Google tests location embeddings to fill gaps in public-health models

Google Research presented five public-health case studies in a post dated October 6, 2026, testing whether reusable representations of places could improve disease forecasts, vaccination estimates and risk screening. The studies put Google Earth AI’s Population Dynamics Foundation Model, or PDFM, alongside existing epidemiological tools rather than replacing them. An accompanying preprint was submitted to arXiv on October 5.

The idea is to spare researchers some of the work of assembling location-specific data for each health problem. PDFM compresses aggregated search trends, information about local places and mobility, weather and air quality into embeddings: numerical representations of a place. Google says these are refreshed monthly and can feed existing statistical and machine-learning models without task-specific fine-tuning.

The results in Google’s post, as captured on October 16, show that those extra inputs were not uniformly useful. In a cholera study using surveillance data from the Democratic Republic of the Congo, PDFM did not significantly improve forecasts one or two weeks ahead, when recent case counts already supplied much of the signal. At eight weeks, however, the model’s five highest-risk zones included an average of 2.10 zones that subsequently had outbreaks, up from 1.78 without the added context, Google reports.

The cardiovascular study offered a different result: a possible substitute for older data, rather than a clear accuracy win. Across 3,091 counties in the contiguous United States, Google reports that models using PDFM estimated current-year cardiovascular deaths about as accurately as models using census-derived inputs. Average errors were 18.7 versus 19.1 deaths per county; the difference was not statistically significant.

For postpartum-depression screening, the reported gains were small. Adding PDFM increased the area-under-the-curve score by 0.0020 over a 0.62 baseline in states represented during training, and by 0.0038 in previously unseen states. That suggests some location context transferred across state boundaries, but the authors draw a firm limit in the preprint abstract: the embeddings did not replace individual socioeconomic information or close demographic gaps in screening.