Loading page…
The research pairs local health records with satellite and population data. Emergency-response tools remain prototypes, while a related dataset is commercially available in Preview.
Listen to this story
Google’s Earth AI effort is moving from population modeling into operational outbreak response: WHO teams in the DRC used a conversational mapping prototype to identify communities around high-risk mining corridors, informing mobile-lab deployment and border surveillance. Separate models estimate Ebola’s potential spread into uninfected zones, with weekly forecasts intended to give coordinators lead time. Google says PDFM data is commercially available in Preview, but the emergency-response prototypes are distinct tools and are not generally available through that dataset offering.
The mapping agent identified 48 settlements and more than 45,500 people considered at risk; Google says the work would usually take weeks.
The DRC’s National Institute of Biomedical Research worked with Google on Ebola spread estimates using mobility flows, historical case trends, and Earth AI data.
Google reports that PDFM improved identification of cholera-prone zones in DRC evaluations.
Public-health responders in the Democratic Republic of Congo used an AI prototype to locate remote communities exposed to Ebola in minutes, Google said in an October 6 research disclosure. The WHO Regional Office for Africa identified 48 settlements and more than 45,500 at-risk people—a mapping task Google says would normally have taken weeks.
Google Earth AI combines environmental signals, satellite imagery and mobility data with models that help analyze places and populations. Local health information supplies the disease-specific context. The research uses AlphaEarth Foundations and the Population Dynamics Foundation Model, or PDFM, alongside a prototype Geospatial Reasoning agent.
PDFM combines aggregated search trends with movement and environmental patterns to describe communities. Partners can integrate those signals into their own systems. For emergency work, Google gave partners access to two research prototypes: an agent for mapping through plain-language conversations, and a planetary prediction engine for automated disease forecasting.
In the Ebola response, the WHO team used the mapping agent to examine remote mining corridors with high exposure risk and human movement. Google says the resulting population map allowed local responders to deploy mobile laboratories proactively and coordinate border surveillance.
Forecasting was a separate task within that response. Working with the epidemic modeling unit at the DRC’s National Institute of Biomedical Research, Google built models estimating Ebola’s risk of spreading into uninfected zones. The models combine mobility flows, historical case trends and Earth AI models and datasets. Weekly estimates are intended to give coordinators planning time before cases arrive.
The disclosure also describes partner research beyond emergency mapping. Each project pairs population signals with information specific to the health question, rather than using geography alone. Google reports the following results:
Access differs between the emergency prototypes and the underlying population dataset. Google says PDFM embeddings—model-generated data representations—are commercially available in Preview as Population Dynamics Insights through Google Maps Platform. Researchers can request no-cost access for selected uses, and eligible organizations can apply for Google Earth credits. That dataset access is distinct from the prototype tools used by outbreak-response partners.
Story updates
Google Research published five public-health evaluations on October 6 showing where location-based AI can help—and where it adds little. Partners tested its Population Dynamics Foundation Model across vaccination, cardiovascular mortality, dengue, postpartum depression and cholera. Google reports better outbreak forecasts in some settings, but no statistically significant advantage over census inputs for estimating current cardiovascular deaths.
The model, known as PDFM, compresses aggregated search trends, movement patterns, local facilities, weather and air quality into embeddings: compact numerical descriptions of places. Google describes the underlying signals as privacy-preserving and refreshes the location representations monthly.
Researchers add those descriptions to the statistical or machine-learning models they already use. The evaluations tested off-the-shelf embeddings without fine-tuning PDFM for each health task, rather than requiring teams to collect and process all the underlying data streams themselves.
For dengue, University of Oxford and Tecnológico de Monterrey researchers combined PDFM with Google’s TimesFM 2.0 forecasting model. They evaluated roughly 2,450 Mexican municipalities over 2020–2025. Google reports that one-month-ahead forecasts improved in up to 72% of municipalities with active transmission—not across every municipality.
The cholera evaluation used surveillance data from 403 health zones in the Democratic Republic of the Congo. For a use case defined by WHO’s Regional Office for Africa, researchers tested a lightweight PDFM version adapted for sparse internet connectivity. At one or two weeks ahead, recent case counts supplied most of the useful information; PDFM did not significantly improve accuracy.
Benefits emerged four to eight weeks ahead, a window Google identifies as useful for moving vaccines and clean-water supplies. Researchers measured how often the five highest-risk zones actually went on to have an outbreak.
NYU Grossman School of Medicine tested current-year cardiovascular mortality estimates across roughly 3,100 U.S. counties. Average errors were 18.7 deaths per county with PDFM versus 19.1 with census inputs. The potential benefit was freshness: the study’s embeddings used one month of data, while American Community Survey inputs pool five years and arrive up to a year after collection.
Other evaluations tested different information gaps. Adding Canadian location context increased the share of vaccination variation explained from 16% to 22% across 146 U.S. border counties. In a postpartum-depression study of 332,970 survey respondents, gains were much smaller: the risk-ranking score rose by 0.0020 in training-represented states and 0.0038 in unseen states.
Google says the embeddings are commercially available in Preview through Google Maps Platform as Population Dynamics Insights. Academics and public-health researchers can request no-cost access for selected non-operational research uses.
research.googleLoading discussion...
Join the conversation
Explain when early preparation would outweigh the uncertainty.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.