Google tests geospatial foundation model across public-health studies

Google Research described five partner-driven public-health case studies using its Population Dynamics Foundation Model, or PDFM, in a post dated October 6, 2026. The associated paper was submitted to arXiv on October 5, 2026. Its abstract describes evaluations across four health domains, five epidemiological tasks and four countries: the United States, Canada, Mexico and the Democratic Republic of the Congo. The work addresses gaps that can delay disease estimates and make it harder to direct resources toward places with limited data.

PDFM uses self-supervised learning to turn diverse signals into compact representations of locations, which Google says are refreshed monthly. The inputs include aggregated search activity, information about local places and mobility, and environmental measures such as weather and air quality. The researchers present these embeddings as a complement to standard epidemiological models, not as a new task-specific system that must be trained for every health question. For AI builders, the approach offers a way to add location context to existing epidemiological machine-learning workflows rather than building separate pipelines to collect and process those raw signals.

One study examined measles, mumps and rubella vaccination across 146 counties along the US–Canada border. Google reports a 36% relative gain in explained variance, which it attributes to capturing cross-border movement and information spillovers. A cardiovascular-disease study covered 3,091 counties in the contiguous United States and assessed mortality nowcasting and interpolation. In Mexico, researchers evaluated dengue forecasts across about 2,450 municipalities; the post reports a statistically significant improvement for one-month forecasts using TimesFM and improved accuracy in as many as 72% of municipalities with active transmission. The paper also describes work on forecasting cholera hotspots.

The postpartum-depression case used data from 332,970 US respondents and assessed individual risk prediction. Google reports small gains in area-under-the-curve scores, including in states not seen by the model, and says the embeddings captured information associated with area poverty. The paper cautions that this signal does not replace individual socioeconomic data or resolve demographic gaps in screening. Taken together, the studies position PDFM as reusable geospatial context for public-health prediction and surveillance, with results reported across different diseases, locations and modeling tasks.