Google Earth AI tracks disease outbreaks before they spread

Tracking a fast-moving virus in a remote mining corridor usually means public health workers hiking for weeks with clipboards in hand. Even in wealthy nations, official disease mortality data is notoriously late, often lagging reality by two or three years.

Google revealed how its Google Earth AI is moving public health from a reactive scramble to a predictive science. By feeding satellite imagery, weather patterns, anonymous search trends, and mobility data into foundation models, the company is giving frontline health teams weeks of advance warning.

Pinpointing 45,500 people in minutes

The most dramatic deployment happened during the ongoing Ebola outbreak in the Democratic Republic of Congo (DRC).

Teams from WHO AFRO and the DRC's National Institute of Biomedical Research (INRB) used Google's prototype Geospatial Reasoning agent. Instead of writing complex GIS queries, epidemiologists simply asked questions in plain English.

Targeting remote mining corridors where infection risk and population mobility collide, the agent identified 48 exposed settlements and located over 45,500 at-risk residents in minutes. Field teams previously spent weeks gathering that same intelligence by hand. Those saved weeks let responders immediately dispatch mobile testing labs and tighten border surveillance.

The system is also clocking wins against other killers. In the DRC, pairing local records with Google's foundation model predicted cholera hotspots up to eight weeks before outbreaks occurred. In Mexico, researchers from Oxford and Tecnologico de Monterrey combined the model with local weather data to forecast dengue spikes, giving sanitation crews enough lead time to eradicate mosquito larvae before swarms took flight.

Turning movement and search into place fingerprints

Underneath these pilots sits Google Earth AI's Population Dynamics Foundation Model, or PDFM.

Traditional epidemiology hits two walls: massive time lags and rigid borders. In the U.S., cardiovascular disease kills over 916,000 people annually, yet official county mortality data lags by one to two years, while census covariates trail by up to three. NYU Grossman School of Medicine tested PDFM as a direct stand-in for outdated census surveys and found its same-year mortality predictions matched or beat traditional methods.

Meanwhile, researchers at Mount Sinai and Boston Children's Hospital used the model along the U.S.-Canada border. Domestic health data stops at national boundaries, but people and viruses cross freely. By adding Canadian mobility and behavioral embeddings to 146 U.S. border counties, the team boosted the model's ability to explain MMR vaccination rates from 16% to 22%, sharpening coverage estimates for 4.7 million border residents.

Google is rolling out these monthly updated embeddings as Population Dynamics Insights in commercial preview on Google Maps Platform, with free access offered to select academic researchers.

Why it matters

Tech companies have promised to transform epidemiology with big data for years, but Google's latest move gets two crucial things right.

First, they turned messy planetary telemetry into plug-and-play embeddings and a natural language agent. Frontline epidemiologists in cash-strapped regions do not have the time or engineering headcount to train bespoke spatial ML pipelines. Letting field doctors type questions and get operational maps in minutes bridges the massive gap between cutting-edge AI and chaotic real-world triage.

Second, it cuts through the artificial blind spots created by sovereign borders. Diseases do not carry passports. By blending cross-border mobility patterns into monthly signals, Google gives health officials an accurate read on how neighboring human behavior spills into domestic outbreak risks.

The lingering question mark is data coverage and equity. In the most vulnerable, offline pockets of the world, search signals and smartphone pings dry up, which risks skewing the model's predictive precision precisely where help is needed most. Privacy watchdogs will also keep a close eye on any commercialized dataset derived from human movement.

Still, buying frontline workers eight weeks of advance notice on cholera and replacing weeks of grueling manual contact tracing with a five-minute query makes this one of the most practical public health applications of foundation models we have seen.