Google reports public health forecasting gains from location embeddings

Google reported five partner evaluations showing that its Population Dynamics Foundation Model can supplement existing public health models without task-specific fine-tuning. The model compresses search trends, mobility and environmental signals into numerical representations of locations that researchers can use as inputs. In dengue forecasting across Mexican municipalities, combining those inputs with TimesFM 2.0 improved one-month forecast accuracy in up to 72% of municipalities with active transmission. In U.S. cardiovascular mortality estimates, partners found no statistically significant difference between the model's inputs and traditional census data. The location dataset is commercially available in Preview through Google Maps Platform. Academics can request no-cost access for select, non-operational research uses.