A Regionally Generalized Machine Learning Framework Towards Census-Enabled Multi-Factor Non-Communicable Disease Analyses

In recent years, non-communicable diseases, diseases intransmittable through human interaction, have cemented themselves as a leading cause of death globally. These diseases are exacerbated by various social, economic and geographical factors within communities. Despite their resounding consequences on community health, the compilation of medical surveys is both costly and time consuming. This study uses a machine learning approach to predict the population proportion with Asthma, High Blood Pressure, Diabetes, and Poor Mental Health within census tract populations in three states. Through an evaluation of 8 different regression models on their predictive capabilities of disease prevalence, the Cubist model held the strongest results for every disease, achieving considerably low relative mean value errors of 3.849%, 6.077%, 7.656%, 7.866%, and 5.531% towards each of the non-communicable diseases, respectively. Out of the four categories split from 41 input features- Sociodemographic, Economic, Environment/Geographic, and Environmental Justice-Sociodemographic features consistently were the highest contributors. A lightweight model with comparable accuracy to the full feature set (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\leq 0.1\%$</tex> difference) was found to require only the 20 most important features from every Cubist model. This study provides insight towards machine learning models' capabilities in enhancing the United States' understanding of domestic epidemiology determinants.