Dimensionality Reduction in Predicting Hospital Readmissions of Diabetics

Hospital readmissions, being expensive, create a negative impact on the quality of care provided by the hospital. If the patients facing high risk of readmissions are identified early and taken an intensive care of, the rate of hospital readmissions can be decreased. Thus reducing the cost of readmissions, this improves the quality of care. Various machine learning approaches have been applied to build models in order to identify the diabetic patients who are much likely to be readmitted. All the features are not equally important in predicting hospital readmissions of diabetics. Dimensionality reduction can come into play in this regard to predict the diabetics with high risk of hospital readmissions more efficiently. This research aims at selecting the critical features that have higher impact on the prediction of hospital readmissions. A comparative analysis on the performances of four classifiers - Naïve Bayes, Random Forests, Logistic Regression, and Multilayer Perceptron is also provided in our study. In this paper, Random Forests-based ensemble method was applied to determine the relative importance of every feature. Performance of Multilayer Perceptron was found more efficient with better accuracy and AUC score for independent test set.