Opening Up Earth AI’s Planet-Scale Geospatial Foundation Models for Global Public Health
Earth AI PDFM supplies place embeddings that help public health teams reduce data gaps and delays across vaccination, disease, and maternal care.
Key Takeaways
- Google Earth AI’s Population Dynamics Foundation Model (PDFM) provides privacy-preserving, monthly-refreshed geospatial place embeddings that plug into existing public-health and epidemiological workflows to reduce data lags, silos, and sparsity.
- Five partner case studies show PDFM can support diverse health tasks, including MMR vaccination coverage, cardiovascular mortality nowcasting, dengue forecasting, postpartum depression risk prediction, and cholera response planning.
- In U.S.-Canada border counties, adding Canadian location embeddings to MMR models improved explained variance by 36% and refined vaccine-coverage estimates for 4.7 million residents.
- For U.S. county-level cardiovascular disease deaths, PDFM-based models performed comparably to census-based models and reduced large-county errors, suggesting they can substitute for outdated official covariates.
- Coupled with TimesFM 2.0, PDFM improved one-month dengue forecasts in up to 72% of active transmission municipalities in Mexico, helping guide vector control and clinical staffing.
Meta description: Google Research's Oct 6, 2026 case studies show Earth AI PDFM supports MMR, cardiovascular, dengue, postpartum depression and cholera public health workflows.
Tags: PDFM, Google Earth AI, public health, epidemiology, geospatial AI
Featured image: abstract friendly illustration of a soft globe made from map tiles, search, mobility, weather and clinic icons connected by gentle arcs, no real people, no logos, no text.
On October 6, 2026, Arbaaz Muslim and Gautam Prasad, software engineers at Google Research, presented five partner-driven case studies showing how Google Earth AI's Population Dynamics Foundation Model, or PDFM, can be inserted into existing epidemiological workflows to reduce data gaps and reporting delays. The work addresses a familiar problem for public health teams: decisions about cardiovascular disease support, maternal mental health screening, dengue response, cholera prepositioning and vaccination coverage depend on timely place-level data, but conventional surveillance is often late, siloed by borders, sparse or difficult to deploy during rapid outbreaks.
For readers tracking AI in health, the question is not whether PDFM replaces epidemiologists. It is whether pre-trained representations of place can act as ready-to-use inputs that improve the statistical and machine learning models health officials already use, without building a custom data pipeline for every new disease or region.
Why PDFM enters public health workflows
Conventional epidemiological surveillance faces three recurring bottlenecks: multi-year reporting lags, data siloed by rigid geopolitical boundaries, and data sparsity. Even when those limitations are reduced, traditional modeling can require extensive task-specific data collection and custom engineering, which may be too slow for acute outbreaks such as dengue or cholera and too costly for resource-constrained settings.
Google's answer is a planetary geospatial foundation model paradigm for global public health. PDFM is part of Google Earth AI, a suite of geospatial models connecting satellite imagery, weather, anonymous search trends, human mobility and other population dynamics. Through self-supervised learning, PDFM compresses privacy-preserving signals into compact location embeddings refreshed monthly. The signals include aggregated search trends reflecting community-level interest in topics and resources; built-environment and mobility data describing the density and busyness of places such as pharmacies, clinics and parks; and environmental determinants such as high-resolution weather and air-quality metrics.
Instead of requiring researchers to collect and process those raw streams themselves, PDFM provides off-the-shelf place embeddings that can be plugged into existing health-science and epidemiological workflows. Prior work showed the embeddings are task-agnostic and performed strongly when filling gaps in a wide variety of CDC health metrics. The new studies raise the burden of proof: they ask whether the same representations can help across different diseases, countries, resource settings and epidemiological tasks.
What the five case studies found
The validation matrix spans immunization, noncommunicable disease, vector-borne disease, mental health and waterborne disease. Each case was evaluated with external partners.
- MMR vaccination, cross-border extrapolation. Mount Sinai Health System and Boston Children's Hospital studied 146 U.S. counties within 150 km of the Canadian border. Domestic-only models often struggle to predict local vaccine uptake, but researchers added Canadian Forward Sortation Area embeddings to U.S. county embeddings to capture cross-border mobility and information spillovers. The relative gain in explained variance was 36%, rising from 0.159 to 0.216, statistically significant. The cross-border context also refined MMR vaccination coverage estimates by at least 3 percentage points for 4.7 million border residents, helping public health officials pinpoint communities at risk of measles outbreaks.
- Cardiovascular disease, mortality nowcasting and interpolation. The Department of Population Health at NYU Grossman School of Medicine tested PDFM across 3,091 counties in the contiguous United States. Cardiovascular disease claims more than 916,000 lives annually in the U.S., yet official county-level mortality data from the National Vital Statistics System typically lags by 1-2 years, and American Community Survey covariates can reflect conditions up to 2-3 years in the past. For 2023 county-level cardiovascular disease deaths, PDFM-based models produced a mean absolute error of 18.7 deaths per county, compared with 19.1 deaths per county for census-based models. Root mean squared error, or RMSE, was 46.00 versus 57.69. The differences were not statistically significant, suggesting PDFM can serve as a substitute for census inputs when data are outdated or unavailable. The PDFM version also reduced large county outlier errors by about 20%, from 57.7 to 46.0.
- Dengue outbreaks, short-horizon probabilistic forecasting. University of Oxford and Tecnológico de Monterrey researchers coupled PDFM with TimesFM 2.0, an open time-series foundation model, to forecast dengue case counts across about 2,450 Mexican municipalities from 2020 to 2025. The one-month forecast gain was statistically significant: the change in Weighted Interval Score was -0.0051, where lower is better. Accuracy improved in up to 72% of active transmission municipalities, and total error reductions were 3.4X larger than degradations. The gains concentrated in active transmission hotspots, where short-term forecasts can support vector control and clinical staffing decisions.
- Postpartum depression, individual risk prediction. The Institute on Human Development and Disability at the University of Washington tested PDFM embeddings using the CDC PRAMS national survey, with 332,970 respondents. Adding the embeddings produced a statistically significant boost to predicting which mothers were at risk: AUC rose by 0.0020 on a 0.62 baseline in states seen during training and by 0.0038 in unseen states. AUC measures how well a model ranks higher-risk cases above lower-risk ones. The embeddings encoded area poverty with R² = 0.45 and recovered about 15% of the predictive signal of income and insurance records when those details were unavailable. In simulations where a health system follows up with the 20% highest-risk mothers, PDFM helped reach 5,640 more rural mothers with postpartum depression each year. In systems aiming to catch 80% of all cases, it cut 17,723 false alarms annually.
- Cholera emergence, outbreak onset prediction. WHO AFRO, the World Health Organization Regional Office for Africa, used national surveillance data from the Democratic Republic of the Congo's Integrated Disease Surveillance and Response reporting, covering 403 health zones over 89 weeks. Cholera early warning is critical for prepositioning oral cholera vaccines and clean water supplies, but outbreaks are rare: in any given week, fewer than 1 in 100 of the country's 403 health zones sees one begin. The study tested a lightweight, low-resource version of PDFM adapted for regions with sparse internet connectivity. Reported relative gains were statistically significant: +9.7% in Area Under the Precision-Recall Curve at 4 weeks, +18.1% Precision@5 at 8 weeks and +19.3% Precision@5 in endemic health zones. Precision@5 measures the share of the model's five highest-risk zones each week that actually went on to have an outbreak.
What the results suggest
The case studies show two different kinds of value. In some settings, PDFM can act as a stand-in when official data are too old: the cardiovascular study found comparable accuracy to census-based nowcasting while using embeddings built from a single month of data, and PDFM is available in 17 countries while ACS covers only the United States. In other settings, it can add context that domestic data miss: the MMR study captured cross-border behavioral and mobility spillovers that can affect local vaccine uptake.
The dengue, maternal mental health and cholera cases point to earlier operational action. Dengue forecasts became more useful in active transmission hotspots, where timing matters for vector control and clinical staffing. Postpartum depression screening moved beyond individual intake data by using community-level context, with transferable gains in unseen states and better ranking in resource-allocation simulations. Cholera onset prediction is harder because events are rare, but the reported Precision@5 gains in endemic health zones suggest the model can help concentrate attention on the most likely places.
The evidence still has limits. The AUC gains for postpartum depression are small in absolute terms, and the cardiovascular differences were not statistically significant compared with census-based models. The work is presented as a proof-of-concept for a broader paradigm, not as an automated public health decision system.
What to watch next
The next question is whether health departments can treat PDFM embeddings as routine geographic inputs alongside local surveillance, laboratory reports, clinical records and operational constraints. That will depend on how the embeddings behave under different privacy assumptions, how useful a monthly refresh cadence is during fast-moving outbreaks, and how well the place signals transfer to regions where search, mobility and environmental data are sparse or unevenly collected.
For now, the story is about integration: PDFM is positioned as a reusable context layer that can enhance existing epidemiological models rather than replacing them. The five case studies show where that layer helped capture cross-border patterns, reduce reliance on stale census inputs, sharpen short-horizon forecasts, support maternal mental-health screening and improve cholera risk ranking.

Comments (0)