Adding location summaries from Google Earth AI’s Population Dynamics Foundation Model (PDFM) helped an existing forecast model rank cholera hotspots eight weeks ahead in the Democratic Republic of the Congo (DRC). In a held-out test of national surveillance records, the model’s weekly list of five high-risk health zones averaged 2.10 areas that later met the study’s outbreak threshold, compared with 1.78 using recent case history alone.
PDFM does not forecast outbreaks by itself. It turns information about places into compact numeric summaries, called embeddings, that another model can use alongside recent case reports. For this test, inputs included aggregated search trends, counts of visits to places and nearby facilities, weather, and air quality. Because digital coverage in the DRC was sparse, the researchers used a custom low-resource version: it pooled search counts over six months and compressed the inputs into 200 features rather than using PDFM’s standard encoder.
The researchers trained the forecast model on surveillance data from before Nov. 1, 2023, then tested it on the next 89 weeks of records across 403 health zones, through July 14, 2025. The paper counted an outbreak emergence as at least 100 suspected cases over three weeks after a week with fewer than 50. Its Precision@5 measure is the average share of each weekly top-five list that later met this threshold. At four weeks, Precision@5 rose from 34.4% to 39.1%; at one or two weeks, recent case counts were already informative and the study found no significant gain.
A second measure, which weighs how many outbreak zones the model finds against how often its high-risk rankings are correct, rose 6.1% relative at eight weeks but did not reach statistical significance (p = 0.064). The evaluation also used one fixed October 2023 embedding snapshot, not a live monthly refresh. It does not show that health teams acted on the rankings or that the forecasts reduced cases. The test indicates that place-based context can improve one longer-range hotspot-ranking measure in sparse surveillance data; real-world warning performance remains untested.
Sources
- Arbaaz Muslim et al., “Planetary Geospatial Foundation Models: A New Paradigm for Global Public Health,” arXiv:2610.05699v1
- Google Research, “Unlocking Earth AI’s planetary geospatial foundation models for global public health”
- Image: Oxfam East Africa via Wikimedia Commons, CC BY 2.0, unmodified; license.
