Google Earth AI location embeddings tested across five public health studies
Google Research reports five partner-led studies testing its Population Dynamics Foundation Model embeddings on vaccination, disease and outbreak tasks.
The story in focus
Google Research describes five partner-led evaluations of its Population Dynamics Foundation Model, which turns search trends, mobility and environmental data into monthly location embeddings. Across vaccination, cardiovascular, dengue, maternal depression and cholera tasks, the off-the-shelf embeddings matched or improved on conventional inputs without task-specific fine-tuning.
- PDFM combines privacy-preserving search trends, mobility, built-environment and environmental signals into location embeddings refreshed monthly
- A Mount Sinai and Boston Children's Hospital study raised explained MMR vaccination variance from 16% to 22% using cross-border Canadian embeddings
- Oxford and Tecnológico de Monterrey coupled PDFM with TimesFM 2.0 to improve one-month-ahead dengue forecasts in up to 72% of active Mexican municipalities
What the model does
Google Research presents its Population Dynamics Foundation Model as a geospatial foundation model within Google Earth AI. It uses self-supervised learning to compress aggregated search trends, place density and mobility, and high-resolution weather and air quality into compact embeddings intended as fingerprints for locations, refreshed monthly. The embeddings are designed as plug-and-play inputs to existing epidemiological models rather than requiring new pipelines or task-specific fine-tuning.
Partner evaluations
Five independent partner evaluations spanned different tasks and settings. At Mount Sinai Health System and Boston Children's Hospital, adding Canadian embeddings to U.S. county models raised explained variation in MMR coverage from 16% to 22% and refined estimates by at least 3 percentage points for 4.7 million border residents. NYU Grossman School of Medicine found PDFM comparable to census inputs when nowcasting 2023 county-level cardiovascular deaths, with average error of 18.7 versus 19.1 deaths per county and root-mean-square error cut 20%.
University of Oxford and Tecnológico de Monterrey combined PDFM with TimesFM 2.0 for dengue forecasts across about 2,450 Mexican municipalities. University of Washington researchers found small but statistically significant screening gains across 332,970 CDC PRAMS participants. A low-resource PDFM variant tested in the Democratic Republic of the Congo improved eight-week cholera shortlists.
Next steps
Google says current static snapshots are driving research into temporally dynamic embeddings and geographic transfer learning for under-connected regions. The embeddings are available in preview commercially through Google Maps Platform, with no-cost access offered for select academic and public health research use cases.
Did the embeddings require task-specific fine-tuning?
No. Google Research says the off-the-shelf location embeddings matched or improved on conventional inputs across disease domains, geographic settings and tasks without task-specific fine-tuning.
Sources & further reading
- Unlocking Earth AI’s planetary geospatial foundation models for global public health ↗research.google
Help keep the record accurate. Send a correction.
