Public-health surveillance often lags because data are sparse, siloed by borders, or slow to arrive. Google Research's Population Dynamics Foundation Model (PDFM) aims to fill those gaps by turning satellite, mobility, search-trend, weather, and built-environment signals into compact location embeddings. The embeddings are refreshed monthly and can be plugged directly into existing epidemiological models, avoiding custom data pipelines.
In a blog post, Google Research and five partner institutions evaluated PDFM across five distinct tasks. For MMR vaccination coverage near the US-Canada border, PDFM added a statistically significant 36% relative gain in explained variance. For cardiovascular mortality nowcasting across US counties, it was comparable to census-based models, making it a timely stand-in when census data are outdated. For dengue in Mexico, it improved one-month probabilistic forecasts, especially in active transmission hotspots.
The model also showed transferable gains in postpartum depression risk prediction, recovering about 15% of the predictive signal of income and insurance records, and improved cholera outbreak onset prediction in the Democratic Republic of the Congo. Because the embeddings are task-agnostic and require no fine-tuning, the authors argue they represent a new paradigm for planetary geospatial foundation models in global health. The source is a single Google Research blog post, so there are no conflicting sources to compare.