Classifying readmissions to a cardiac intensive care unit

Abstract

Research has associated intensive care unit (ICU) readmissions with increased risk of morbidity and mortality. Readmitted patients are also exposed to complications as they are transferred between hospital units. Moreover, due to their unexpected nature, readmissions increase ICU costs and the complexity of managing ICUs. Existing studies on ICU readmissions have mainly used logistic regression for identifying patients who are more likely to be readmitted. However, such studies do not account for the imbalanced nature of the data where the class of interest (readmitted patients) is the minority group. This paper empirically compares three approaches for handling the imbalanced ICU readmissions data: misclassification cost ratio, synthetic minority oversampling technique (SMOTE), and random under-sampling. We used three classification techniques for identifying patients who are more likely to be readmitted to the ICU within the same hospital stay: support vector machines, C5.0, and logistic regression. We evaluated the classification performance of the three methods using recall, specificity, accuracy, F-measure, G-mean, confusion entropy, and area under the receiver operating characteristic curve. Our results showed that SMOTE is the best approach for addressing the imbalanced nature of the data. The sensitivity analysis identified prolonged ventilation, renal failure, and pneumonia as the top three predictors of ICU readmissions. Our findings can be used to develop a decision support tool to help ICU clinicians and administrators in identifying patients who are more likely to be readmitted and hence provide the patients with the appropriate care to minimize their risk of readmission.