After studying this material, students should be able to:

1. Explain what planetary geospatial foundation models are and why they matter for public health
2. Describe how PDFM works and what data it uses
3. Identify the core limitations of traditional epidemiological surveillance
4. Analyze real-world case studies where PDFM improved health outcomes
5. Evaluate when and why geospatial embeddings outperform conventional approaches

---

Step-by-Step Teaching Guide

STEP 1: Understanding the Problem First

Why does public health data fail us?

Before appreciating the solution, you must understand what is broken.

Traditional epidemiological surveillance suffers from three critical failures:

Think of it this way: Imagine trying to navigate traffic using a map that was printed two years ago. That is essentially what public health officials do with conventional data.

Additionally, building custom models for each disease requires:

• Extensive task-specific data collection
• Custom engineering pipelines
• Significant time — which is unavailable during rapid outbreaks

---

STEP 2: Introducing the Core Concept — Foundation Models

What is a foundation model?

A foundation model is a large AI model trained on broad data that can be applied across many different tasks without retraining from scratch.

You likely already know examples:

• GPT → trained on text, applied to writing, coding, summarizing
• PDFM → trained on geospatial signals, applied to health, disease, demographics

The key insight: Instead of building a new model for every health problem, you build one powerful model of place and plug it into existing workflows.

---

STEP 3: Understanding PDFM Specifically

What is the Population Dynamics Foundation Model (PDFM)?

PDFM is Google Earth AI's geospatial foundation model. It creates location embeddings — think of these as compact digital fingerprints for any geographic location.

What data goes into these fingerprints?

PDFM Inputs → Location Embedding (Fingerprint)
─────────────────────────────────────────────
• Anonymous search trends
• Human mobility patterns
• Built-environment density
• Environmental determinants

Three critical properties of PDFM embeddings:

1. Privacy-preserving — No individual-level data is used
2. Monthly refresh cadence — Far fresher than census data
3. Task-agnostic — Same embedding works across completely different health problems

Analogy: A location embedding is like a neighborhood's personality profile. It captures whether an area is urban or rural, wealthy or underserved, well-connected or isolated — all without identifying any individual person.

---

STEP 4: The Paradigm Shift — Plug-and-Play vs. Build From Scratch

This is the conceptual core of the article.

Old paradigm:

Health Problem → Collect Data → Build Custom Pipeline → Train Model → Deploy
                 (months/years)   (expensive)

New paradigm:

Health Problem → Plug PDFM Embeddings Into Existing Model → Deploy
                 (days/weeks)

PDFM does not replace epidemiologists or their models. It enhances them by providing ready-made geospatial context as an input variable.

---

STEP 5: Case Study Analysis — Learning Through Evidence

Now apply your understanding to five real-world cases. For each, ask: What problem existed? What did PDFM add? What was the measurable result?

---

Case Study 1: Cross-Border Vaccination (U.S.-Canada)

Problem:

• 146 U.S. counties sit within 150 km of the Canadian border
• People cross borders for work, shopping, healthcare
• U.S.-only models completely ignore this behavioral reality

What PDFM did:

• Added Canadian location embeddings alongside U.S. county embeddings
• Captured cross-border mobility and behavioral spillovers

Result:

• Explained variation in MMR vaccination coverage increased from 16% → 22% (a 36% improvement)
• Coverage estimates refined by ≥3 percentage points for 4.7 million border residents

Key lesson: Disease outcomes do not respect political borders. Models should not either.

---

Case Study 2: Cardiovascular Disease Nowcasting

Problem:

• CVD kills 916,000 Americans annually
• Official mortality data lags 1-2 years
• Census covariates reflect conditions 2-3 years in the past
• Health departments are making decisions based on severely outdated information

What PDFM did:

• Substituted for census-based inputs in estimating current-year CVD deaths across ~3,100 counties

Result:

• No statistically significant difference in accuracy compared to traditional census data
• But PDFM is fresher and available in more locations

Key lesson: PDFM can match the accuracy of slow, expensive census data while being dramatically more current and accessible.

---

Case Study 3: Dengue Forecasting in Mexico

Problem:

• Dengue is a vector-borne disease requiring urgent, time-sensitive response
• Forecasting across ~2,450 Mexican municipalities is complex
• Short-term forecasts (1 month ahead) are most operationally valuable

What PDFM did:

• Combined with TimesFM 2.0 (a time-series foundation model)
• Forecasted dengue case counts from 2020-2025

Result:

• Improved forecast accuracy in 72% of active transmission municipalities
• Error reductions 3.4× larger than degradations
• Greatest gains at the 1-month horizon — exactly when vector control decisions must be made

Key lesson: Combining geospatial and temporal foundation models creates compounding benefits for outbreak forecasting.

---

Case Study 4: Postpartum Depression Screening

Problem:

• PPD screening relies on clinical intake data
• Community-level risk factors (poverty, isolation, access to care) are rarely captured
• Health systems must prioritize limited follow-up resources efficiently

What PDFM did:

• Evaluated across 332,970 participants in the CDC PRAMS national survey
• Added community-level context to individual-level screening models

Results:

Critical finding: The model worked in states it had never seen during training — demonstrating genuine geographic generalizability.

Key lesson: Geospatial context can partially substitute for missing individual-level data, and it transfers to new regions.

---

Case Study 5: Cholera Early Warning in DRC

Problem:

• Cholera outbreaks are rare — fewer than 1 in 100 health zones affected in any given week
• Rarity makes prediction extremely difficult
• Early warning is critical for prepositioning vaccines and clean water
• The DRC has sparse internet connectivity, limiting data availability

What PDFM did:

• Used a lightweight, low-resource version of PDFM adapted for low-connectivity regions
• Tested across 403 health zones

Results by time horizon:

1-2 weeks ahead:  Recent case counts dominate → PDFM adds little
4-8 weeks ahead:  PDFM significantly improves hotspot forecasting
                  → Enough lead time to move supplies

Key lesson: Foundation models capture underlying structural determinants (environment, connectivity, population density) that become increasingly valuable as you look further into the future.

---

STEP 6: Synthesizing the Pattern Across All Cases

Now step back and identify what all five cases share:

---

STEP 7: Understanding Current Limitations and Future Directions

Honest assessment of what PDFM cannot yet do:

1. Static snapshots — Current embeddings do not fully capture rapid temporal change within a month
2. Under-connected regions — Sparse internet means sparse signal; the DRC required a special lightweight version
3. Short-term outbreak tracking — As seen in cholera, very recent clinical data still outperforms PDFM at 1-2 week horizons

Active research directions:

• Temporally dynamic embeddings (capturing change over time more granularly)
• Geographic transfer learning for under-connected regions

---

STEP 8: Conceptual Summary

TRADITIONAL APPROACH                    PDFM APPROACH
─────────────────────────────────────────────────────────
Slow census data          →             Monthly refreshed embeddings
Geopolitical boundaries   →             Cross-border behavioral signals
Task-specific pipelines   →             Plug-and-play inputs
Reactive (lagged)         →             Proactive (current/forecasting)
Data-rich regions only    →             Adaptable to low-resource settings

---

Self-Assessment Questions

Test your understanding:

1. What are the three main limitations of conventional epidemiological surveillance?

2. In your own words, what is a location embedding and why is it useful?

3. Why did PDFM improve CVD nowcasting even though it matched — rather than beat — census data accuracy?

4. Why did PDFM help more at the 4-8 week cholera forecast horizon than at 1-2 weeks?

5. What does it mean that the PPD model worked in "unseen states," and why does that matter practically?

6. What is the fundamental paradigm shift this article proposes for public health modeling?

---

Bottom line: PDFM represents a shift from building custom, slow, data-hungry models for each health problem toward a world where a single pre-trained model of place can be plugged into any health workflow — making public health responses faster, broader, and more equitable across the globe.