Understanding the effect of specialization on hospital performance through knowledge-guided machine learning

Abstract

The healthcare spending in the United States accounts for about 17% of US GDP, making the US the highest healthcare expenditure per capita in the world. There are areas for improvement such as hospital care, which represents the single largest national health expenditure by the type of services. Because hospital operations and manufacturing processes share some similarities at systems level, we believe that some of the theories and techniques developed for manufacturing processes are also applicable to hospital operations. In this work, we examine whether the so-called focused factory theory (i.e., factories that concentrate on narrow range of services or operations produce better products at low costs) is applicable to hospital operations. Specifically, we examine whether the hospitals that are specialized in certain diseases achieve better results in terms of costs and patient outcomes using a large national healthcare cost and utilization project (HCUP) dataset. Challenges in analyzing the HCUP data will be discussed. Pure data-driven machine learning (ML) approaches based on multiple linear regression (MLR), principal component regression (PCR), partial least squares (PLS), Fisher discriminant analysis (FDA), ordinal regression (OR) and logistic regression (LR) will be used to investigate the effects of hospital specialization on hospital performance in terms of cost (measured by total charge) and patient outcome (measured by death of patient during hospitalization). We show that, without domain knowledge, the results from ML approaches are incomplete. The full effects of specialization are revealed only when ML is applied to a model structure that is defined based on domain knowledge. Our results suggest that domain knowledge can play a significant role in machine learning applications and should be incorporated whenever possible.