Google describes Earth AI prototypes for outbreak mapping and population-health forecasts

Google is applying its Earth AI models to public-health planning, combining local health records with satellite imagery, mobility information and environmental signals. Its October 6 account describes outbreak-mapping and forecasting prototypes used by health partners, along with research into broader population-health patterns.

The Google Research and Google Health announcement separates two research prototypes from a commercially available preview dataset. The Geospatial Reasoning agent supports conversational mapping, while a planetary prediction engine supports disease forecasting. Population Dynamics Insights, based on Population Dynamics Foundation Model embeddings, is available in Preview through Google Maps Platform.

These tools support planning at a community or regional level. Google's account does not present them as a way to diagnose an individual patient or replace local health professionals. The work depends on pairing the models with local information and the people responsible for interpreting it.

Mapping exposure during the DRC Ebola outbreak

Google says the WHO Regional Office for Africa's Emergency Preparedness and Response Hub and the DRC National Institute of Biomedical Research used its prototypes during the ongoing Ebola outbreak. The WHO AFRO team asked the Geospatial Reasoning agent to map remote mining corridors with high exposure risk and mobility.

According to Google, the team identified 48 exposed settlements and more than 45,500 at-risk people in minutes, compared with a process that would normally take weeks. The company says the mapping helped responders deploy mobile laboratories and coordinate border surveillance. These are reported partner uses, not independently verified timing measurements by Franklin.

Working with INRB's Epidemic Modeling and Intelligence Unit, Google also built models to estimate the risk of Ebola spreading into uninfected zones. The models combine mobility flows, historical case trends and Earth AI information to provide weekly planning insights. A forecast of spread remains distinct from confirmation that cases have appeared in a location.

Local records give population signals a purpose

The Population Dynamics Foundation Model combines aggregated search trends with mobility and environmental patterns. Google describes partners integrating those signals into their own systems, rather than replacing epidemiological records with a standalone general model.

NYU Langone Health researchers used population signals in models of same-year cardiovascular mortality. Google says they achieved performance similar to, and in some cases better than, conventional approaches. The announcement gives no single universal improvement figure, so the result should not be turned into a blanket accuracy claim for cardiovascular forecasting.

Researchers at Mount Sinai Health System and Boston Children's Hospital examined behavioral patterns on both sides of the US-Canada border to model vaccination rates in US counties. Google says those patterns helped reveal information traditional models could miss. The example concerns population-level coverage estimates, not a record of any particular person's immunizations.

Seasonal and waterborne disease studies

For dengue, researchers at the University of Oxford and Tecnológico de Monterrey combined PDFM with local climate models to forecast outbreaks in Mexico. Google describes more accurate forecasts and potential lead time for interventions such as controlling mosquito larvae. The announcement does not supply a measured reduction in cases resulting from those forecasts.

For cholera in the DRC, Google reports that combining epidemiological records with PDFM improved identification of outbreak-prone health zones up to eight weeks ahead. That is a reported forecasting horizon in the evaluation, rather than a guarantee of eight weeks' warning for every future outbreak.

Both examples illustrate why a health model needs evaluation against a specific outcome and location. Identifying a high-risk zone, forecasting cases and improving an intervention are different results. A useful deployment review should ask which one the evidence establishes.

Access is available through specific programs

Researchers can request no-cost access to Population Dynamics Insights for selected uses, and eligible organizations can apply for Google Earth credits through Google Maps Platform Public Programs. Google.org has also funded INRB to help modernize testing and surveillance.

The announcement describes a collection of research collaborations and access routes. It does not mean every prototype is a generally available finished product. Health teams evaluating the work need to identify the dataset or prototype being offered, its local data requirements and the evidence for the particular planning task they intend to support.