Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Google Earth AI launches planetary geospatial foundation models for public health - OpenSmartRoute
Google Earth AI launches planetary geospatial foundation models for public health
Google released a new model that turns search trends and weather data into location fingerprints. These inputs improve disease forecasting without needing custom training.
Key points
PDFM embeddings matched or improved conventional inputs across five disease domains.
Cross-border context increased measles vaccination coverage variance explained by 36%.
PDFM reduced county-level cardiovascular death prediction errors by 20%.
The model recovered 15% of income and insurance predictive signal in unseen states.
Why it matters: Runners can plug these embeddings into existing ML workflows to get fresher data than census reports.
By OpenSmartRoute editorial · written through the router by writer-small
From Google Research blog - “Unlocking Earth AI’s planetary geospatial foundation models for global public health”
Google Earth AI released Population Dynamics Foundation Model (PDFM). This new tool turns search trends and weather data into location fingerprints. Public health officials can now use these inputs for disease forecasting. They do not need to train custom models for every specific task. The system works without requiring extensive custom training pipelines. Existing epidemiological workflows can adopt this technology immediately.
Google researchers developed PDFM to fix data gaps in public health. Traditional surveillance often suffers from multi-year reporting lags. Data silos also restrict analysis across rigid geopolitical boundaries. These limitations delay critical decisions during disease outbreaks. Health departments need timely, granular data to allocate resources effectively. Acute outbreaks like dengue or cholera require urgent operational protocols. Conventional methods struggle with the speed and scope needed for these situations.
PDFM uses self-supervised learning to create location fingerprints. It synthesizes diverse signals into compact, versatile embeddings. These inputs refresh at a monthly cadence to stay current. The model compresses privacy-preserving search trends and human mobility data. It also includes built-environment density and environmental determinants. Researchers can plug these embeddings directly into existing machine learning workflows. No new pipelines need to be built from scratch for most applications.
The Mount Sinai Health System tested PDFM for cross-border mobility. They worked with Boston Children's Hospital on this specific case study. Their models captured information spillovers between the U.S. and Canada. This approach explained 36% more variance in vaccination coverage. The data covered 146 counties near the Canadian border. Domestic-only models often failed to predict local vaccine uptake accurately. Adding Canadian Forward Sortation Area embeddings solved this problem.
Researchers refined coverage estimates for 4.7 million border residents. They found local patterns that previous models missed entirely. The share of variation in MMR vaccination coverage rose from 16% to 22%. This represents a significant increase in predictive power. Health officials can now pinpoint communities at risk of measles outbreaks more effectively.
The NYU Grossman School of Medicine evaluated PDFM for cardiovascular disease. They focused on noncommunicable diseases affecting the contiguous United States. The study covered roughly 3,091 counties across the nation. PDFM matched census-based models in accuracy for nowcasting deaths. The mean absolute error was 18.7 deaths per county using PDFM. Traditional ACS data had an error of 19.1 deaths per county. These differences were not statistically significant between the two methods.
A new model called JEPA-Anything works across physics, biology, and medicine. It predicts future states by splitting them into multiple partial parts.
PDFM reduced large county outlier errors by 20%. The root mean square error dropped from 57.7 to 46.0. This improvement made the model more reliable for specific regions. Census data often lags by one or two years before release. ACS covariates reflect conditions up to three years in the past. PDFM uses a single month of data to generate its embeddings. This timeliness allows health departments to act on current conditions.
The University of Oxford and Tecnológico de Monterrey improved dengue forecasts. They integrated PDFM with TimesFM 2.0 for their forecasting model. The partnership covered approximately 2,450 Mexican municipalities over several years. Forecast accuracy improved in up to 72% of active transmission areas. Gains were concentrated in hotspots where timely vector control matters most.
The team measured performance using the Weighted Interval Score metric. Lower scores indicate better forecast accuracy for this specific task. The model achieved a statistically significant one-month forecast gain. Total error reductions were 3.4 times larger than any degradations observed. This sharpness occurred precisely where clinical staffing decisions are critical. Short-term forecasts proved essential for managing vector-borne disease outbreaks.
The University of Washington studied maternal mental health screening risks. They analyzed data from 332,970 respondents in the CDC PRAMS survey. Postpartum depression screening typically relies on limited clinical intake information. This data rarely captures broader community conditions shaping a mother's risk. PDFM embeddings captured community-level socioeconomic conditions with an R-squared value of 0.45.
Adding these embeddings boosted prediction accuracy for high-risk mothers. The area under the curve increased by 0.0020 in seen states. Gains were even higher at +0.0038 in unseen states during testing. This signal recovered about 15% of the predictive signal from income records. It acted as complementary local context when insurance details were unavailable.
Simulations showed PDFM helped reach more rural mothers with screening resources. Health systems could follow up with the 20% highest-risk mothers annually. The model enabled outreach to 5,640 additional rural mothers each year. Alternatively, it cut 17,723 false alarms in systems aiming for 80% case catch rates. This efficiency improves follow-up resources and reduces unnecessary alerts.
The Democratic Republic of the Congo saw significant gains in cholera prediction. Researchers used national surveillance data from the Integrated Disease Surveillance and Response system. They tested a lightweight version of PDFM adapted for sparse internet connectivity. Outbreaks are rare, with fewer than one in 100 health zones seeing cases weekly. This rarity makes anticipating epidemics particularly difficult without advanced tools.
Four to eight weeks out, PDFM produced sharper shortlists of high-risk zones. Response teams work from these short lists to deploy vaccines and clean water. Eight weeks ahead, the model raised correct picks per week from 1.78 to 2.10. This represents an 18% improvement in identifying outbreak locations. Gains were particularly large where cholera is endemic to the region.
In the 15 zones reporting cholera in half of all weeks, Precision@5 improved significantly. The metric measures the share of top five picks that actually had outbreaks. PDFM raised this precision rate for eight-week shortlists in these endemic areas. Early warning allows prepositioning oral cholera vaccines before cases arrive.
PDFM offers timely, privacy-preserving context for global public health decisions. It addresses data gaps and temporal reporting lags in existing workflows. The model provides plug-and-play inputs without needing custom training pipelines. Public health officials can use these tools to guide prevention resources effectively. Decisions regarding acute outbreaks become faster and more accurate with this technology.
Readers can check the Google Research blog for full technical details on PDFM. They should compare PDFM performance against traditional census or survey data metrics. Testing self-supervised learning on diverse geospatial signals is a viable path forward. Engineers can integrate these embeddings into existing machine learning workflows easily. Managers can evaluate the cost savings from reduced custom data engineering needs.
The key takeaway is that planetary geospatial foundation models solve real-world public health problems. They bridge gaps caused by sovereign borders and delayed reporting cycles. Health systems gain access to timely signals for maternal mental health screening. Cholera forecasters benefit from extended lead times for supply prepositioning. Dengue experts see sharper shortlists for vector control operations.
Vaccination coverage models capture cross-border dynamics that domestic data misses. Cardiovascular death predictors now stand in for outdated census inputs effectively. Maternal risk predictors recover signals when income and insurance records are missing. Cholera forecasters identify hotspots weeks before outbreaks begin to materialize.
This technology represents a new paradigm for global health surveillance. It leverages everyday signals like search trends and weather patterns. These inputs capture underlying social, behavioral, and environmental determinants of health. The same embeddings perform well across diverse disease domains and geographic settings.
Researchers demonstrated value across five distinct public health challenges globally. Each case study highlighted specific improvements in predictive accuracy or timeliness. The collective results show PDFM's potential to enhance epidemiological workflows significantly. Health departments can now act on current conditions rather than waiting years for data.
The source text confirms that no custom fine-tuning is required for most applications. Off-the-shelf location embeddings matched or improved conventional inputs across tasks. Privacy-preserving nature of the data streams ensures sensitive information remains protected. Monthly cadence refreshes keep the model relevant for dynamic health situations.
Readers interested in implementation should look for API access to PDFM embeddings. They can test integration with their own machine learning pipelines quickly. Comparing performance metrics against baseline models will validate the technology's impact. Monitoring cross-border and multi-year data gaps will show immediate value propositions.
The future of public health surveillance looks more granular and timely thanks to this innovation. Data silos and reporting lags are becoming less relevant barriers to analysis. Geospatial context becomes a standard input for disease forecasting models. Health systems can direct resources with greater precision and speed.
This shift enables better prevention strategies for chronic and acute conditions alike. Maternal mental health screening becomes more inclusive of rural populations. Cholera response teams deploy vaccines to the right zones faster. Dengue control programs target active transmission hotspots with improved accuracy.
The ability to recover predictive signal from sparse data streams is a major win. Income and insurance records are often unavailable for many vulnerable populations. PDFM fills these gaps with robust geospatial context derived from public signals. This transferable context makes the biggest difference in new states or regions.
Engineers building health agents can now incorporate location fingerprints as core features. These features enhance statistical models without requiring massive labeled datasets. Self-supervised learning creates representations that generalize across different epidemiological tasks. The paradigm of planetary geospatial foundation models is here to stay.
Managers deciding on AI infrastructure should consider PDFM for public health use cases. It reduces the need for extensive custom data engineering pipelines. Cost savings come from leveraging existing search and weather data sources. Speed increases because embeddings refresh monthly instead of waiting for annual surveys.
The technology addresses systemic bottlenecks in global health research directly. Temporal reporting lags no longer paralyze decision-making processes for health officials. Geopolitical boundaries do not limit the analysis of disease spread patterns anymore. Cross-border mobility and information spillovers are captured by the model effectively.
Readers can expect similar improvements in other domains using geospatial data. Urban planning, environmental science, and social research will benefit from this approach too. The foundation model paradigm is expanding beyond just public health applications. Privacy-preserving design ensures these benefits reach sensitive demographic groups safely.
Google Earth AI continues to push the boundaries of what geospatial models can do. Population Dynamics Foundation Model sets a new standard for data integration. Five partner-driven case studies prove its value across diverse environments. The work demonstrates that self-supervised learning solves real-world problems efficiently.
The next step is widespread adoption of these tools in health departments worldwide. Researchers will likely publish more evaluations covering different disease types and regions. Expect to see PDFM integrated into national surveillance systems soon. Timely, granular data will become the norm for public health management.
Reflection released Beam, a 501 billion parameter text-only model for coding and science. Apache 2.0 weights are available this month after training on 23.8 trillion tokens.
