Google Earth AI's PDFM location embeddings matched or improved conventional inputs across five public-health partner studies, lifting cross-border vaccine modeling explanatory power from 16% to 22%

PDFM compresses privacy-preserving search trends, human mobility, built-environment density, and environmental determinants into monthly-refreshed location embeddings that can be plugged into existing statistical and ML models as ready-to-use geospatial context without task-specific fine-tuning. Relative to traditional approaches requiring extensive task-specific data collection and custom data engineering pipelines, the work proposes pretrained representations of place as plug-and-play inputs rather than building new pipelines from scratch. PDFM serves as a proof-of-concept, with independent evaluations by global health research partners across five distinct disease domains, resource settings, and epidemiological tasks; the text reports embeddings matched or improved on conventional inputs across a wide variety of tasks.

Cross-border context corrects modeling fragmented by sovereign borders: in the 146 U.S. counties within 150 km of the Canadian border, supplementing with Canadian Forward Sortation Area embeddings raised the share of variation in MMR vaccination coverage explained by models from 16% to 22% (a 36% increase) and refined coverage estimates by at least 3 percentage points for 4.7 million border residents. Domestic-only models struggle to capture cross-border behavioral and mobility spillovers; adding neighboring-country embeddings revealed local patterns domestic-only models may miss. Conducted by researchers at the Mount Sinai Health System and Boston Children's Hospital, reporting specific county counts, distance thresholds, explanatory-power changes, and affected resident population.

When estimating current-year cardiovascular deaths across roughly 3,100 U.S. counties, PDFM showed no statistically significant differences from census-based inputs on nowcasting, while being much fresher and available in far more places; official county-level mortality data typically lag by 1-2 years and ACS covariates reflect conditions up to 2-3 years in the past. This suggests health departments can guide prevention resources using current conditions rather than waiting on multi-year survey cycles. Tested by partners at NYU Grossman School of Medicine, reporting the no-significant-difference comparison and the magnitude of data lags.

For dengue forecasting across roughly 2,450 Mexican municipalities (2020-2025), coupling PDFM with TimesFM 2.0 produced its greatest gains at one month ahead, improving forecast accuracy in up to 72% of active transmission municipalities with total error reductions 3.4X larger than degradations; in the Democratic Republic of the Congo's 403 health zones, a lightweight low-resource PDFM did not significantly enhance accuracy one to two weeks out but helped four to eight weeks out. Benefit depends on forecast horizon: short-term tracking still relies on recent case counts, while foundation model embeddings capture underlying environmental, connectivity, and population determinants that supplement historical data, enabling proactive planning one to two months in advance. The dengue work was done with public health researchers at the University of Oxford and Tecnológico de Monterrey; the cholera work addressed a use case defined by the WHO Regional Office for Africa using national IDSR surveillance data from the DRC.

The paradigm targets public health and epidemiological workflows that need timely, granular geospatial context: health departments, clinical and screening programs, and planners in resource-constrained or sparsely connected regions. Applicable settings include cross-border vaccine coverage modeling, current-year cardiovascular death nowcasting, one-month-ahead dengue forecasting, postpartum depression screening resource allocation, and four-to-eight-week cholera hotspot early warning. Embeddings refresh monthly and serve as ready-to-use inputs to existing statistical and ML models without task-specific fine-tuning.

The text notes current limitations such as static snapshots and points to active research into temporally dynamic embeddings and geographic transfer learning for under-connected regions. In the cholera task, no significant gain appeared one to two weeks ahead, indicating benefit depends on forecast horizon. Postpartum depression AUC gains were small (+0.0020 and +0.0038), with practical value shown in screening resource simulations. Some results are listed as bullet points without full detail in the text, so specific numbers and figures require the original paper.