Predicting Hospital Readmission Patterns of Diabetic Patients using Ensemble Model and Cluster Analysis

Abstract

Diabetes is a chronic illness that affects around 425 million people globally in 2017, and this is predicted to increase to 629 million by the end of 2045. The ability to analyze and predict the readmission patterns of diabetic patients would allow the optimization of hospital resources and assessment of treatment effectiveness. This paper proposes an ensemble model to predict hospital readmission by choosing from a pool of 15 models, made up of variants of Logistic Regression, Decision Trees (DT), Neural Network (NN) and Augmented Naive Bayes (NB) networks. The final ensemble model was assembled using the five best models, determined based on individual model accuracy and the Jaccard distance between them, to maximize overall accuracy and sensitivity. The final ensemble contained DT (CHAID), Tree Augmented Naive Bayes network, DT (CHAID with boosting), Neural Network with bagging and DT (CART with boosting). Compared against existing predictive models, the proposed ensemble was able to achieve improved sensitivity at 56% while maintaining comparable accuracy at 63.5%. Cluster analysis after performing principal component analysis on the dataset revealed 4 distinct clusters of patients. Patients with a history of in-patient visits or if they received a high amount of treatment in their current visit were found more likely to be readmitted.