AI supports population health data by combining satellite imagery, phone data and existing surveys to estimate things like poverty, population size and vaccination coverage in places and years that surveys miss. These estimates can guide campaigns and aid. But they are only as good as the ground-truth surveys used to train and check them, and without new surveys they drift into confident guesswork.
For forty years, much of what the world knows about the health of women and children in poor countries has come from a single source. The Demographic and Health Surveys programme, funded mainly by the United States since 1984, sent trained interviewers into households in more than 90 countries to ask about births, deaths, vaccinations, nutrition, contraception and HIV. Child mortality estimates, vaccination coverage maps and a vast share of global health research rest on its data.
In early 2025, the US government shut the programme down along with other foreign aid data work. At that point, surveys were set up or already collecting data in around two dozen countries, including large ones such as Nigeria and the Democratic Republic of the Congo. The Gates Foundation later stepped in to fund ICF, the lead implementer, to stabilise the programme and complete surveys that were in progress. But the long-term future of the most important household health data system in the world is now uncertain.
This is the moment that makes AI’s role in population data matter. Machine learning can estimate poverty from satellite images, map settlements from space and fill gaps between surveys. It is tempting to see it as a replacement for costly fieldwork. That would be a serious mistake. AI estimates of population health are a way of stretching surveys, not replacing them, and the less new survey data exists, the more their errors hide.
How is AI used to estimate population health data?
The approaches fall into three families.
Mapping where people are. High-resolution satellite imagery combined with machine learning can identify buildings and settlements, and from those estimate how many people live where. This matters because many countries’ censuses are years or decades old, and planning a vaccination campaign with the wrong denominator means missing people.
Estimating conditions from proxies. The best-known example is a 2016 study in Science by Neal Jean and colleagues at Stanford. They trained a model to predict household consumption and assets from daytime satellite imagery, using night-time lights as an intermediate signal, across five African countries: Nigeria, Tanzania, Uganda, Malawi and Rwanda. The model explained up to 75% of the variation in local economic outcomes. Later work extended similar methods, adding mobile phone and connectivity data, to produce wealth maps for entire regions, and researchers have used phone-data models to help target cash transfers, for example during Togo’s pandemic aid programme.
Filling in between measurements. Statistical and machine learning models estimate indicators such as child mortality or vaccination coverage for places and years with no direct survey, by borrowing strength from neighbouring areas, earlier surveys and related indicators.
All three families share one feature. They are trained and validated against real measurements, usually household surveys like DHS. The satellite model in the Stanford study learned what poverty looks like from above by being shown survey data about poverty on the ground.
Has AI population mapping actually improved health programmes?
The strongest example comes from polio eradication in northern Nigeria. In the early 2010s, Nigeria accounted for more than half of the world’s polio cases, and vaccination teams worked from hand-drawn settlement maps. When officials overlaid a sample of those maps with satellite imagery, many settlements were misplaced, misnamed or missing altogether.
The response, which grew into the GRID3 programme, used high-resolution satellite imagery and, later, machine learning to map settlements and estimate populations, while GPS tracking of vaccination teams showed which areas they had actually visited. The tracks revealed that teams routinely missed whole clusters of households. Better maps and tracking helped direct teams to missed settlements. Nigeria went a full year without a wild polio case in 2017, and the WHO African region was certified free of wild poliovirus in 2020. Many factors contributed, but the mapping work is widely credited as one of them.
Note what made this work. The satellite maps were checked by field teams with local guides who confirmed settlement names and locations. The AI found places; people confirmed them. And the output was used immediately, by vaccination teams, who would find out quickly if the map was wrong.
When does AI-filled data become guesswork?
The danger is not that AI estimates are inaccurate. It is that they can look exactly as precise when they are inaccurate as when they are accurate.
Models drift when the world changes. A model trained on the relationship between satellite features and poverty in 2018 assumes that relationship still holds. After a drought, a conflict or a currency collapse, roofs do not change quickly but household consumption does. Without a new survey, the model cannot know that it is wrong.
Errors concentrate where it matters most. Satellite-based models tend to be better at distinguishing wealthier areas from poorer ones than at making fine distinctions among very poor households. Informal settlements, nomadic populations and displaced people are hard to see from space and often underrepresented in training data. These are exactly the people health programmes most need to find.
Models get trained on their own outputs. If modelled estimates are published and then used as inputs for the next generation of models, the system can drift away from reality while appearing internally consistent. Only fresh measurement breaks that loop.
Estimates become facts through repetition. A modelled vaccination coverage figure appears in a report, is cited by a donor, lands in a funding formula and becomes, in effect, the official number. The uncertainty interval disappears somewhere along the way.
This is why the pause in DHS matters beyond the surveys themselves. Every modelled estimate that leans on DHS for calibration loses its anchor as the data ages.
A worked example: two ways to fill a gap
This is a hypothetical example to illustrate the difference between stretching and replacing surveys.
A health ministry needs district-level estimates of childhood vaccination coverage to decide where to run catch-up campaigns. Its last national household survey is five years old, and no new one is funded.
In the first approach, the ministry commissions a model that combines the old survey, satellite-based population estimates, facility records and mobile connectivity data to produce a coverage map for every district. The map is striking and precise to one decimal place. The ministry targets the 20 lowest-coverage districts. Nobody checks the map on the ground. Two years later, a measles outbreak begins in a district the model rated as well covered, where a large displaced population had settled since the last survey.
In the second approach, the ministry uses the same model but treats its output as a hypothesis. It runs rapid, small-sample coverage surveys in 15 districts: the ones the model rates lowest, a handful it rates highest, and several where the model’s uncertainty is widest. Those checks show the model is broadly right, but badly wrong in two districts with recent displacement. The ministry corrects its targeting and feeds the new data back to recalibrate.
The second approach costs more than the first. It is far cheaper than a national survey, and far cheaper than a measles outbreak.
The strongest counterargument: are surveys really better?
A fair critic would point out that household surveys have their own serious problems. They are expensive and slow, often reporting years after fieldwork. They are too small to give reliable estimates for most districts. They depend on recall and self-report, which is unreliable for things like vaccination history. And they are shaped by the priorities of the donors who fund them, which is part of why the DHS programme’s collapse was so abrupt. If AI can provide timely, granular estimates at a fraction of the cost, insisting on surveys may simply mean making decisions on older, coarser data.
This is a fair challenge, and the answer is not survey fundamentalism. AI estimates are genuinely better than surveys for some questions: where people live, how settlements change, and where to look first. And surveys will never be granular enough on their own for district decisions.
But the argument misses the dependency. AI’s granularity is borrowed from the surveys it is trained on. Remove the surveys and you do not get cheaper data; you get a model with no way to find out it has gone wrong. The right design is a smaller, more frequent survey system that is designed from the start to calibrate models, not an either-or choice between them.
What this means for governments, donors and data scientists
For national statistics offices and health ministries, the lesson is to own the ground truth. That means building domestic capacity and budget for household surveys, even lighter and more frequent ones, rather than depending on a single external programme. As our piece on why “who owns patient data?” is the wrong question argued, the practical question is who controls the use and continuity of data, and the DHS pause has made that question urgent at population level.
For donors, the priority is funding measurement as infrastructure, not as a series of projects. AI mapping and modelling deserve support, but a dollar spent on models without the surveys to calibrate them buys a more precise-looking version of ignorance. This is especially important for the diseases and populations that already receive little attention, a point we made in our piece on how the biggest killers in emerging markets get the smallest share of aid.
For data scientists and modellers, the obligation is transparency. Publish uncertainty prominently, not in an appendix. Label modelled estimates clearly as modelled. State when the most recent calibration data was collected. And design models to tell users where a new survey would most reduce uncertainty, so that scarce fieldwork goes where it counts.
Frequently asked questions
What happened to the Demographic and Health Surveys?
The Demographic and Health Surveys programme, funded mainly by USAID since 1984 and active in more than 90 countries, was shut down by the US government in early 2025 as part of foreign aid cuts. Surveys were underway in around two dozen countries at the time. The Gates Foundation subsequently funded ICF, the lead implementer, to help complete in-progress surveys and restore access to data, but the programme’s long-term future remains uncertain.
Can satellite images really measure poverty?
Satellite imagery combined with machine learning can estimate relative wealth and poverty across areas with useful accuracy. A 2016 study in Science explained up to 75% of the variation in local economic outcomes in five African countries. But these models learn from household survey data, are better at ranking areas than measuring absolute poverty, and can miss rapid changes and hard-to-see populations.
What is GRID3?
GRID3, which stands for Geo-Referenced Infrastructure and Demographic Data for Development, is an initiative that produces geospatial data such as settlement maps and population estimates for use in planning. It grew out of work supporting polio vaccination campaigns in northern Nigeria, where satellite-based maps and GPS tracking of vaccination teams helped identify settlements that had been missed.
Are modelled health estimates reliable?
Modelled estimates can be reliable when they are calibrated against recent, good-quality measurements and when their uncertainty is reported honestly. They become unreliable when the underlying data is old, when conditions change quickly, or when uncertainty is dropped as numbers move into reports and funding formulas. Always check when the most recent data behind a modelled estimate was collected.
The takeaway
AI can help health systems see the gaps between surveys more clearly than ever. It cannot see what nobody has measured. The countries and donors that treat models as a reason to stop counting will end up with beautiful maps of a world that no longer exists.
Where have you seen a modelled estimate treated as a measured fact, and what did it cost?
