PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Hamed Sadeghi, Sarah Conderino, Stone Jiang, Ben Spoer, Daniel Neill, Mimi Sun, Joydeep Paul, Yun Liu, Tomer Shekel, Amy Chung-Yu Chou, Anna Carter, Charles Elliott, Aaron Bell, Reuven Sayag, Avinatan Hassidim, Niv Efron, Yossi Matias, Kasumi Widner, Eduardo Lopez Ortiz, José Alberto Díaz-Quiñonez, Moritz U. G. Kraemer, John Brownstein, Isaias Fernandes Co, Etien Luc Koua, Monica Bharel, Benjamin Rader, Lorna Thorpe, Marc Gourevitch, Shravya Shetty, Gautam Prasad

arXiv:2610.05699v1cs.LGcs.CY

TL;DR

Conventional public-health prediction faces spatial gaps and temporal lags that constrain resource allocation. This paper evaluates PDFM geospatial representations across global health cases, finding surveillance benefits while noting that monthly-refresh gains and individual-level screening boundaries remain important qualifications.

  • Problem

    Public-health prediction is constrained by spatial gaps and temporal lags, including delayed fine-scale mortality data that hinder allocation of preventive resources.

    Source: §1 · p. 1, §2.2 · p. 6

  • Method

    The paper evaluates PDFM geospatial representations, built from multimodal signals, as covariates across five independent public-health case studies.

    Source: §3 · pp. 16, 18

  • Results

    36% relative increase in explained variance for near-border vaccination prediction accompanied broader surveillance gains across five cases, including corrections in dengue hotspots and endemic cholera zones.

    Source: §3 · pp. 16–17

  • Takeaways & Limitations

    Across the cases examined, modest error reductions can materially improve allocation of scarce public-health resources and correct high-stakes outlier predictions.

    Source: §3 · p. 16

    The experiments do not demonstrate that monthly refreshed covariates capture within-location change during rapid shifts, so their operational advantage remains a hypothesis.

    Source: §3 · p. 19

Abstract

from arXiv · show

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.

1. Introduction

Conventional epidemiological data can be stale, spatially coarse, or bounded by national borders, constraining timely local surveillance. The section presents PDFM as a transferable source of geospatial covariates and previews evaluations across diseases, tasks, and settings.

1. Introduction

  • Conventional data streams are often updated infrequently, lack fine-resolution disease outcomes in real time, and stop at national borders, limiting disease tracking and prediction.These gaps affect cross-border vaccination surveillance and access to timely county-level cardiovascular mortality data.Source: §1 · p. 2
  • PDFM uses self-supervised pretraining on multimodal behavioral, mobility, built-environment, meteorological, and air-quality signals to produce transferable place representations.Its geospatial feature vectors can be added as covariates to existing public-health statistical and machine-learning workflows.Source: §1 · p. 2
  • Independent partner evaluations span epidemiological tasks, geographic and resource settings, and communicable and noncommunicable diseases across four countries.The cases include the United States, Canada, Mexico, and the DRC.Source: §1 · pp. 2–3
  • For border vaccination extrapolation, PDFM captures behavioral spillovers that domestic-only models miss; for cardiovascular disease, monthly covariates offer a timely update between census cycles.The cardiovascular case examines spatial interpolation and temporal nowcasting of mortality.Source: §1 · p. 2
  • In Mexico, PDFM corrects large dengue prediction errors in outbreak hotspots, while in the DRC it supplements historical data for cholera hotspot forecasts with 4- to 8-week lead time.Source: §1 · p. 3
  • For postpartum-depression risk prediction, PDFM adds a signal in unseen US states but does not replace individual socioeconomic data or reduce demographic biases.Source: §1 · p. 3

2. Results

The Results section evaluates fixed PDFM embeddings, without task-specific fine-tuning, as covariates across five case studies, epidemiological tasks, and statistical and machine-learning paradigms. It examines hotspot prediction in endemic zones and postpartum-depression risk prediction in unseen states, including whether the embeddings substitute for individual socioeconomic data or close demographic performance gaps.

2.1. Case Study 1: Cross-border geospatial embeddings capture behavioral spillovers to improve localized immunization forecasting

  • In 146 U.S.–Canada border counties, adding Canadian embeddings increased R2 from 0.159 to 0.216, a roughly 36% relative gain in explained variance.Predictions shifted by at least 5 percentage points for 24 counties containing approximately 2.5 million residents (13.4% of the border-county population); Canadian embeddings alone performed poorly nationally, so they complemented rather than replaced U.S. embeddings.Source: §2.1 · pp. 5–6, Table 2 · p. 6, Figure 3 · p. 7

2.2.1. Interpolation

  • For spatial interpolation, PDFM-augmented Bayesian models reduced MAE versus baseline (44.95 vs. 48.44), with performance comparable to ACS-augmented models.ACS also reduced MAE (42.07 vs. 48.44; p<0.01); RMSE differences between covariate sets were not significant. PDFM uses one month of data and monthly updates, whereas ACS relies on 60-month pooling and may be released up to 12 months after collection.Source: §2.2.1 · p. 8, Table 3 · p. 9

2.2.2. Nowcasting

  • For cardiovascular-disease nowcasting, combining ACS and PDFM covariates significantly improved both MAE and RMSE over the Bayesian baseline, while PDFM alone produced the lowest GBDT MAE and RMSE among individual covariate configurations.PDFM- and ACS-augmented models did not differ significantly in pairwise comparisons; monthly PDFM embeddings offer a timely alternative or complement to delayed multi-year survey covariates.Source: §2.2.2 · pp. 8–9, Table 4 · p. 9

2.3. Case Study 3: Population dynamics covariates improve forecasting in active vector-borne outbreak hotspots

  • PDFM significantly improved one-month dengue forecasts (ΔWIS = −0.0051), but showed no robust overall improvement at three or six months.Benefits were concentrated in municipalities with active transmission: 71.9%–72.1% of municipalities in the low- and medium-case strata benefited, versus 38.9% in zero-case months. Overall, 1,170 of 2,457 municipalities (47.6%) improved; the benefit was robust across embedding snapshots.Source: §2.3 · pp. 10, 13, Figure 4 · p. 11

2.4. Case Study 4: Population dynamics for hotspot prediction in data-scarce contexts

  • At 4 and 8 weeks, PDFM raised PR-AUC from 0.2832 to 0.3107 and from 0.2733 to 0.2899, respectively, and increased Precision@5 at both horizons.At one week, recent case history performed similarly and Precision@5 was slightly lower with embeddings, a nonsignificant drop; at four weeks, the top five zones captured 2.10 true events per week versus 1.78 using case history alone.Source: §2.4 · p. 14
  • In endemic zones, 4-week PR-AUC rose from 0.4209 to 0.4729, while 8-week Precision@5 increased from 0.3333 to 0.3975.Sporadic zones had a 4-week PR-AUC gain from 0.2127 to 0.2397, but it fell short of significance after correction (p = 0.059).Source: §2.4 · pp. 14–15

2.5. Case Study 5: Scope conditions of geographic context in individual-level clinical risk prediction for postpartum depression

  • PDFM added a small transferable signal to PPD risk prediction: ΔAUC +0.0020, 0.0008–0.0031, in seen states and +0.0038, 0.0004–0.0072, in unseen states.Real embeddings outperformed random embeddings (ΔAUC +0.0047, 0.0013–0.0081), but the gain did not persist in later years (S3: +0.0006, −0.0005 to 0.0015).Source: §2.5 · p. 15
  • PDFM embeddings did not replace individual socioeconomic data: income and Medicaid status outperformed PDFM alone (ΔAUC −0.0116, −0.0144 to −0.0089).PDFM recovered 15% (6–24%) of the gain from individual socioeconomic data and added little beyond it (+0.0011, 0.0000–0.0022; Holm-adjusted p = 0.10).Source: §2.5 · p. 15
  • PDFM shifted rural screening trade-offs by policy, while none of the 21 pre-registered race × ethnicity × urbanicity performance gaps closed.At 20% screening capacity, it found 5,640 more rural cases per year (2,576–8,859) at the cost of 23,474 more false alarms (8,964–39,309); at an 80% sensitivity target, it averted 17,723 false alarms (6,952–32,033) while flagging 1,722 fewer rural cases (650–3,127). The largest baseline disparity, Asian mothers’ under-detection at 80% sensitivity, remained unchanged; all models detected 39.8% of cases.Source: §2.5 · pp. 15–16
  • In exploratory transport scenarios, a 20% screening capacity shifted detection toward rural mothers, scaling to approximately 445,000 more rural and 283,000 fewer urban cases annually worldwide.The estimated net change in total cases found was not distinguishable from zero in any scenario except South America; the scenarios assume US score distributions transfer to other settings, an assumption not testable here.Source: §2.5 · p. 16
  • PPD findings were limited by a time gap between the 2022–2024 embeddings and 2012–2021 births, and the PDFM release omitted Alaska and Hawaii.The reported gain did not vary with gap length (slope 0.0000/yr, −0.0004 to 0.0004); the omissions excluded most Native Hawaiian and Pacific Islander respondents.Source: §2.5 · p. 16

3. Discussion

PDFM is presented as a complementary, reusable place-based signal for public health workflows, with value in cross-border prediction, timely disease estimates, and targeted risk decisions. Its performance and utility depend on the prediction task, available data, and how place-based context is combined with individual information.

3. Discussion

  • Across five independent case studies in four countries, pre-trained place representations supported diverse prediction tasks without task-specific retraining or requiring practitioners to abandon existing tools.The paper frames PDFM as a plug-in layer for existing analytical workflows, not a replacement for established methods.Source: §3 · pp. 16, 18
  • 36% relative increase in explained variance came from adding Canadian contextual covariates to near-border vaccination prediction; PDFM- and ACS-augmented CVD nowcasts showed no statistically significant differences.The discussion presents PDFM as a more timely and easily available alternative for CVD nowcasting.Source: §3 · p. 17
  • PDFM disproportionately corrected large dengue errors in outbreak hotspots and improved cholera forecasts especially in endemic zones, while postpartum screening effects varied by policy.In a verified simulator, fixed screening capacity shifted detection toward rural mothers in rural-majority, high-prevalence settings, while a fixed detection target mainly avoided false alarms; the trade-off is a policy choice.Source: §3 · pp. 17–18
  • Most retrospective analyses used static embedding snapshots, and the forecast gains described for dengue and cholera clustered in an approximately one-to-two-month window.In the DRC cholera analysis, sparse search activity required six-month pooling and PCA, limiting how rapidly embeddings could reflect acute shocks.Source: §3 · pp. 18–19
  • PDFM embedding dimensions lack the clear meanings of labeled census covariates, making them less suited to descriptive or causal analyses of place-based drivers.The paper proposes measuring captured variation in census measures or attributing gains to modality-specific dimension blocks as future work.Source: §3 · p. 19
  • Place-based context should not be assumed to complement individual signals: one cited perinatal study found geospatial context did not rank individuals, while a speech model did.The discussion says fusion pays off only when each component supplies information the other lacks, and that this must be tested empirically.Source: §3 · p. 19

4.1. PDFM Embeddings

PDFM builds location representations from search, built-environment, and environmental data, then uses self-supervised reconstruction to produce embeddings for downstream statistical models.

4.1.1. Data Sources

  • PDFM combines privacy-preserving search trends, POI density and visits, meteorological data, and air-quality measurements as its primary data modalities.Search features use Knowledge Graph entities; built-environment data include 1,192 POI categories and visit counts for 683 categories.Source: §4.1.1 · p. 20
  • Most analyses used embeddings computed from October 2023 data, while postpartum-depression screening used the original PDFM public release.The dengue sensitivity analysis used October 2025 embeddings; cholera search data covered six months through October 2023.Source: §4.1.1 · p. 20

4.1.2. PDFM Architecture and Training

  • The encoder processes each location independently, concatenates its three modalities, and compresses them into a 330-dimensional location embedding.Source: §4.1.2 · p. 20
  • Self-supervised training reconstructs the input modalities from a shared embedding, using dataset-specific heads and a balanced weighted sum of reconstruction losses.The resulting embeddings are used as covariates in downstream statistical models.Source: §4.1.2 · pp. 20–21
  • Cholera embeddings instead used PCA to reduce inputs to 200 dimensions because the DRC training set was an order of magnitude smaller than other countries’ datasets.Source: §4.1.2 · p. 21

4.2. Extrapolating MMR Vaccination Uptake on the US-Canada Border

This section tests whether Canadian context improves prediction of US county MMR vaccination uptake. In permutation tests, improvements in R2, MAE, and RMSE exceeded all 999 random reassignments.

4.2.1. Data

  • The target is county-level MMR vaccination uptake among US children under five, estimated from 2023–2024 parent surveys using multilevel regression with poststratification.Source: §4.2.1 · p. 21

4.2.2. Experimental Setup

  • Canadian FSA embeddings within 150 km of each US county centroid are aggregated using inverse-distance weights and appended to US embeddings as contextual predictors.Connecticut is excluded because it uses planning regions rather than counties.Source: §4.2.2 · p. 21

4.2.3. Models

  • Ridge regression compares US-only, Canadian-only, and combined embeddings, with regularization selected by 10-fold cross-validation.Ridge regression is used to stabilize estimation with high-dimensional correlated predictors.Source: §4.2.3 · p. 22

4.2.4. Evaluation Metrics

  • Performance is assessed with RMSE, MAE, Pearson correlation, and R2.Source: §4.2.4 · p. 23
  • Improvements in R2, MAE, and RMSE were reproduced across 100 repeated cross-validation runs and exceeded all 999 random reassignments (empirical permutation p = 0.001).Source: §4.2.4 · p. 23

4.3. Interpolating and Nowcasting Cardiovascular Disease Risk in US Counties

The section evaluates county-level cardiovascular disease mortality prediction in the contiguous United States using nowcasting and spatial interpolation. Bayesian spatiotemporal and gradient-boosted models are compared with and without ACS and PDFM covariates.

4.3.1. Data

  • CVD mortality counts from 2010–2023 are harmonized to 3,091 county-equivalent units, with counties under 500 residents in any year and several areas lacking usable embeddings excluded.Deaths are defined by underlying-cause ICD-10 codes I00–I99.Source: §4.3.1 · p. 23

4.3.2. Experimental Setup

  • Nowcasting predicts 2023 counts from 2013–2022 data, while interpolation with 2023 data holds out county folds for prediction.The nowcasting set contains 30,910 county-year observations; interpolation uses five folds.Source: §4.3.2 · p. 24

4.3.3. Models

  • Bayesian negative-binomial models and XGBoost gradient-boosted trees compare baseline specifications with ACS covariates, PDFM embeddings, or both.The Bayesian models use expected-count offsets and spatial random effects; nowcasting also includes temporal and county-year effects.Source: §4.3.3 · pp. 24–25

4.3.4. Evaluation Metrics

  • RMSE and MAE assess performance, with paired county errors tested using state-clustered sign-flip permutations and bootstrap intervals.A difference is considered significant only when the permutation test is significant and the bootstrap confidence interval excludes zero.Source: §4.3.4 · p. 25

4.4. Forecasting dengue in Mexico

The section evaluates probabilistic dengue forecasts for Mexican municipalities at 1-, 3-, and 6-month horizons. PDFM embeddings are incorporated into TimesFM forecasts through horizon- and quantile-specific residual models.

4.4.1. Data

  • The analysis covers monthly dengue counts for approximately 2,450 Mexican municipalities from January 2020 through August 2025.Forecast horizons are 1, 3, and 6 months.Source: §4.4.1 · p. 25, §4.4.2 · p. 26

4.4.2. Experimental Setup

  • PDFM embeddings adjust TimesFM forecasts through horizon- and quantile-specific residual models trained in causal 24-month sliding windows.Predicted residuals are added to TimesFM forecasts to produce adjusted predictive distributions.Source: §4.4.2 · p. 26

4.4.3. Models

  • The baseline TimesFM forecasts use historical dengue incidence and dynamic epidemiological and environmental covariates, including climate indicators and population size.Source: §4.4.3 · p. 26

4.4.4. Evaluation Metrics

  • Forecasts are assessed with Weighted Interval Score, where negative ΔWIS indicates improvement after PDFM adjustment; Diebold–Mariano tests and moving-block bootstrap intervals assess differences.The reported bootstrap uses three-month blocks, with six- and twelve-month block sensitivity analyses.Source: §4.4.4 · pp. 26–27, §4.4.2 · p. 26

4.5. Cholera Hotspot Forecasting

The DRC cholera analysis compares history-only features with history augmented by PDFM embeddings in prospective health-zone emergence forecasts. Because sparse digital coverage prevents standard self-supervised training, the embeddings use a custom low-resource construction.

4.5.1. Data

  • PDFM inputs were compressed to 200 dimensions with PCA and used as a static snapshot, because sparse digital coverage ruled out the standard self-supervised reconstruction framework.Search counts were pooled over the six months ending in October 2023; all inputs predated the 1 November 2023 evaluation boundary.Source: §4.5.1 · p. 27
  • History-only models used twelve lagged surveillance, population, and seasonal covariates; the comparison set added PDFM embeddings.Features were shifted to avoid same-week leakage.Source: §4.5.1 · p. 28

4.5.2. Models

  • Prospective evaluation began 1 November 2023, with training truncated by forecast lead time to prevent labels from depending on evaluation-period surveillance.Models were not refitted during evaluation, and evaluation data did not inform feature construction or endemicity assignment.Source: §4.5.2 · p. 28

4.5.3. Evaluation Metrics

  • Emergence meant an outbreak onset within the forecast horizon, defined as at least 100 cumulative suspected cases in a three-week window; already-outbreak zone-weeks were excluded.An eligible zone-week also required at least two weeks of prior and h weeks of subsequent surveillance.Source: §4.5.3 · pp. 28–29
  • The analysis used PR-AUC and weekly Precision@5, with paired inference and Benjamini–Hochberg correction across the specified unstratified and stratified test families.Precision@5 was reported only for strata with at least ten health zones; dormant zones were excluded from stratified analysis because no emergence events occurred during evaluation.Source: §4.5.1 · p. 28, §4.5.3 · p. 29

4.6. Postpartum Depression Screening (PRAMS)

The PRAMS analysis tests whether PDFM location embeddings add predictive signal to individual-level postpartum depressive-symptom screening models. It evaluates prediction, screening, and fairness across respondent, jurisdiction-held-out, and later-birth test settings, while noting that US-only data cannot establish performance elsewhere.

4.6.1. Data

  • PDFM’s 330-dimensional ZCTA embeddings were mapped to 87 state-by-urban/rural locations and population-weighted before being joined one-to-one to respondents.State-level, random, and shuffled embeddings served as comparators.Source: §4.6.1 · p. 30

4.6.2. Models

  • Models compared individual predictors with PDFM replacing state, and tested whether embeddings could replace or add to income and Medicaid predictors.Race and ethnicity were excluded from model inputs by design.Source: §4.6.2 · p. 30, §4.6.1 · p. 30, §4.6.3 · p. 31

4.6.3. Evaluation Metrics

  • Evaluation used a respondent holdout, jurisdiction-held-out folds, and a 2020–2021 birth test set, with threshold-free metrics and screening comparisons at fixed capacity.The primary respondent holdout contained 66,594 people; 20% were flagged at the main screening capacity, with 10% and 30% as sensitivity analyses.Source: §4.6.2 · p. 31, §4.6.3 · p. 31
  • The fairness audit assessed race-by-ethnicity-by-urban/rural performance gaps, counting a gap as closed only if it narrowed without another group losing more than two sensitivity points.Race and ethnicity defined evaluation groups but were never model inputs.Source: §4.6.3 · p. 31
  • The exploratory screening simulator assumes US score distributions carry over to other settings, and US-only PDFM and PRAMS data cannot show that PDFM works elsewhere.Scenarios were defined after the primary results were known.Source: §4.6.3 · p. 32

4.7. Use of AI in Manuscript Preparation

The authors report that portions of the manuscript were drafted manually and then refined using LLMs, with final manual accuracy checks.

4.7. Use of AI in Manuscript Preparation

  • Portions of the manuscript were manually drafted, refined using LLMs, and checked manually for accuracy.The authors take full responsibility for the content.Source: §4.7 · p. 32

Data Availability

The data availability statement identifies sources and access conditions for the case-study datasets. Protected individual-level data were handled under access agreements or national data-sharing frameworks, while some contextual and embedding resources are public.

Data Availability

  • Academic and institutional collaborators independently curated and analyzed the non-Google empirical health datasets other than cholera, without Google accessing protected individual- or patient-level records.The cholera dataset was the exception to that arrangement.Source: Data Availability · p. 32
  • MMR estimates used OutbreaksNearMe participatory-surveillance data, with Canadian spatial context from publicly available Statistics Canada boundary files.The estimates were generated by academic collaborators.Source: Data Availability · pp. 32–33
  • Cardiovascular mortality counts came from CDC/NCHS vital-statistics microdata, while county covariates came from Census ACS estimates.Unsuppressed subnational tabulations require a data-use agreement.Source: Data Availability · p. 33
  • The cholera analysis used DRC national surveillance records covering 403 health zones, with subnational access governed by national data-sharing frameworks.The records were compiled with WHO AFRO collaboration.Source: Data Availability · p. 33
  • PRAMS individual responses require CDC approval and cannot be redistributed by the authors; PDFM v0 embeddings and specified Census and NCHS files are publicly available.Researchers may request PRAMS data through the CDC ARF system or PRAMS proposals process.Source: Data Availability · p. 33

Code Availability

The section describes what is and is not publicly available for reproducing the analyses.

Code Availability

  • PDFM pre-training code is proprietary and not publicly available.Public tutorials, starter code, and benchmark replication examples are available.Source: Code Availability · p. 33
  • Collaborating institutions ran downstream analyses independently using standard off-the-shelf open-source software rather than custom algorithms.The listed tools include Python and R libraries.Source: Code Availability · p. 33
  • Downstream scripts are not publicly released because they depend on institution-specific pipelines and restricted datasets; Methods provide specifications to support replication.The Methods include mathematical specifications, feature definitions, and hyperparameter settings.Source: Code Availability · p. 34

Extended Data

The Extended Data materials report pre-registered postpartum-depression analyses, define stratification and simulation conditions, and describe the scope of supplementary tables. The reported comparisons include no claimed improvement from adding PDFM to income and Medicaid, and no closure of pre-registered intersectional gaps in S1 or S2.

Extended Data

  • Supplementary tables cover border-county MMR prediction changes by threshold, endemicity-stratified performance, unweighted test-set performance, and cohort flow.The supplied captions identify table topics but do not provide their cell values.Source: Table 1 · p. 38, Table 2 · p. 38, Table 3 · p. 39
  • Endemic zones required at least 156 surveillance weeks and reported cases in at least 50% of weeks; 15 zones were endemic and 340 sporadic, while 48 dormant zones were omitted.Table p-values were Benjamini–Hochberg adjusted.Source: Extended Data · p. 38, Table 2 · p. 38
  • Weighted observed/expected ratios in S1 were 0.92–0.93; PDFM denotes the population-weighted v0 embedding for the respondent’s state × urban/rural location.Source: Extended Data · p. 39
  • The pre-registered “PDFM added to income and Medicaid” comparison was not claimed (lower CI bound 0.0000; Holm-adjusted p = 0.104), and none of 21 pre-registered intersectional gaps closed in S1 or S2.A comparison was claimed only when its Holm-adjusted p-value was below 0.05 and its confidence interval supported the pre-defined reading.Source: Table 4 · p. 40, Extended Data · p. 40
  • The label-permutation control gave AUC 0.475 (0.468–0.481); a post-hoc null of 20 permutations centered on 0.499 (0.479–0.522).Results outside primary S1 and all secondary analyses are reported without adjustment.Source: Extended Data · p. 40
  • The exploratory screening simulator reproduces US estimates at US composition and assumes US score distributions within urban/rural × PPD strata carry over to target populations.The simulator rebuilds each target population and reruns the screening rule; the table compares PDFM with baseline in transported settings.Source: Extended Data · p. 41, Table 5 · p. 41
  • PDFM v0 does not cover Alaska or Hawaii.Source: Extended Data · p. 41
Loading 2610.05699v1…