Identifying influential individuals and predicting future demand of chronic kidney disease patients

Abstract

To ensure high service quality, managers need to personalize treatment options and meet their customer demands. Our research is motivated by the need to better anticipate and prepare for that. We develop a generalizable framework that is the first to address two healthcare risk management goals: (1) identifying high risk and stable-demand customers and (2) predicting the medium-term demand for services of stable-demand customers. We also design a model-agnostic method for variable evaluation. It can rank predictors based on their global impact, and highlight their effect on a model's local accuracy. In this research, we leverage a large electronic medical records' data set, which comprised of 48,344 chronic kidney disease patients treated across geographically diverse Veterans Affairs regions. Our framework indicates that although only 1.3% of the examined individuals are high-risk patients, it can correctly identify 35% of them and highlight an additional 8.9% as having important demand implications. Identifying high-risk individuals can be used in (1) monitoring prioritization, (2) patients' motivation, and (3) patients' stabilization. Furthermore, our model accurately predicts the monthly need for care of stable-demand individuals up to 3 years into the future and outperforms popular statistical and data mining models. This information is especially critical for hospital management in identifying future hiring needs.