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

Arbaaz Muslim; Aviv Slobodkin; Katherine Wheeler-Martin; Eric Zhou; John Brittain; Martin Frasch; Hamsa Subramaniam; Jacob Bien

cs.LGOct 5, 2026

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

Authors: Arbaaz Muslim, Aviv Slobodkin, Katherine Wheeler-Martin, Eric Zhou, John Brittain, Martin Frasch, Hamsa Subramaniam, Jacob Bien, +30 more

Organizations: Google, Mountain View, CA, USA · Department of Population Health, NYU Grossman School of Medicine, New York, NY, USA · Center for Child Health Services Research, Icahn School of Medicine at Mount Sinai, New York, NY, USA · Department of Pediatrics, Icahn School of Medicine at Mount Sinai, New York, NY, USA · Department of Population Health Science and Policy, Icahn School of Medicine at Mount Sinai, New York, NY, USA · Computational Epidemiology Lab, Boston Children’s Hospital, Boston, MA, USA · Department of Biology, University of Oxford, Oxford, UK · Pandemic Sciences Institute, University of Oxford, Oxford, UK · Institute on Human Development and Disability, University of Washington, Seattle, WA, USA · Health Stream Analytics, LLC, Seattle, WA, USA · KHAI Ventures, Boston, MA, USA · National Autonomous University of Mexico, Mexico City, Mexico · Dirección General de Calidad y Educación en Salud, Secretaría de Salud, Mexico City, Mexico · Departments of Pediatrics and Biomedical Informatics, Harvard Medical School, Boston, MA, USA · World Health Organization Regional Office for Africa (WHO AFRO), Nairobi, Kenya · Emergency Preparedness and Response, World Health Organization Regional Office for Africa (WHO AFRO), Brazzaville, Republic of the Congo · Department of Anaesthesia, Harvard Medical School, Boston, MA, USA

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.

Figures & tables

Figure 1: Overview of Google Earth AI for global public health. Phase 1 encodes multimodal, privacy-preserving data streams (aggregated search trends, maps, busyness, and weather and air quality) via the Population Dynamics Foundation Model (PDFM) into generalizable regional fingerprint embeddings. Phase 2 integrates these place representations with domain-specific labeled health datasets and downstream statistical or machine-learning models across five epidemiological tasks: spatial extrapolation (MMR vaccination coverage), spatial interpolation and nowcasting (cardiovascular disease mortality), probabilistic forecasting (dengue outbreaks), prospective binary forecasting (cholera emergence), and individual-level risk stratification (postpartum depression).

ML Problem	Geographic Scope	Thematic Domain	Benchmark Dataset	Prediction Targets	Train / Test Granularity	Train / Test Size
Spatial Extrapolation	Global North (US, Canada)	Vaccine-Preventable (Immunization)	MMR Vaccination Uptake (ONM/MRP)	County-level MMR vaccination rate	US County / Canadian FSA (postal)	3,080 US counties (10-fold CV); 146 border counties subset
Interpolation	Global North (US)	Non-communicable (Cardiovascular)	NVSS CVD Mortality	County-level CVD death counts	US County (Census-equivalent)	3,091 counties (5-fold CV)
Nowcasting	Global North (US)	Non-communicable (Cardiovascular)	NVSS CVD Mortality	County-level CVD death counts	US County (Census-equivalent)	30,910 county-years (2013–2022) / 3,091 counties (2023)
Probabilistic Forecasting	Global South (Mexico)	Communicable (Dengue)	Mexico Dengue Surveillance for the Mexican Ministry of Health	Monthly dengue case counts	GADM Admin Level 2 (municipality)	$\sim$ 2,450 municipalities $\times$ 68 months (Jan 2020–Aug 2025); 24-month sliding window
Prospective Forecasting	Global South (DRC)	Communicable (Cholera)	DRC National Integrated Disease Surveillance	Binary cholera emergence at 1-, 2-, 4-, and 8-week leads	Health zone $\times$ week	403 health zones; train: 722 weeks (4 Jan 2010–31 Oct 2023) / test: 89 weeks (1 Nov 2023–14 Jul 2025)
Risk Stratification	Global North (US)	Maternal Mental Health (PPD)	CDC PRAMS, Phases 7–8 (births 2012–2021)	Binary postpartum depressive symptoms (CDC indicator: MH_PPDPR or MH_PPINT “always” or “often”)	Respondent; PDFM v0 embedding at state $\times$ urban/rural (87 locations)	332,970 respondents; sealed tests: 66,594 holdout (S1), 10 jurisdiction-held-out folds (S2), 2020–21 births (S3)

Table 1: Summary of Evaluation Benchmarks and Dataset Specifications. Overview of target variables, geographic scopes, and spatial resolutions across the five evaluation paradigms.

Geography	Model	Correlation	RMSE	MAE	$R^{2}$
All US counties	US embeddings only	0.611	0.058	0.045	0.373
	Canadian context only	0.113	0.072	0.059	0.013
	US + Canadian context	0.614	0.057	0.044	0.381
US–CA border counties	US embeddings only	0.399	0.070	0.055	0.159
	Canadian context only	0.381	0.071	0.056	0.146
	US + Canadian context	0.465	0.068	0.051	0.216

Table 2: Predictive performance of models using U.S. and Canadian contextual information to predict county-level MMR vaccination coverage. We show results for all 3,080 US counties (top three rows) and for all 146 counties on the US–Canada border (bottom three rows). For each of these test cases, the best results on each of the reported metrics are in bold.

Figure 2: Spatial distribution of cross-border predictive influence between Canadian postal regions and U.S. counties. The map illustrates where Canadian contextual embeddings contribute to improved prediction of U.S. county vaccination rates. Orange shading indicates U.S. counties where incorporating Canadian contextual embeddings reduces prediction error relative to a model using U.S. embeddings alone; darker shades correspond to larger reductions in absolute prediction error. Blue shading indicates Canadian postal regions whose embeddings exert the greatest influence in the cross-border prediction model, measured by the increase in prediction error that occurs when the information from that region is removed from the model through a leave-one-out procedure.

Figure 3: Cumulative population living in U.S. border counties whose predicted vaccination coverage changes by at least a given threshold after incorporating Canadian contextual information. For each threshold, the corresponding affected population is presented as both population count and the percentage of the total border-county population.

	Bayesian spatial model		Gradient boosted decision tree model
Model configuration	MAE	RMSE	MAE	RMSE
Baseline model	48.44 (44.36, 52.87)	126.87 (111.42, 153.35)	74.64 (62.10, 94.37)	368.57 (171.94, 565.82)
+ ACS	42.07 ∗∗ (36.53, 45.94)	125.89 (87.04, 152.19)	74.02 (60.07, 98.91)	406.02 (174.89, 585.12)
+ PDFM	44.95 (40.59, 50.12)	133.70 (96.90, 165.19)	79.58 (63.74, 107.56)	411.00 (185.25, 597.87)
+ ACS + PDFM	47.77 (38.87, 52.47)	179.09 (101.72, 245.96)	75.81 (60.90, 96.62)	403.91 (166.70, 583.31)

Table 3: Spatial interpolation results showing mean MAE and pooled mean RMSE (in county-level counts of cardiovascular disease deaths; lower values indicate better accuracy) from 20% county holdout tests in 2023. Values in parentheses denote the minimum–maximum range across 5 cross-validation folds for Bayesian models and the 2.5th–97.5th percentile interval across 100 bootstrap replicates for GBDT models. Bold reflects the lowest MAE and RMSE across all models. ** indicates p<0.01 versus Baseline (clustered sign-flip test significant and clustered bootstrap 95% confidence interval for paired error differences excludes zero); all other differences from Baseline, as well as all pairwise differences between ACS and PDFM configurations, are not statistically significant.

	Bayesian spatial model		Gradient boosted decision tree model
Model configuration	MAE	RMSE	MAE	RMSE
Baseline model	23.99	71.22	26.18 (25.51, 27.17)	216.38 (193.67, 240.93)
+ ACS	24.30	72.83	19.13 ∗ (18.63, 19.86)	57.69 ∗∗ (48.79, 71.58)
+ PDFM	24.28	72.65	18.66 ∗ (18.43, 19.22)	46.00 ∗∗ (44.96, 47.61)
+ ACS + PDFM	23.59 ∗∗∗	69.62 ∗∗∗	18.52 ∗∗ (18.39, 19.15)	46.02 ∗∗ (45.14, 49.18)

Table 4: Nowcasting results showing MAE and pooled RMSE (in county-level counts of cardiovascular disease deaths; lower values indicate better accuracy) for the 100% county holdout test in 2023 (trained on 2013–2022). Values in parentheses for GBDT models denote the 2.5th–97.5th percentile interval across 100 random-seed replicates. Bold reflects the lowest MAE and RMSE across all models. * indicates p<0.05, ** indicates p<0.01, and *** indicates p<0.001 versus Baseline (clustered sign-flip test significant and clustered bootstrap 95% confidence interval for paired error differences excludes zero); pairwise differences between ACS and PDFM configurations are not statistically significant.

Figure 4: PDFM improves short-horizon dengue forecasting during active transmission in Mexico. A) Monthly reported dengue cases during the January 2020–August 2025 evaluation period. B) Mean difference in Weighted Interval Score ( $\Delta\text{WIS}=\text{WIS}_{\text{PDFM}}-\text{WIS}_{\text{baseline}}$ ) between PDFM-adjusted and baseline forecasts by forecast target month for 1-, 3-, and 6-month forecast horizons. Negative values indicate improved probabilistic forecast performance with PDFM. C) Mean $\Delta$ WIS stratified by municipality-level dengue case burden. Municipality-months with zero reported cases are shown separately; non-zero observations are divided into low-, medium-, and high-case groups. Shaded regions indicate 95% confidence intervals estimated using a temporal moving-block bootstrap with three-month blocks. D) Mean $\Delta$ WIS by calendar month for 1-, 3-, and 6-month forecast horizons. Shaded regions indicate 95% confidence intervals estimated using the same temporal moving-block bootstrap.

Figure 5: Distribution and cumulative magnitude of time-averaged forecast accuracy differences across Mexican municipalities. A) Histogram displaying the difference in mean Weighted Interval Score ( $\Delta\text{WIS}_{m}$ ) at Horizon 1 between the TimesFM model augmented with PDFM and the baseline TimesFM without PDFM. Differences are computed over the evaluation period from January 2020 to August 2025 across 2,457 municipalities. The solid black vertical line marks zero difference; negative values ( $\Delta\text{WIS}_{m}<0$ ) indicate a lower average forecast error for the model utilizing PDFM. The solid and dashed blue vertical lines indicate the mean and median, respectively. The distribution has a mean difference of $-$ 0.0051 and a median of 0.0013. Overall, the inclusion of PDFM improved average forecast performance in 47.6% (1,170 out of 2,457) of the evaluated municipalities. B) Cumulative magnitude of municipality-level mean $\Delta$ WIS after separately ranking municipalities with improved and worsened performance from largest to smallest effect magnitude. Gross improvement, defined as the sum of absolute mean $\Delta$ WIS among municipalities with $\Delta$ WIS < 0, was 17.61, compared with gross deterioration of 5.10 among municipalities with $\Delta$ WIS > 0. Thus, although deterioration occurred in slightly more municipalities, improvements were substantially larger in magnitude.

Figure 6: Geographic distribution of reported dengue burden and change in Weighted Interval Score (WIS) applying PDFM embeddings in Mexico. A) Mean monthly reported dengue case burden by municipality over the evaluation period. Values are displayed as log(1 + mean monthly cases). B) Mean difference in Weighted Interval Score ( $\Delta\text{WIS}=\text{WIS}_{\text{PDFM}}-\text{WIS}_{\text{baseline}}$ ) for one-month-ahead forecasts by municipality over the same evaluation period. Negative $\Delta$ WIS indicates improved forecast performance with PDFM relative to baseline. Across 2,457 municipalities, 1,170 (47.6%) had negative mean $\Delta$ WIS; the median municipality-level $\Delta$ WIS was 0.0013 (interquartile range $-$ 0.0111 to 0.0035), indicating that the overall improvement was driven by larger gains in a subset of municipalities rather than uniform improvement across locations.

Lead Time	Metric	History-only	History+PDFM	$\Delta$ (%)	95% CI	p
1 wk	PR-AUC	0.2931	0.2969	$+$ 0.0038 ( $+$ 1.3)	$-$ 0.0161 to $+$ 0.0258	0.72
	Precision@5	0.2841	0.2614	$-$ 0.0227 ( $-$ 8.0)	$-$ 0.0432 to $-$ 0.0023	0.103
2 wk	PR-AUC	0.2990	0.2860	$-$ 0.0129 ( $-$ 4.3)	$-$ 0.0304 to $+$ 0.0055	0.23
	Precision@5	0.3011	0.3126	$+$ 0.0115 ( $+$ 3.8)	$-$ 0.0138 to $+$ 0.0368	0.52
4 wk	PR-AUC	0.2832	0.3107	$+$ 0.0275 ( $+$ 9.7)	$+$ 0.0089 to $+$ 0.0479	0.019
	Precision@5	0.3435	0.3906	$+$ 0.0471 ( $+$ 13.7)	$+$ 0.0141 to $+$ 0.0776	0.019

Table 5: Prospective emergence forecasting performance, all health zones. History+PDFM versus History-only, predictions pooled over five random seeds. Lead time is the prediction horizon.

\Delta

is History+PDFM minus History-only. CIs are unadjusted percentile bootstraps while p-values are Benjamini–Hochberg adjusted within this table.

Figure 7: PDFM in individual-level postpartum-depression risk prediction (PRAMS, 332,970 mothers). A) Change in discrimination ( $\Delta$ AUC, 95% CI) from adding PDFM embeddings. Rows: seen states (S1, pre-registered primary); unseen states (S2); later years (S3); real vs random embedding in unseen states; PDFM added beyond individual income and Medicaid; and PDFM in place of them. B) Rural mothers in unseen states (exploratory; US-weighted, per year): change in PPD cases found and in false alarms under two screening policies. Fixed screening capacity flags the 20% of mothers at highest predicted risk. The fixed target uses a threshold set on training data for $\geq$ 80% sensitivity. C) Sensitivity by group under the 80% rule, without (blue) and with (orange) PDFM. The dashed line marks 80%. Error bars are 95% bootstrap CIs; S2 CIs resample jurisdictions.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Min. change (pp)	Border counties	% Border counties	Affected population	% Border pop.	% Total pop.
10.0	3	2.1	241,045	1.3	0.1
7.5	9	6.2	1,672,517	9.0	0.5
5.0	24	16.4	2,502,339	13.4	0.8
4.0	35	24.0	3,529,624	19.0	1.1
3.0	55	37.7	4,706,150	25.3	1.5
2.0	74	50.7	8,540,213	45.9	2.6

Appendix

Extended Data Table 1: Cumulative population in U.S. border counties affected by incorporating Canadian contextual information, by threshold of change in predicted MMR vaccination coverage.

Regime	Lead Time	Metric	History-only	History+PDFM	$\Delta$ (%)	95% CI	p
Endemic	1 wk	PR-AUC	0.4679	0.4748	$+$ 0.0069 ( $+$ 1.5)	$-$ 0.0388 to $+$ 0.0494	0.99
Endemic	1 wk	Precision@5	0.2318	0.2250	$-$ 0.0068 ( $-$ 2.9)	$-$ 0.0159 to $+$ 0.0023	0.74
Endemic	2 wk	PR-AUC	0.4637	0.4444	$-$ 0.0193 ( $-$ 4.2)	$-$ 0.0611 to $+$ 0.0251	0.75
Endemic	2 wk	Precision@5	0.2621	0.2644	$+$ 0.0023 ( $+$ 0.9)	$-$ 0.0046 to $+$ 0.0115	1.00
Endemic	4 wk	PR-AUC	0.4209	0.4729	$+$ 0.0521 ( $+$ 12.4)	$+$ 0.0172 to $+$ 0.0887	0.024
Endemic	4 wk	Precision@5	0.3082	0.3200	$+$ 0.0118 ( $+$ 3.8)	$-$ 0.0024 to $+$ 0.0259	0.47

Appendix

Extended Data Table 2: Performance stratified by pre-evaluation endemicity regime

Setting	Model	AUC (95% CI)	Brier score	Observed / expected	Calibration slope
S1: seen states (primary)	Baseline	0.622 (0.616–0.629)	0.1174	1.00	1.05
	PDFM (replaces state)	0.624 (0.618–0.630)	0.1173	1.00	1.02
	Baseline + income, Medicaid	0.636 (0.630–0.642)	0.1168	1.00	1.06
	Income, Medicaid + PDFM	0.637 (0.631–0.643)	0.1167	1.00	1.10
S2: unseen states	Baseline	0.615 (0.604–0.625)	0.1177	1.00	1.02
	PDFM (replaces state)	0.618 (0.607–0.630)	0.1176	1.01	1.06

Appendix

Extended Data Table 3: Test-set performance (unweighted)

Setting	Comparison	Estimate (95% CI)	Pre-defined reading	Holm p (S1)
S1: seen states (primary)	PDFM vs baseline	$+$ 0.0020 ( $+$ 0.0008 to $+$ 0.0031)	improvement	0.012
	PDFM instead of income and Medicaid (non-inferiority, margin $-$ 0.005)	$-$ 0.0116 ( $-$ 0.0144 to $-$ 0.0089)	harm	1.000
	Share of the income/Medicaid gain recovered by PDFM ( $R$ )	0.15 (0.06 to 0.24)	$R$ not above 0.5	—
	PDFM vs area poverty share	$+$ 0.0022 ( $+$ 0.0011 to $+$ 0.0033)	improvement	—
	PDFM added to income and Medicaid	$+$ 0.0011 ( $+$ 0.0000 to $+$ 0.0022)	improvement	0.104
S2: unseen states	PDFM vs baseline	$+$ 0.0038 ( $+$ 0.0004 to $+$ 0.0072)	improvement	—

Appendix

Extended Data Table 4: Pre-registered comparisons (

\Delta

AUC unless stated)

		Fixed screening capacity (20%)		80% sensitivity target
Setting (rural, prev.)	Group	Cases found	False alarms	Cases found	False alarms
US (observed)	rural	$+$ 157 ( $+$ 72 to $+$ 246)	$+$ 653 ( $+$ 249 to $+$ 1,093)	$-$ 48 ( $-$ 87 to $-$ 18)	$-$ 493 ( $-$ 891 to $-$ 193)
	urban	$-$ 70 ( $-$ 171 to $+$ 32)	$-$ 740 ( $-$ 1,188 to $-$ 321)	$+$ 50 ( $-$ 38 to $+$ 134)	$+$ 288 ( $-$ 579 to $+$ 1,040)
	all births	$+$ 87 ( $-$ 13 to $+$ 191)	$-$ 87 ( $-$ 191 to $+$ 12)	$+$ 3 ( $-$ 114 to $+$ 102)	$-$ 205 ( $-$ 1,452 to $+$ 735)
World (42.7%, 17.2%)	rural	$+$ 337 ( $+$ 137 to $+$ 545)	$+$ 948 ( $+$ 329 to $+$ 1,574)	$-$ 175 ( $-$ 305 to $-$ 63)	$-$ 1,300 ( $-$ 2,097 to $-$ 521)
	urban	$-$ 214 ( $-$ 379 to $-$ 65)	$-$ 1,070 ( $-$ 1,667 to $-$ 433)	$+$ 45 ( $-$ 35 to $+$ 119)	$+$ 187 ( $-$ 371 to $+$ 676)

Appendix

Extended Data Table 5: Screening simulator: PDFM vs baseline transported to other settings (exploratory)

Step	Remaining	Excluded
Phases 7–8 respondents	390,320	—
Both outcome items answered	380,689	9,631
Core predictors complete	356,626	24,063
NCHS urban/rural known (1 or 2)	350,723	5,903
Jurisdiction present in PDFM (absent: AK, HI)	332,970	17,753
Urban/rural stratum present in PDFM	332,970	0

Appendix

Extended Data Table 6: Cohort flow

Explore similar work

May 3, 2026cs.LG

Geospatial foundation-model embeddings improve population estimation unevenly across space and scale

Reliable subnational population estimates are essential for applications, yet remain difficult where censuses are sparse, outdated or spatially coarse. Existing population-mapping workflows rely on hand-built geospatial covariates, such as settlement extent, night-time lights, and environmental conditions, which must be assembled and harmonised across scales and geographies. Geospatial foundation models offer an alternative by learning reusable representations of place from more multifaceted and heterogeneous data sources. Here, we benchmark Population Dynamics Foundation Model (PDFM) embeddings against the harmonised geospatial covariates for subnational population estimation in Brazil, Nigeria and the United States. Under geographically structured validation, PDFM increased predictive fit by a median of 20.1% (IQR: 10.0-33.2%, across country-model comparisons) reduction in unexplained variance, and reduced Kullback-Leibler divergence by 23.2% (9.2-26.2%). However, these gains were uneven. PDFM was most advantageous where the geospatial covariates weakly characterised settlement context, such as larger and less-developed subnational areas. Moreover, PDFM performance was scale-coupled with embeddings providing less flexible transfer across spatial aggregations than geospatial covariates. These findings showed that geospatial foundation-model representations of place can improve population estimation in data poor settings, but their benefits break down predictably under spatial scale mismatch, revealing a fundamental limitation of current geospatial AI.

Wenbin Zhang, Eimear Cleary, Francisco Rowe +4

WorldPop, School of Geography and Environmental Sciences, University of Southampton, United Kingdom · Geographic Data Science Lab, Department of Geography and Planning, School of Environmental Sciences, University of Liverpool, United Kingdom

Geospatial Foundation Models Populations Geospatial Covariates Population Dynamics Geometric Bias

Ask PDF

Sep 10, 2026stat.AP

Geospatial Foundation Models Capture Health-Relevant Dimensions of Place Beyond Conventional Social Risk Indices

Area-based social risk indices summarize residents' socioeconomic conditions but incompletely capture physical features of place that may affect health. We evaluated whether numerical representations of physical place produced by four geospatial foundation model families from 2022 satellite data explained residual variance in tract-level associations between the Area Deprivation Index, Social Deprivation Index, and Social Vulnerability Index with health outcomes. We used LightGBM to predict variables from the American Community Survey and 40 chronic disease and health-behavior outcomes from CDC PLACES across 82,646 census tracts in the contiguous United States, evaluating performance across 10 held-out states. Among survey variables, models were moderately predictive of some variables including housing type (R-squared up to 0.54) but weak for disability, unemployment, and income disparity. For health outcomes, models explained up to 54% of variance left unexplained by social risk indices, with the largest gains for annual checkups, arthritis, and high blood pressure. Mean total variance explained by geospatial foundation models across the 40 health-related outcomes increased from 0.31 in the smallest tract-size decile to 0.39 in the largest. Geospatial foundation models capture health-relevant features of place not represented by conventional social risk indices and may usefully augment them in epidemiological analyses.

Nathaniel Hendrix, Carl Y. Zhang, Chris Heitzig +2

Center for Professionalism and Value in Health Care, American Board of Family Medicine 1016 16th St NW Ste 800 Washington, DC 20036 United States of America · Stanford Center on Longevity, Stanford University Palo Alto, CA USA · Center for Population Health Sciences, School of Medicine, Stanford University Palo Alto, California, United States of America +1

Geospatial Foundation Models Cities

Ask PDF

Jun 20, 2026cs.LG

Beyond Time Series: Spatial Reasoning for Epidemic Forecasting via Multimodal Learning

Epidemic forecasting models typically rely on surveillance data reported over administrative regions, treating them as atomic units, thereby obscuring sub-regional spatial structure that shapes disease dynamics. We introduce a spatially structured multimodal epidemic forecasting setting that integrates region-level temporal surveillance data with spatially localized auxiliary signals that are misaligned in resolution and structure, reflecting realistic public health reporting constraints. Building on this formulation, we propose M-SPICE (Multimodal SPatIal Context for Epidemic Forecasting), a structure-aware spatiotemporal forecasting framework that performs joint reasoning over temporal disease dynamics and spatial context via attention-based multimodal fusion, allowing spatial signals to selectively condition temporal representations across forecast horizons. We evaluate our approach on real-world COVID-19, influenza, and influenza-like illness (ILI) forecasting tasks under realistic real-time evaluation protocols. Across all forecasting settings, our method consistently outperforms state-of-the-art multivariate time-series, multimodal, and epidemiological forecasting baselines while maintaining strong probabilistic forecasting performance. Finally, interpretability analyses reveal when, where, and how spatial signals are leveraged, highlighting settings in which purely temporal, region-aggregated models are most likely to fail.

Diana Guadalupe Gomez, Chenwei Wu, Zhiyi Wang +2

University of Michigan Ann Arbor, MI, USA

Epidemiological Models Spatio-Temporal Forecasting Temporal Dynamics Time Series Multimodal Learning Spatial Reasoning

Ask PDF