PDFM is a geospatial embedding model that aggregates privacy-preserving signals—such as aggregated search trends, built environment and mobility data, and environmental determinants—into monthly-updated location embeddings that can plug into existing epidemiological workflows.

In the DRC, lightweight PDFM variants improved early warning for cholera by four to eight weeks, increasing correct identification of high-risk zones and enhancing precision in endemic areas.

Among 332,970 CDC PRAMS participants, PDFM embeddings added predictive value for postpartum depression risk, modestly increasing AUC and recovering about 15% of the signal from income or insurance data when those inputs are unavailable.

PDFM reduced lag in chronic disease surveillance by matching census-based inputs for 2023 county-level cardiovascular deaths while delivering timelier and more geographically available data across 17 countries.

In cross-border contexts, incorporating Canadian location context with US border regions raised explained variance in MMR vaccination models from 16% to 22%, underscoring cross-border spillovers.

Overall, geospatial embeddings provide broader and timelier context for resource allocation, intervention planning, and outbreak alerting at a planetary scale, though they have static snapshot limitations and ongoing work on temporally dynamic embeddings and transfer learning.

PDFM is demonstrated as task-agnostic and capable of enhancing standard health models without task-specific fine-tuning, across diseases and geographic settings.

Access to PDFM embeddings is available commercially as Population Dynamics Insights in Preview, with no-cost access for select non-operational research use requests.

When paired with TimesFM, PDFM improved dengue forecasts in Mexican municipalities, notably one month ahead, with up to 72% of hotspots showing gains and up to 3.4x reduction in total forecast error.