Capability benchmark · Medicine · Computer science
Earth AI place embeddings improve selected public-health prediction tasks with uneven gains
Five public-health prediction case studies across four countries, including temporally held-out cholera hotspot forecasting.
Summary
Five studies assess PDFM place representations for vaccination, cardiovascular mortality, dengue, cholera and postpartum-depression prediction. In the temporally held-out DRC cholera study, eight-week top-five precision rose from 0.3556 to 0.4198. Gains vary across horizons and tasks; population context does not replace individual socioeconomic records.
AI role
PDFM encodes aggregated search, mobility and environmental context as regional covariates for downstream predictive models.
Narrative role
Documents task-level use of geospatial foundation representations in epidemiological research, with a temporally separated outbreak evaluation and explicit boundaries on clinical transfer.
Caveat
Preprint with collaborating investigators, not independent replication or measured deployment impact. Most embeddings use October 2023 data; historical dengue and postpartum evaluations have temporal misalignment. No causal research-productivity or patient-benefit estimate.