Data mining based predictive models for 30-day hospital readmission

It has been observed that a significantly large number of readmissions of congestive heart failure (CHF) patients are preventable. Due to high penalty cost from the Centers for Medicare and Medicare Services (CMS) readmission reduction program, many hospitals are striving to reduce their readmission rates to a level below the national average. Accurate identification of high-risk readmitted patients is a prioritized task, which can be facilitated by predictive analytics techniques. The goal of this research was focused on developing a predictive framework of the CHF readmissions. Data from 2010-2013 from a community hospital in Upstate New York containing 1,167 patients characterized by demographic, medical, and laboratory features was used, in which there are apparently challenging issues: data overlapping and data imbalance, which were not taken into account previously in most literature. In the proposed framework, logistic regression and support vector machine were employed integrating feature selection and sampling techniques as preprocessing steps to resolve the abovementioned issues. The results showed that of the 22 features, a logistic regression model with seven significant features outperforms and provide sensitivity of 70.7%, specificity of 69%, and accuracy of 70.2% on the testing dataset. Moreover, the developed algorithms can be easily generalized for other critical diagnostic related groups, such as Acute Myocardial Infarction and Pneumonia.