Determining Personalized Community Health Needs by Feature Selection and Clustering

Abstract

The Center for Disease Control, through the Community Health Data Initiative (CHDI), has released a large dataset by county detailing the overall health indicators, demographics, and major risk factors and causes of morbidity and mortality in the US. In order to address the heterogeneity of community healthcare in the US, k-Means clustering was performed on the CHDI dataset to determine community subtypes in terms of health challenges and outcomes. The optimal number of eight clusters was determined by the Elbow Method, and clusters were analyzed to determine significant differences in demographic. In order to determine community-specific healthcare solutions and directions, feature selection and modeling of healthcare outcomes was performed for each of the eight subtypes using LASSO regression. It was determined that different features significantly impact health outcomes in the different clusters, providing information about the unique health challenges faced by these different types of communities. LASSO regression using the entire unclustered dataset yielded significantly poorer results on the sub-clusters in terms of model performance, further supporting the claim that modeling community-specific needs is a vital step for delivering accurate and adequate community healthcare. These results have the potential to inform policymaking at the local/municipal level, as well as inform the approaches taken by primary practitioners to address community needs.