploy a black-box maternal health tool in a clinic without local validation, escalation pathways, and a named clinician responsible for the output. I would not let a vendor call a pilot successful because it had high technical accuracy while nurses were ignoring the alerts. And I would not confuse novelty with equity just because the software can speak in polished language.
That refusal is not technophobia. It is clinical discipline.
It is also why physician identity matters in the AI age. The profession does not need more people who can repeat the pitch. It needs clinicians who can test the pitch against the room, the patient, and the workflow. When I think about my own career, I do not want to be remembered as someone who admired the future from a distance. I want to be remembered as someone who asked whether the future helped the patient in front of me.
Global Health, Personal History
My own attachment to global health is personal. It comes from family stories, from seeing how much a stable diagnosis can mean when resources are thin, and from the recurring truth that illness is never only biological. It is logistical. It is financial. It is relational. Sometimes it is political. If you have ever watched a family arrange transport, childcare, language interpretation, and work leave just to make one appointment happen, you understand why global health is not a side interest. It is a measure of whether medicine is serious about everyone.
I first trained myself to think of global health as charity. That was wrong. I now think of it as a stress test for the profession. If a system can serve the least-resourced patient well, it is probably serving everyone else badly enough to notice. AI can help expose that truth, or it can hide it. The difference is design and ethics.
This is where I keep returning to the fetal ultrasound example. If AI can help a frontline clinician identify a risky pregnancy earlier, interpret images more consistently, or prioritize referral when the specialist is miles away, then it belongs in the conversation. If it only adds another layer of dependency on internet access and vendor maintenance, it does not deserve the name progress. The machine has to meet the mission.
When I want to think more about the human side of that mission, I sometimes revisit the values behind my own work at my clinical and editorial home at sinabarimd.com, and the background behind my practice at Dr. Sina Bari’s training and perspective. Credentials do not solve the problem. They do, however, remind me that judgment is earned in the room, not in the abstract.
Back in the Exam Room
By the time I finished with that patient last Tuesday, the question had changed. She still needed reassurance, but she also needed a system that would not ask her to come back four times for information that could have been available sooner. That is where AI belongs in global health, in the seam between scarcity and decision-making.
I told her what I could see, what I could not, and what the next step should be. I did not pretend certainty where there was none. I did not outsource judgment to a machine. But I left that room thinking differently than I would have years ago. The best AI for global health will not make clinicians disappear. It will make their care more reachable.
That is the promise worth keeping.
FAQ
How could AI help with prenatal ultrasound in a rural clinic?
AI can help by making ultrasound interpretation more consistent when a specialist is not available on site. In a rural clinic, that can mean earlier identification of malpresentation, growth concerns, or other findings that should trigger referral. The useful version of the tool is one that fits the clinic workflow and still works with limited connectivity.
What is the biggest risk of using AI in global health programs?
The biggest risk is deploying a tool that looks accurate in validation but fails in real-world conditions. That can happen because of poor local fit, missing escalation pathways, biased training data, or weak maintenance. If clinicians cannot trust the output, the tool becomes another burden instead of help.
What does Dr. Sina Bari think AI should do in low-resource healthcare settings?
Dr. Sina Bari’s approach is to treat AI as support for clinical judgment, not a replacement for it. The right tool should reduce friction, improve access, and survive scarcity. If it needs ideal infrastructure to work, it is not ready for the settings that need it most.
Why does the Gates Foundation fetal ultrasound AI project matter?
It matters because fetal ultrasound is one of the places where earlier information can change care in pregnancy. If AI helps frontline clinicians get a reliable read faster, patients may receive referrals and monitoring sooner. That can be especially important when repeat visits are difficult or impossible.
How do hospitals know if an AI tool is actually equitable?
They should look for local validation, transparent performance by subgroup, clinician oversight, and a clear plan for what happens when the tool is uncertain. Equity also means checking whether the system works for the patients with the least bandwidth, the least time, and the least access to specialists. A tool is only equitable if it is usable where the need is greatest.