🩺 endocrinology

Predicting county-level diagnosed diabetes prevalence in the United States using explainable gradient boosting and geographic interpretation

This study utilizes an explainable LightGBM framework integrated with diverse public data sources to accurately predict and geographically interpret county-level diagnosed diabetes prevalence across the United States, identifying poverty and food insecurity as dominant structural drivers while emphasizing that the resulting insights serve as model-based explanations rather than causal effects.

Original authors: Yahaya, Y., Khan, S., Rani Saha, P., Meia, M. A. A.

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yahaya, Y., Khan, S., Rani Saha, P., Meia, M. A. A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: Mapping the "Diabetes Landscape"

Imagine the United States as a giant patchwork quilt made of 3,000 different counties. Some patches are bright and healthy, while others are dark and struggling. This study wanted to figure out why some patches (counties) have much higher rates of diagnosed diabetes than others.

Instead of looking at individual people (like checking a single person's blood sugar), the researchers looked at the entire neighborhood of each county. They asked: "If we look at the whole town, what makes it more likely that people there have diabetes?"

The Tool: A Super-Intelligent Weather Forecaster

To solve this puzzle, the researchers built a computer model. Think of this model as a super-advanced weather forecaster.

The Goal: Just as a weather forecaster uses temperature, wind, and humidity to predict rain, this model used data about poverty, food, jobs, and demographics to "predict" the diabetes rate in a county.
The Contest: They didn't just use one forecaster. They held a competition between four different types of computer algorithms (Elastic Net, Random Forest, XGBoost, and LightGBM).
The Winner: LightGBM won the first round because it was the most accurate at guessing the rates on a "practice test" (validation set). However, XGBoost actually did slightly better on the final "real exam" (the held-out test set). The researchers kept LightGBM as the main star because that's how they set the rules, but they kept XGBoost as a backup to make sure the results were solid.

The Ingredients: What Goes Into the Prediction?

To make their forecast, the model ate a massive buffet of public data. They mixed together:

Food Environment: How many fast-food restaurants vs. grocery stores are there? How many people live in "food deserts" (areas far from fresh food)?
Money & Jobs: What is the poverty rate? How many people are unemployed? How much money does the average family make?
Demographics: How old is the population? What is the racial makeup?
Health Habits: How many people smoke, are obese, or don't exercise?

The Surprise: The researchers found that even if you took away the specific health habits (like smoking or obesity) and only looked at the "structural" stuff (poverty, food access, jobs), the model could still guess the diabetes rates pretty well. It's like saying, "You don't need to know the exact recipe of the cake to know the kitchen is messy; the mess itself tells you a lot."

The "Why": The Magic of SHAP Maps

The hardest part of using AI is that it's often a "black box"—you get an answer, but you don't know why. To fix this, the researchers used a tool called SHAP (which sounds like "shap" but stands for a complex math concept).

Think of SHAP as a detective's magnifying glass. It breaks down the final prediction for every single county and says: "Here is exactly how much poverty added to the diabetes rate, and here is how much food insecurity added to it."

They created a map where every county is colored by its "main culprit."

In many counties in the South, Poverty was the biggest driver.
In others, Food Insecurity (not having enough money for food) was the main driver.
In some places, Unemployment or SNAP participation (food stamps) was the key factor.

What the Map Tells Us (and What It Doesn't)

The researchers drew a very important line in the sand: This map shows patterns, not causes.

The Analogy: Imagine you see a map where every time it rains, people carry umbrellas. The map shows a strong link between rain and umbrellas. But the map doesn't prove that carrying an umbrella causes rain.
The Paper's Warning: The researchers emphasize that their map tells us what is associated with high diabetes rates in a specific town, but it does not prove that fixing poverty will automatically fix diabetes. It is a tool for hypothesis generation—it tells public health officials, "Hey, look at this county; something about its poverty or food access is strongly linked to high diabetes rates. Go investigate further."

The Results in a Nutshell

Accuracy: The model was incredibly good at predicting diabetes rates across the US (about 96% accurate in terms of statistical fit).
Geography Matters: The model worked great when testing random counties, but struggled a bit when trying to predict a whole region it hadn't seen before (like the Northeast). This suggests that every region has its own unique "flavor" of risk factors.
The Main Drivers: Nationally, Poverty was the most frequent "main culprit" for high diabetes rates, followed closely by Food Insecurity.
Spatial Clustering: Diabetes rates aren't random; they cluster together. High rates tend to be surrounded by other high rates (mostly in the South), and low rates cluster together (mostly in the Midwest and West).

The Bottom Line

This study is like creating a high-definition heat map of diabetes risk. It uses powerful math to show us that where you live, how much money you make, and what food is available in your neighborhood are huge factors in diabetes rates.

However, the authors are very careful to say: This is a starting point, not a solution. The map helps us ask the right questions and find the right places to look deeper, but it doesn't tell us exactly how to fix the problem or what specific action will work best. It's a tool for understanding the landscape, not a magic wand for changing it.

Technical Summary: Predicting County-Level Diagnosed Diabetes Prevalence in the United States Using Explainable Gradient Boosting and Geographic Interpretation

Problem Statement
Diagnosed diabetes affects approximately 38.4 million Americans, with a burden that is geographically uneven across U.S. counties. While existing machine learning studies have successfully predicted individual-level diabetes risk using biometric and clinical variables, they are less suited to explaining why prevalence differs geographically at the county level. Furthermore, traditional spatial regression models often struggle to capture nonlinear relationships while simultaneously producing locally interpretable outputs. There is a need for a framework that utilizes public county-level data to predict diabetes prevalence and provide explainable, geographically specific insights into the structural and contextual conditions driving these disparities.

Methodology
The study employed an ecological cross-sectional design analyzing 2,957 U.S. counties. The unit of analysis was the county, chosen to capture variations in food access, socioeconomic disadvantage, and local health behaviors that exist within states.

Data Sources: The analysis integrated data from five public sources linked via Federal Information Processing Standards (FIPS) codes:
- Outcome: 2023 age-adjusted diagnosed diabetes prevalence (CDC PLACES, based on BRFSS).
- Predictors: Food access and environment (USDA Food Access Research Atlas, USDA Food Environment Atlas), socioeconomic and demographic indicators (U.S. Census ACS 5-year estimates), and rurality (USDA Rural-Urban Continuum Codes).
- Covariates: Health behaviors and clinical conditions (obesity, physical inactivity, hypertension, smoking) were also included from CDC PLACES.
Feature Engineering: A Food Desert Index (FDI) was calculated by aggregating tract-level data to the county level using a population-weighted formula. A Food Environment Composite Index (FECI) was constructed via Principal Component Analysis (PCA) combining fast-food density, grocery density, and SNAP participation. Missing data were handled via iterative multivariate imputation.
Modeling Approach: Four regression models were trained and compared: Elastic Net, Random Forest, XGBoost, and LightGBM.
- Selection: Models were tuned via cross-validated grid search. The primary model was selected exclusively based on Root Mean Squared Error (RMSE) on a validation set (15% of data).
- Validation: A held-out test set (15%) was used for final evaluation. Region-held-out validation was also performed to assess geographic generalizability.
- Interpretability: The selected model was interpreted using SHapley Additive exPlanations (SHAP) with the TreeExplainer algorithm. SHAP values were used to rank global feature importance and identify the "dominant positive structural contributor" for each county.
Sensitivity and Diagnostics: Nine sensitivity analyses tested robustness against alternative predictor sets (e.g., excluding health behaviors, removing the FECI, reducing collinearity). Spatial autocorrelation was assessed using Global Moran's I on both observed outcomes and model residuals.

Key Results

Model Performance: LightGBM was selected as the primary model based on the lowest validation-set RMSE (0.442). On the held-out test set, it achieved an RMSE of 0.423 percentage points, an $R^2$ of 0.964, and a MAPE of 2.76%. XGBoost performed slightly better on the test set ( $R^2$ = 0.968, RMSE = 0.399) but was retained only as a secondary benchmark.
Feature Importance: Global SHAP analysis identified physical inactivity as the most influential predictor overall. Among structural and contextual predictors, poverty rate, food insecurity rate, SNAP participation rate, unemployment rate, and median household income were the most significant contributors. The food desert index and retail density indicators had lower global importance after accounting for broader socioeconomic factors.
Geographic Interpretation:
- Dominant Contributors: Nationally, poverty rate was the most frequent dominant positive structural contributor (26.1% of counties), followed by food insecurity (23.9%).
- Regional Variation: Patterns varied by Census region. Poverty and food insecurity were dominant in the South and Midwest; unemployment was most common in the Northeast; and SNAP participation was prominent in the West.
- Spatial Autocorrelation: The observed outcome showed strong spatial clustering (Global Moran's $I$ = 0.665). After model fitting, residual spatial autocorrelation decreased significantly to $I$ = 0.069, though it remained statistically significant, indicating some unexplained geographically structured variation.
Robustness: A sensitivity model using only structural and contextual predictors (excluding health-behavior covariates) retained substantial predictive power ( $R^2$ = 0.827), suggesting these factors contain considerable independent predictive information. Region-held-out validation revealed weaker transportability, particularly in the Northeast ( $R^2$ = 0.336), highlighting that random-split accuracy does not guarantee geographic generalizability.

Significance and Claims
The paper claims to demonstrate that explainable machine learning, when applied to public county-level data, can effectively characterize geographic variation in diagnosed diabetes prevalence. The primary significance lies in the ability to translate complex model predictions into geographically specific explanations using SHAP.

The authors emphasize that these SHAP-derived maps and dominant-contributor analyses serve as tools for local hypothesis generation and identifying counties and structural conditions that warrant further investigation. However, the paper explicitly states that these outputs describe predictive patterns and model explanations, not causal effects. The authors caution against interpreting SHAP values as causal determinants or as direct recommendations for intervention targets. Instead, the framework is presented as a reproducible approach to support geographically informed diabetes-prevention planning, provided it is combined with local surveillance data, stakeholder knowledge, and causal research.