Anchoring Public Health Forecasts: High-Precision Obesity Prevalence Estimation Via Hybrid Geodemographic Ensemble Regression

Obesity remains a major driver of chronic disease and rising healthcare costs in the United States. Yet, most existing predictive models rely on single-model classifiers and rarely incorporate spatial structure or ensemble-based regression, thereby limiting their ability to capture geodemographic variation. This study introduces a hybrid unsupervised–supervised ensemble framework that integrates K-Means spatial profiling with a Stacking Regressor to address these limitations. A geodemographic cluster feature is engineered from latitude, longitude, and population density to capture latent regional health patterns. The supervised ensemble, comprising HistGradientBoosting, Random Forest, and Multi-Layer Perceptron base learners with a RidgeCV meta learner, is trained on CDC BRFSS data cleaned to 19,078 detailed records to forecast obesity prevalence as a continuous regression target. The final ensemble achieves an R² of 0.8045, an accuracy of 91.95%, a correlation (r) of 0.897, and a Mean Absolute Error of 2.20%, outperforming all individual models and related literature. To evaluate robustness, the framework is stress-tested on secondary datasets, where it consistently maintains superior accuracy. These findings demonstrate the value of hybrid spatial ensemble modeling for high-precision public health forecasting.