Public health decisions rely on timely, granular data, but conventional surveillance is hindered by reporting lags, data silos, and sparsity. Traditional modeling also requires extensive task-specific data collection. To address this, we introduce a paradigm using planetary geospatial foundation models. Using Google Earth AI’s Population Dynamics Foundation Model (PDFM) as proof-of-concept, we integrate self-supervised representations of 'place' into existing health workflows as plug-and-play inputs, enhancing models without new pipelines. PDFM compresses privacy-preserving search trends, human mobility, built-environment density, and environmental determinants into location embeddings. Without fine-tuning, these embeddings matched or improved on conventional inputs across diverse disease domains and tasks.
PDFM uses self-supervised learning to synthesize signals like aggregated search trends, built environment and mobility data, and environmental determinants into monthly-updated embeddings. Rather than requiring raw data processing, PDFM embeddings plug into ML workflows for geospatial context. Prior work showed these embeddings are task-agnostic, capturing social, behavioral, and environmental health determinants. To meet a higher burden of proof, we validated PDFM across five public health challenges with global partners.
Validation included MMR vaccination extrapolation across the US–Canada Border, achieving a +36% relative gain in explained variance. For cardiovascular disease mortality nowcasting across the Contiguous United States (3,091 counties), PDFM showed comparable accuracy to census-based models. These results demonstrate PDFM's value for diverse epidemiological tasks and resource settings.
Source: Google AI Blog · Summarized by HeadlinesBriefing