Predicting Food Insecurity Across U.S. Census Tracts: A Machine Learning Analysis Using the USDA Food Access Research Atlas

Food insecurity still poses a serious public-health and social-equity problem in the United States. The USDA Food Access Research Atlas (N = 72,531 census tracts) served as the basis for this study, which not only created but also evaluated machine-learning models to predict the level of food insecurity in a certain tract, which is determined by the lack of supermarkets being accessible to low-income people. The Logistic Regression, Random Forest, and XGBoost classifiers went through training and standard metric comparison. The tree ensemble models (Random Forest and XGBoost) reached remarkable performance (accuracy ≈ 97%, ROC-AUC ≈ 0.99) far above the logistic regression baseline (ROC-AUC ≈ 0.89). SHAP-based model interpretability recognized the poverty rate, median family income, SNAP participation, and vehicle access as the most critical determinants of food insecurity. These results affirm the value of interpretable machine learning in revealing important socioeconomic factors that contribute to food access inequalities, thereby providing a basis for data-informed interventions. The entire analytic process made use of publicly accessible national data, so the results can be reproduced, and future research and policy applications made more transparent.