Planetary Geospatial Foundation Models: A New Paradigm for Global Public Health

Abstract: The efficacy of traditional disease prediction is limited by spatial gaps and temporal lags, which impact the timing and targets of resource deployments. Outbreaks escalate undetected, chronic disease burdens are quantified years later, and at-risk populations in data-sparse regions remain unaddressed. Planetary geospatial foundation models complement existing epidemiological workflows to provide operational improvements, encoding multimodal search, mobility, and environmental signals into generalizable place representations. As illustrations of this complementarity, we present independent global health case studies of Google Earth AI's Population Dynamics Foundation Model (PDFM) -- a foundation model for geospatial inference -- across four domains (vaccine-preventable, communicable, noncommunicable, maternal mental health), five tasks (spatial extrapolation, interpolation/nowcasting, probabilistic forecasting, prospective forecasting, risk stratification), and four countries (USA, Canada, Mexico, and the Democratic Republic of the Congo). Across these case studies, PDFM addresses critical surveillance gaps across domains: improving US-Canada border MMR vaccination coverage predictions by capturing cross-border behavioral spillovers domestic models miss; nowcasting cardiovascular disease to accelerate data availability; enhancing short-term municipal Mexican dengue forecasts for timely outbreak vector control; improving forecasts of cholera hotspots; and adding a transferable signal to individual-level postpartum-depression risk prediction in US states the model had never seen, while not replacing individual socioeconomic data or closing demographic screening gaps. Together, these results showcase capabilities of geospatial foundation models for public health surveillance.

• Planetary Geospatial Foundation Models (PDFMs) as reusable, multimodal location embeddings for public-health algorithms offer predictive and transferable benefits in cross-border vaccination, cardiovascular disease mortality, dengue forecasting, and cholera hotspot prediction, improving statistical accuracy by 13 to 25%.
• On individual-level risk stratification for postpartum depression showed a small but significant improvement of 0.0020 in AUC, dependent on socioeconomic and data availability factors.
• Their improved performance across five disease-related tasks proved greatest at ‘short’ time horizons, ideally 1-2 months, and performed relatively less so at longer-horizon or low-income and varied disease-related tasks such as dengue.

Problem formulation and methodological contribution

The paper presents planetary geospatial foundation models as reusable covariate generators for public-health surveillance. Its central empirical question is not whether a single foundation model can replace epidemiological models, but whether a fixed representation of place can improve heterogeneous downstream tasks when integrated into established statistical and machine-learning pipelines. The evaluation is organized across five tasks: cross-border spatial extrapolation, spatial interpolation, temporal nowcasting, probabilistic forecasting, prospective outbreak forecasting, and individual-level risk stratification. The applications span MMR vaccination in the United States and Canada, cardiovascular disease mortality in the United States, dengue in Mexico, cholera in the Democratic Republic of the Congo, and postpartum depression in the United States (2610.05699).

PDFM encodes aggregated search trends, built-environment structure, mobility-derived place busyness, meteorology, and air-quality measurements into a 330-dimensional location embedding. The representation is learned through self-supervised reconstruction: modality-specific prediction heads reconstruct the original input streams from the shared embedding. Search-related features occupy 128 dimensions, built-environment and busyness features occupy another 128, and environmental determinants occupy 74. Downstream models use these embeddings as fixed covariates without task-specific fine-tuning. This design is operationally important because it preserves compatibility with ridge regression, Bayesian disease-mapping models, gradient-boosted trees, and time-series forecasters rather than requiring public-health agencies to replace their existing modeling infrastructure.

The paper’s conceptual contribution is therefore architectural as much as predictive. It separates representation learning from disease-specific supervision and treats geospatial context as a transferable, multimodal latent variable. The underlying premise is consistent with the earlier PDFM formulation, which proposed general geospatial inference from population-dynamics signals (Agarwal et al., 2024). The present work tests whether that premise survives across disease classes, geographic scales, outcome types, and resource settings.

Figure 1: PDFM converts multimodal, privacy-preserving signals into regional embeddings that are supplied to task-specific public-health models.

Cross-border vaccination prediction

The first case study tests whether geospatial representations can cross national boundaries in a way that conventional domestic surveillance cannot. The target is county-level MMR vaccination coverage among children under five in 3,080 U.S. counties, with particular attention to 146 counties near the Canadian border. For each U.S. county, Canadian Forward Sortation Area embeddings within 150 km are aggregated using inverse-distance weights and appended to the local U.S. embedding. Ridge regression is evaluated using U.S. information alone, Canadian contextual information alone, or both.

At the national scale, Canadian context alone is predictably weak: its correlation with vaccination coverage is 0.113 and its $R^2$ is 0.013, compared with 0.611 and 0.373 for the U.S.-only model. This result is important because it rejects the strongest possible interpretation of cross-border transfer: Canadian representations do not substitute for domestic information. Their value is conditional on geographic proximity and an already informative domestic representation.

Among border counties, however, adding Canadian information increases correlation from 0.399 to 0.465, reduces RMSE from 0.070 to 0.068, reduces MAE from 0.055 to 0.051, and increases $R^2$ from 0.159 to 0.216. The approximately 36% relative increase in explained variance is the strongest result in this case study. Permutation tests over 999 reassigned contextual embeddings yielded empirical $p=0.001$ for the improvements in $R^2$, MAE, and RMSE, and the gains were reproduced across 100 alternative cross-validation assignments.

The operational effect is larger than the aggregate error changes might suggest. Incorporating Canadian context changes predicted vaccination coverage by at least five percentage points in 24 counties containing approximately 2.5 million residents, or 13.4% of the border-county population. At a one-percentage-point threshold, 13.1 million residents, or 70.5% of the border-county population, are affected. The result implies that boundary-constrained surveillance can materially mischaracterize local vaccination risk even when the neighboring country is not itself used as a direct outcome source.

Figure 2: Cross-border predictive influence is concentrated in geographically connected regions of southern Ontario, southern Quebec, southwestern British Columbia, and adjacent U.S. counties.

Figure 3: The population affected by incorporating Canadian contextual information increases sharply as the threshold for a prediction change decreases.

The spatial influence analysis identifies southern Ontario, southern Quebec, and southwestern British Columbia as major contributors to predictive performance. These are plausible regions for cross-border behavioral spillovers, but the analysis remains predictive rather than causal. The embeddings demonstrate useful cross-border dependence; they do not establish which mobility, media, economic, or environmental mechanisms generate it.

Cardiovascular mortality interpolation and nowcasting

The second case study addresses a different surveillance failure: temporal latency. County-level CVD mortality data can lag by one to two years, while ACS covariates are based on five-year pooled surveys and may be released up to 12 months after collection. The authors compare PDFM embeddings with ten ACS socioeconomic and demographic variables in 3,091 contiguous U.S. county-equivalent units. Two model classes are used: negative-binomial Bayesian spatial models with BYM2 effects and XGBoost gradient-boosted trees.

For spatial interpolation of 2023 mortality counts, the Bayesian model with ACS covariates obtains the lowest MAE, 42.07 deaths, compared with 44.95 for PDFM and 48.44 for the no-covariate baseline. The ACS improvement over baseline is statistically significant, whereas the PDFM improvement is not. Direct ACS–PDFM comparisons are also not statistically significant. Thus, the paper’s claim is not that PDFM outperforms demographic surveys in cross-sectional interpolation. Rather, it is that a single-month representation can achieve statistically comparable performance to substantially slower conventional covariates.

The result is weaker and more nuanced for RMSE. County mortality counts are highly population-skewed, so a small number of large counties dominate squared error. Adding all 330 PDFM dimensions without sufficient shrinkage increases variance in some configurations, particularly in the Bayesian model with ACS and PDFM jointly. This demonstrates that the embedding’s dimensionality is not automatically benign when the downstream sample size is only a few thousand spatial units.

Nowcasting produces a more favorable result. In the Bayesian model, the combination of ACS and PDFM reduces MAE from 23.99 to 23.59 and RMSE from 71.22 to 69.62, with both improvements significant at $p<0.001$. In GBDT models, PDFM alone reduces MAE from 26.18 to 18.66 and RMSE from 216.38 to 46.00. The combined configuration obtains MAE 18.52 and RMSE 46.02. These large GBDT gains indicate that PDFM can provide information useful for updating mortality estimates when contemporaneous outcome counts are unavailable.

The interpretation must remain bounded. The direct comparisons between ACS-augmented and PDFM-augmented models are not statistically significant. PDFM therefore appears to be a timely alternative or complement to ACS rather than a demonstrated replacement in predictive quality. Its main advantage is temporal availability: a single-month representation can be refreshed monthly, whereas ACS measures for smaller counties depend on 60 months of pooled data and delayed release.

Dengue forecasting and the temporal scale of utility

The dengue analysis evaluates approximately 2,450 Mexican municipalities from January 2020 through August 2025 at one-, three-, and six-month horizons. TimesFM supplies the baseline probabilistic forecasts, while PDFM embeddings enter horizon- and quantile-specific residual models trained with a causal 24-month sliding window. Performance is measured using WIS, which jointly evaluates calibration and sharpness.

PDFM significantly improves one-month forecasting. The mean difference in WIS is $-0.0051$, with a 95% moving-block bootstrap interval from $-0.0081$ to $-0.0027$ and a Diebold–Mariano test of $p<0.001$. The effect is robust to three-, six-, and twelve-month block lengths. At three months, the difference reverses direction to $+0.0028$ with $R^2$0; at six months it is $R^2$1 with $R^2$2. The evidence therefore supports a short-horizon benefit but not a general improvement across forecasting horizons.

The heterogeneity analysis explains why the overall result is not uniform. At one month, PDFM has little effect in municipality-months with zero reported cases, but reduces WIS by approximately 0.0172–0.0266 in active-transmission strata. The benefit is concentrated where operational intervention is most relevant. Nevertheless, only 1,170 of 2,457 municipalities, or 47.6%, improve on average. The median municipality-level change is slightly adverse at 0.0013, while the aggregate mean is favorable because improvements are larger than deteriorations: gross improvement is 17.61 versus gross deterioration of 5.10.

Figure 4: PDFM improves dengue forecasts primarily at short horizons and in municipalities with active transmission.

Figure 5: Municipality-level effects are heterogeneous; a minority of municipalities account for disproportionately large gains.

Figure 6: Dengue burden and PDFM-associated forecast improvement exhibit substantial geographic heterogeneity across Mexican municipalities.

The paper offers a useful disease-relative interpretation of this horizon dependence. Static embeddings can encode durable spatial suitability for dengue but cannot anticipate future meteorological or vector-population transitions at six months. At intermediate horizons, they can also induce overprediction during zero-case periods. The October 2025 embedding sensitivity analysis produces nearly the same one-month improvement as the October 2023 embedding, suggesting that the observed signal is relatively stable spatial structure rather than a transient snapshot effect. However, because historical embeddings were unavailable, the study does not demonstrate that monthly time-varying embeddings improve forecasts during rapidly changing local conditions.

Prospective cholera hotspot prediction in the DRC

The cholera analysis tests whether geospatial context is useful in a low-resource setting and at weekly operational lead times. The dataset covers 403 health zones and 811 weeks of surveillance from 2010 through July 2025. Because sparse digital coverage made the full self-supervised PDFM infeasible, the authors use a custom 200-dimensional representation obtained through PCA after pooling search signals over six months. The evaluation is strictly prospective, with a cutoff of 1 November 2023 and no evaluation-period information entering the embeddings or feature construction.

The target is emergence of a three-week outbreak window containing at least 100 suspected cases after a quiet week. Emergence is rare, comprising 0.92% of zone-weeks at a one-week lead and 2.40% at an eight-week lead. Models use either recent surveillance history alone or history augmented with PDFM embeddings.

At one and two weeks, PDFM provides no statistically reliable advantage. Recent case history is sufficiently informative at these short horizons. At four weeks, however, PR-AUC increases from 0.2832 to 0.3107, a 9.7% relative improvement with a corrected $R^2$3. Precision@5 rises from 0.3435 to 0.3906, a 13.7% relative improvement, also with $R^2$4. At eight weeks, PR-AUC increases from 0.2733 to 0.2899, although the corrected $R^2$5 is 0.064, while Precision@5 increases from 0.3556 to 0.4198, an 18.1% improvement with $R^2$6.

These ranking results are operationally more interpretable than probability discrimination alone. At an eight-week horizon, the model identifies approximately 2.10 true events per week among the five highest-risk zones, compared with 1.78 for history alone. The largest gains occur in endemic zones: four-week PR-AUC rises from 0.4209 to 0.4729, and eight-week Precision@5 rises from 0.3333 to 0.3975. The result supports a specific temporal pattern shared with dengue: geospatial context adds most value after short-term persistence in case counts has decayed but before longer-term seasonal and environmental transitions dominate.

The DRC implementation also exposes a significant constraint. Internet penetration and digital activity are limited, requiring six-month search aggregation and PCA rather than the standard encoder. Consequently, the DRC experiment is not a direct test of the same PDFM representation used elsewhere. It is evidence for the usefulness of a related multimodal geospatial representation under sparse-data adaptation, not evidence that the full standard model can be deployed unchanged in all low-connectivity settings.

Individual-level postpartum-depression risk prediction

The final case study establishes the boundary between place-level context and individual-level risk. The analysis uses 332,970 PRAMS respondents from 2012–2021, linked to PDFM embeddings at state-by-urban/rural resolution across 87 locations. The prediction task is binary postpartum depressive symptoms. Individual predictors include education, smoking during pregnancy, pre-pregnancy BMI, marital status, and state; income and Medicaid status are added in prespecified comparisons. Race and ethnicity are excluded from model inputs but used for performance auditing.

The primary gain is statistically significant but small. In states represented during training, PDFM increases AUC by 0.0020, from a baseline near 0.62, with a confidence interval of 0.0008–0.0031 and Holm-adjusted $R^2$7. In states held out entirely from training, the gain is larger at 0.0038, with a 0.0004–0.0072 interval. Real embeddings outperform random embeddings by 0.0047 AUC, supporting the claim that the transferred signal is attributable to embedding content rather than dimensionality or regularization alone.

The model does not substitute for individual socioeconomic information. Replacing income and Medicaid status with PDFM decreases AUC by 0.0116, and PDFM recovers only 15% of the socioeconomic-information gain. Adding PDFM beyond income and Medicaid yields only +0.0011 AUC, with Holm-adjusted $R^2$8. The finding is therefore explicitly contradictory to any claim that coarse geographic context can replace patient-level socioeconomic variables.

Figure 7: PDFM adds a small transferable signal to postpartum-depression prediction but does not replace individual socioeconomic data or eliminate demographic performance disparities.

The fairness analysis is similarly restrictive. None of 21 preregistered race-by-ethnicity-by-urbanicity gaps closes in seen or unseen states. PDFM changes the allocation of screening primarily across rural and urban populations. Under a fixed capacity that flags the 20% highest-risk mothers, the model identifies an estimated 5,640 additional rural cases per year but produces 23,474 additional false alarms. Under a fixed 80% sensitivity threshold, it avoids 17,723 false alarms but identifies 1,722 fewer rural cases. These are not intrinsic model benefits; they are policy-dependent reallocations induced by the chosen screening rule.

The Asian-mother sensitivity gap remains particularly severe: under the 80% sensitivity rule, all models detect fewer than half of cases, with sensitivity of 39.8%. This result directly limits the paper’s broader claims about equitable screening. Place-level context redistributes risk between locations, but most individual-level signal and the largest disparities remain within locations.

The exploratory international screening simulations should be interpreted cautiously. They estimate how the U.S.-trained score distribution would behave under alternative rural shares and PPD prevalences, but both PRAMS and PDFM v0 are U.S.-restricted. The simulations assume transportability of score distributions and therefore cannot establish performance in Southern Asia, Southern Africa, or other settings.

Limitations and open questions

The principal limitation across the case studies is temporal misalignment. Most analyses use static embeddings, primarily from October 2023, because historical PDFM sequences were unavailable. The observed results therefore demonstrate the value of cross-sectional and relatively stable place representations, not the claimed operational benefit of monthly updates during acute shocks. The dengue snapshot sensitivity analysis is reassuring, but it does not substitute for longitudinal evaluation in which embeddings are aligned with each outcome period.

The embedding dimensions are also difficult to interpret epidemiologically. Unlike ACS variables, individual dimensions do not correspond to explicit constructs such as income, unemployment, or age structure. The embeddings support prediction but are poorly suited to direct etiologic interpretation without modality-level attribution or analyses of their relationship to labeled socioeconomic variables. This limits their use in causal analysis, descriptive epidemiology, and policy explanation.

The cross-study comparisons are not fully standardized. Outcomes, spatial units, forecast horizons, baselines, and representation construction differ across applications. The cholera analysis uses PCA-compressed features rather than the standard self-supervised model, and the postpartum-depression analysis uses an older PDFM release with embeddings that postdate the births. These design choices are defensible given data constraints, but they mean that the paper establishes a family of related capabilities rather than a single uniform performance benchmark.

Finally, the paper relies on proprietary pretraining code and restricted health datasets. Public replication materials include downstream examples and tutorials, but complete independent replication is not possible from the released artifacts alone. Questions therefore remain about performance under longitudinal embedding availability, robustness to changes in digital-platform behavior, calibration under distribution shift, coverage in low-connectivity regions, and the comparative value of modality-specific representations.

Conclusion

The paper provides a broad empirical assessment of geospatial foundation-model covariates in public-health surveillance. PDFM improves cross-border vaccination prediction, matches ACS covariates in several CVD settings while offering substantially fresher inputs, improves short-horizon dengue forecasting in active-transmission locations, increases four- to eight-week cholera hotspot ranking performance in the DRC, and contributes a small transferable signal to postpartum-depression prediction. These gains are conditional rather than universal: they depend on spatial scale, disease dynamics, forecast horizon, data availability, and the information already present in individual-level predictors.

The most defensible conclusion is that planetary geospatial foundation models function as complementary representations of place. Their strongest demonstrated value lies in filling spatial and temporal surveillance gaps within existing epidemiological workflows, particularly over the one-to-two-month interval in which recent case history has weakened but longer-term seasonal transitions have not yet dominated. The paper does not show that such representations replace demographic surveys, individual socioeconomic data, longitudinal outcome surveillance, or fairness-oriented screening design.