Abstract
Despite industrialized nations' increases in medical investment, care organizations have been unable to consistently deliver quality healthcare to all patients. Healthcare decision makers must find ways to improve delivery during a patient's stay at a hospital while also assuring they are discharged at the proper time. Oftentimes, decision makers rely on care variation studies to understand trends in the consistency of patient care across many patient cohorts. Decision makers also measure the quality of transition from discharge to home through unplanned thirty-day readmissions. Unplanned thirty-day readmissions has become the legislated metric for quality of care in the United States. This dissertation provides metrics to measure and assess care organizations' variation in care both between and within patient cohorts. This dissertation also provides predictive models and modeling recommendations across a broad set of data, variable, and model choices to improve unplanned thirty-day readmission prediction. These two combined provide a set of care quality tools for addressing consistency in care during a visit and data driven decision tools for deciding when that visit should end.
This research provides six multidisciplinary contributions. First, this dissertation provides a data framework for simplifying and utilizing a complicated, high-dimensional data structure consisting of many categorical variables. Second, this dissertation demonstrates a statistically viable method for measuring variation between two columns of the data framework. Third, this research demonstrates metrics for measuring variation within columns of the framework and provides validation for their utility. Fourth, this dissertation assesses the performance and parameterization of three algorithms on high-dimensional data sets. Fifth, for high-dimensional data, this research assesses the utility of methods commonly used to address class imbalance. Finally, this research provides evidence that careful selection of variable representation when deriving new variables for predictive models has compounding effects on model performance.
This dissertation also provides six healthcare contributions. First, this research demonstrates the utility of our between cohort variation method across both principal and all procedures for 2,383 comorbidities as well as additional cases. Second, this research provides evidence for increased chance of variation due to lack of diagnostic specificity. Third, this research derives more than a dozen new variable representations, including within cohort variation metrics, and presents their predictive performance of unplanned thirty-day readmissions. Fourth, this research has demonstrated a method for ranking procedures for analysts to explore in order to reduce thirty-day readmissions and care variation. Fifth, this dissertation has developed the best performing model of thirty-day readmissions to date. Finally, this dissertation provides multiple recommendations for shifting the thirty-day readmissions modeling paradigm to markedly improve predictive performance.
