The Pulse
Google Offers Earth AI Data for Public-Health Research
Google Research presented five public-health studies using its Population Dynamics Foundation Model, which turns aggregated search, mobility and environmental signals into geographic data. The embeddings are available commercially in Previe
AI.info Team · Written with AI ·
On October 6, Google Research detailed five public-health studies built around its Population Dynamics Foundation Model, or PDFM, and said researchers can request no-cost access to its location embeddings for select non-operational work. The announcement puts the model’s data to the test across vaccination estimates, cardiovascular disease, dengue, postpartum depression and cholera. It does not make the model’s underlying weights openly downloadable: Google says the dataset is available commercially in Preview, while academics and public-health researchers must request access.
Monthly location profiles, not a new surveillance system
PDFM converts aggregated search trends, patterns of human mobility, characteristics of the built environment and environmental measurements such as weather and air quality into compact representations of places. Google says those embeddings refresh monthly and can be added to statistical or machine-learning systems that epidemiologists already use, without task-specific fine-tuning.
The approach is aimed at familiar gaps in health data: official figures can arrive years late, datasets often stop at national borders, and some regions lack detailed records. “Together, these results showcase capabilities of geospatial foundation models for public health surveillance,” write Arbaaz Muslim et al., authors of the paper.
Cross-border data improves one vaccine estimate
In one study, researchers combined U.S. county embeddings with data from nearby Canadian postal areas to estimate measles, mumps and rubella vaccination coverage in 146 U.S. counties near the border. The share of variation explained by the models rose from 15.9% to 21.6%, a relative improvement of about 36%, according to the paper. Google says the added context changed coverage estimates by at least three percentage points for 4.7 million residents in border counties.
A separate study with NYU Grossman School of Medicine examined cardiovascular disease deaths in roughly 3,100 U.S. counties. For nowcasting, models using PDFM had a mean absolute error of 18.7 deaths per county, compared with 19.1 for models using American Community Survey data. Google says the difference was not statistically significant; the comparison suggests the embeddings could offer a more frequently refreshed alternative when census measures are stale, not that they outperform them conclusively.
Dengue gains concentrate in active hotspots
For dengue, researchers paired PDFM with Google’s TimesFM 2.0 forecasting model and evaluated case counts across about 2,450 Mexican municipalities from 2020 to 2025. The largest reported benefit came in one-month forecasts: accuracy improved in as many as 72% of municipalities with active transmission. The results point to a possible aid for decisions such as where to direct vector control or clinical staff, though the reported gain is concentrated in places already experiencing transmission.
In a cholera study with the World Health Organization’s Regional Office for Africa, the team tested forecasts across 403 health zones in the Democratic Republic of the Congo. Eight weeks ahead, the number of outbreaks appearing among the model’s five highest-risk picks rose from 1.78 to 2.10 per week—an 18% improvement. The paper says PDFM did not significantly improve accuracy one or two weeks ahead, when recent case counts already carried much of the signal.
Google’s access offer has limits
The five evaluations also include postpartum-depression risk prediction using survey responses from 332,970 participants. Adding location embeddings produced a small improvement in prediction scores, including in states absent from model training. The authors stress that place-level signals do not replace a person’s own socioeconomic information or resolve demographic gaps in screening.
That distinction matters to the release. Google’s no-cost research access is request-based and restricted to select non-operational uses; commercial customers can access the embeddings through Google Maps Platform’s Preview offering. The studies make a case for testing location data alongside conventional health records, but they do not establish that the same gains will hold in routine clinical or outbreak-response operations.