📄 medicine

Using Explainable Machine Learning to Examine Community, Food-Access, and Built-Environment Factors Associated with Census-Tract Diabetes Prevalence in Collin and Denton Counties, Texas

This study demonstrates that public community, food-access, and built-environment data can explain approximately half of the variation in census-tract diabetes prevalence in Texas counties, though the model's predictive accuracy is significantly lower in low-income areas, highlighting a critical equity gap for public health applications.

Original authors: Eshaan Nidee¹

Published 2026-08-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Eshaan Nidee¹

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Invisible Map of Health

Imagine your neighborhood as a giant, living puzzle. For a long time, doctors have looked at the individual pieces—the person's age, their diet, and their family history—to understand why some people get sick while others stay healthy. But there's a whole other layer to the puzzle: the picture the pieces make when they fit together. This is the world of social determinants of health. Think of it as the "weather" of a community. Just as a stormy sky makes it harder to grow a garden, certain neighborhood conditions—like how easy it is to walk to a grocery store, how many parks are nearby, or how much money the average family makes—can make it harder for people to stay healthy, regardless of their personal choices.

Scientists have been trying to build a "weather forecast" for disease using computers. They use machine learning, which is like teaching a computer to find hidden patterns in a mountain of data, much faster than a human could. The big question they are asking is: Can we look at a neighborhood's map, its food stores, and its schools, and predict where diabetes (a serious condition where the body struggles to manage sugar) will be most common? If we can, we might be able to fix the "weather" before the storm hits, rather than just treating the people who get soaked.

The Neighborhood Detective: A Story of Maps, Models, and Mistakes

In the bustling suburbs of Collin and Denton Counties in Texas, a high school researcher named Eshaan Nidee decided to play detective. The case? Figuring out why diabetes rates vary so wildly from one neighborhood (called a "census tract") to the next, even when they are right next door to each other.

Eshaan didn't just look at medical records; instead, they built a digital detective kit using only public information. They gathered data on community life (like how many people have college degrees or how much money families make), food access (using a special map from the USDA to see where "food deserts" might be), and the built environment (counting how many grocery stores, clinics, and parks exist per square mile, using a giant open-source map called OpenStreetMap).

They fed this information into four different computer "brains" (algorithms) to see which one could best guess the diabetes rate for each of the 412 neighborhoods. The goal was to see if these "actionable" community factors alone could predict the health of a neighborhood, without incorporating other health problems first.

The Two Models: The "Real" Test vs. The "Easy" Baseline

To make sure the computer wasn't just getting lucky, Eshaan set up a two-part challenge.

Model A was the "Real" test. It only used the community, food, and park data. It was like trying to guess a student's test score based only on their neighborhood's library and school funding, without knowing if they studied or slept well.

Model B was the "Easy" baseline. It took Model A's data and added three other health stats: obesity, physical inactivity, and high blood pressure (hypertension). This was like being allowed to peek at the student's previous test scores to guess their next one.

The Results:
When Model A (the community-only detective) tried to predict diabetes, it got about half of the story right. Specifically, it explained 51.9% of the differences between neighborhoods (a number called $R^2$ ). When the researchers tested it on a completely different county it had never seen before, it actually did slightly better, explaining 59.2% of the differences. This suggests the computer learned some real rules about how neighborhoods work, not just memorized the specific towns it studied.

However, when Model B (the baseline) was used, the accuracy skyrocketed to 93.5%. It seemed like a miracle! But here is the twist: the computer wasn't using the parks or the grocery stores to get that high score. It was almost entirely relying on hypertension (high blood pressure). In fact, high blood pressure and diabetes are so closely linked in this data (a correlation of 0.87) that the computer basically just said, "If they have high blood pressure, they probably have diabetes."

The paper argues that Model B is a "ceiling" of predictability, not a discovery. It's like a weather app that predicts rain perfectly because it's looking at the rain that is already falling, rather than predicting the storm before it starts. The real takeaway is that while community factors matter, they only explain about half the picture on their own.

The "Unfair" Map: When the Detective Gets It Wrong

Here is where the story gets a bit serious. The researchers checked if their "Real" model (Model A) was fair to everyone. They split the neighborhoods into low, middle, and high-income groups.

The result? The computer was much worse at predicting diabetes in low-income neighborhoods.

In high-income areas, the average error was only 0.63 percentage points.
In low-income areas, the error jumped to 1.29 percentage points—roughly twice as bad.

This is a crucial finding. It means that if a city planner used this computer model to decide which neighborhoods needed help the most, they might accidentally miss the poorest areas because the model is less reliable there. The paper suggests this happens because the "rules" of the neighborhood might be different for poor communities, or because the data doesn't capture the full story of those areas. It's a warning that even smart computers can have blind spots, especially for the people who need help the most.

The Final Clue: The Neighborhood Effect

Finally, the researchers looked at the "mistakes" the computer made. They found that when the computer guessed wrong, it tended to make the same kind of mistake in neighboring towns. If it overestimated diabetes in one town, it likely overestimated it in the town next door. This suggests that there are invisible forces—like shared bus lines, hospital boundaries, or local history—that affect whole regions, not just single blocks.

The Bottom Line

This paper teaches us that we can use public maps and data to predict where diabetes is likely to be a problem, but we can only see about half of the picture using community factors alone. The other half is hidden in the complex web of other health issues like high blood pressure.

Most importantly, the study warns us that these computer models are not perfect. They work best for wealthy neighborhoods and struggle with poorer ones. So, while we can use these tools to start a conversation about how to build healthier cities, we have to be careful not to trust them blindly. We need to remember that the map is not the territory, and the computer's guess is just a starting point, not the final answer.

Technical Summary: Explainable Machine Learning for Census-Tract Diabetes Prevalence in Texas

Problem Statement
Diabetes prevalence in the United States varies significantly between neighboring communities, a variation largely driven by social determinants of health (SDOH) rather than individual clinical risk factors alone. While predictive modeling in this domain often relies on individual-level clinical data or combines community-level predictors with downstream comorbidities (e.g., obesity, hypertension), there is a lack of research that isolates the predictive power of actionable community characteristics (socioeconomic status, food access, built environment) from comorbidities that are themselves strong correlates of diabetes. Furthermore, existing community-level models rarely assess whether predictive accuracy is distributed equitably across income groups or whether models generalize geographically beyond their training data.

Methodology
This study employed a cross-sectional ecological machine learning analysis across 412 census tracts in Collin and Denton Counties, Texas. The research utilized a framework designed to separate actionable community predictors from health comorbidities:

Data Sources: The study integrated data from the American Community Survey (ACS) for socioeconomic indicators, the USDA Food Access Research Atlas for food-access measures, OpenStreetMap (OSM) for built-environment densities, and CDC PLACES for diabetes prevalence and comorbidity estimates.
Feature Sets: Two distinct feature sets were constructed:
- Model A (Primary/Actionable): Comprising 12 variables: five ACS socioeconomic/demographic variables, two USDA food-access variables, and five OSM built-environment variables.
- Model B (Comorbidity Comparison): Model A plus three comorbidity variables (obesity, physical inactivity, and hypertension prevalence).
Algorithms: Four algorithms were evaluated: a mean-prediction baseline, Ordinary Least Squares (OLS) regression, LASSO regression, and Random Forest regression.
Validation Strategy: To ensure robustness and generalizability, models underwent four validation procedures:
1. 80/20 held-out split.
2. 5-fold out-of-fold cross-validation (with bootstrap confidence intervals).
3. Nested 5-fold cross-validation with hyperparameter tuning for Random Forest.
4. Geographic external validation: Training exclusively on Collin County (219 tracts) and testing on Denton County (193 tracts).
Interpretability and Fairness:
- Explainability: SHAP (SHapley Additive exPlanations) values were used to quantify the marginal contribution of predictors, specifically to determine how much Model B's accuracy was driven by comorbidities versus community factors. OLS coefficients provided linear interpretations for Model A.
- Fairness Auditing: Prediction error (Mean Absolute Error) was stratified by income terciles to assess equity.
- Spatial Analysis: Residual spatial structure was analyzed using Moran scatterplots to detect clustering of prediction errors.

Key Results

Predictive Performance of Actionable Factors (Model A): Public community, food-access, and built-environment data explained a modest but consistent share of tract-level diabetes variation.
- Cross-validated $R^2$ was 0.519 (95% CI: 0.402–0.624).
- Nested cross-validation yielded an $R^2$ of 0.523.
- Geographic external validation (Collin to Denton) resulted in an $R^2$ of 0.592, indicating the model generalizes well across distinct demographic and food-retail landscapes within the region.
Impact of Comorbidities (Model B): Adding obesity, physical inactivity, and hypertension increased the out-of-fold $R^2$ to 0.935. However, SHAP analysis revealed that hypertension (correlated with diabetes at $r=0.87$ ) dominated the predictive signal. The author concludes that Model B represents a "comorbidity ceiling" rather than evidence that community factors alone explain 93.5% of the variance.
Equity and Error Distribution: Predictive accuracy was not equity-neutral. Model A's error was significantly higher in low-income tracts (MAE = 1.29) compared to high-income tracts (MAE = 0.63), with a gap of 0.66 points ( $p=0.0001$ ).
Spatial Structure: Residuals exhibited positive spatial autocorrelation, suggesting that regional factors (e.g., healthcare system boundaries, transportation networks) not captured by tract-level predictors influence diabetes prevalence.
Variable Importance: Among Model A predictors, educational attainment showed the strongest univariate association with diabetes prevalence. Poverty and unemployment rates showed weak univariate associations in this specific context.

Significance and Claims
The paper claims to be among the first census-tract-level studies to explicitly separate actionable community predictors from downstream comorbidities while integrating geographic external validation, explainability, and fairness auditing within a single framework.

Actionable Insight: The study demonstrates that public data alone can predict roughly half of the variation in diabetes prevalence, providing a baseline for community-level intervention without relying on protected health information or comorbidity data.
Cautionary Finding on Accuracy: The author emphasizes that high accuracy in community models is often an artifact of including closely correlated comorbidities (like hypertension). They argue that models relying on these variables do not identify new, modifiable targets but rather exploit existing risk clustering.
Equity Warning: A central contribution is the demonstration that prediction error is systematically higher in low-income communities. The author asserts that any "Community Priority Score" derived from this model must be tempered by the knowledge that it is least reliable in the very populations it is most likely to flag for intervention.
Generalizability: The successful transfer of the model from Collin to Denton County suggests that the relationships captured are robust to shifts in educational composition and food-retail geography, though the study remains limited to a specific North Texas metropolitan region.

The study concludes that while public data offers a moderate predictive signal for diabetes prevalence, the resulting tools must be applied with an awareness of their limitations regarding equity and the distinction between actionable community factors and downstream health outcomes.