Predicting Early Readmission of Diabetic Patients: Toward Interpretable Models

摘要

Hospital readmission among diabetic patients is a common phenomenon throughout the world. Predicting such patients with a high risk of readmission at the time of discharge even before can help us to provide better health care to them. It can minimize the cost associated with readmission too. This study aims at creating a decision support system that can find diabetic patients who are prone to early readmission. To do that, several data mining techniques have been used. Two regular classifiers using the decision tree and random forest have been developed. After that, two rule-based classifiers using Repeated Incremental Pruning to Produce Error Reduction (RIPPER) and PART algorithms have been developed to provide better interpretability and understandability of the support system. Between two regular classifiers, the random forest has shown a better performance with 89.5% accuracy and 89.5% recall. And, between rule-based classifiers, PART has demonstrated promising performance with 86.6% accuracy and 84.6% recall. Using these classifiers, smart and improved health care can be ensured.