Google Earth AI anticipates public health risks

Google is testing Google Earth AI to anticipate outbreaks and other public health risks by combining maps, satellite imagery, human mobility and environmental data. The goal is to detect problems before they reach hospitals, rather than respond after an emergency has already spread.

The tool brings together several AI models, including AlphaEarth Foundations and the Population Dynamics Foundation Model (PDFM). The latter analyzes aggregated patterns in searches, travel and environmental conditions to build a detailed picture of how a population is changing.

Google has also created a prototype called Geospatial Reasoning, an agent that lets you query maps and geographic data in everyday language. Instead of manually preparing layers of information over several days, a health team can ask it to locate exposed areas, cross-reference mobility data or identify hard-to-reach communities.

How it was tested during an Ebola outbreak

During the Ebola outbreak in the Democratic Republic of the Congo, teams from the World Health Organization and the country’s National Institute for Biomedical Research tested these prototypes.

The geospatial agent analyzed remote mining corridors where high mobility coincided with a greater risk of exposure. Within minutes, it identified 48 exposed settlements and more than 45,500 people at risk. According to Google, an analysis of this kind would normally have taken weeks.

With that information, local teams were able to plan the deployment of mobile laboratories and better coordinate border surveillance. The system also combined population movements with the historical progression of cases to estimate when the virus might reach areas that had not yet been affected.

This is not an infallible prediction tool, nor is it meant to replace epidemiologists. These are research models that can give teams extra time to decide where to send staff, tests and supplies.

Beyond emergencies

Google’s tests cover diseases and problems that do not always appear as sudden crises:

• Cardiovascular disease: Researchers at NYU Langone Health used recent population signals to project cardiovascular mortality for the same year. Its performance was similar to that of conventional methods and better in some cases.
• Vaccination: Teams at Mount Sinai and Boston Children’s Hospital incorporated mobility patterns between the United States and Canada to estimate measles, mumps and rubella vaccination rates in U.S. counties more accurately.
• Dengue: Researchers from the University of Oxford and Tecnológico de Monterrey combined the population model with local climate data to anticipate outbreaks in Mexico. That lead time could help eliminate mosquito breeding sites before infections rise.
• Cholera: In assessments conducted in the Democratic Republic of the Congo, combining epidemiological histories with PDFM data helped identify areas prone to outbreaks up to eight weeks in advance.

What changes for health teams

The practical value lies in filling the gaps left by traditional systems: remote areas without recent data, reports that take time to arrive and populations that cross borders or move frequently.

But AI does not turn incomplete data into certainty. The results depend on the quality of local information and must be interpreted alongside the knowledge of doctors, epidemiologists and authorities in each territory.

Google says PDFM representations are already available in commercial preview as Population Dynamics Insights, a geospatial dataset for Google Maps Platform. Some organizations can request free access for selected uses, as well as Google Earth credits through Google Maps Platform’s public programs.

The key question is whether these predictions work consistently outside research trials and reach the teams working on the ground. The promise is not to guess the next outbreak, but to gain days or weeks to act before a scattered signal becomes an emergency.