Reliable prediction of anti-diabetic drug failure using a reject option

The medical care for patients with type 2 diabetes generally involves ingestion of oral hypoglycemic agents in order to lower their glucose level. When predicting the result of the medication using a classification approach, high prediction accuracy of the classifier is essential because of high misclassification costs. The application of a reject option to this approach supports more accurate prediction, allowing for human experts to examine when the classifier is unreliable to predict. In this paper, we propose a reject option framework based on heterogeneous ensemble learning through a two-phase fusion. The first phase is to calculate confidence scores, which are used to determine whether to predict, and the second phase is to derive final prediction results by fusing the outputs from multiple heterogeneous classifiers. We confirm the effectiveness of the proposed method to the anti-diabetic drug failure prediction problem through experiments on actual electronic medical records data of type 2 diabetes. The proposed method yields a better trade-off between accuracy and rejection than other reject options with statistical significance. A lower prediction error is obtained for the same degree of rejection. We obtained desirable accuracy for the anti-diabetic drug failure problem by applying the proposed reject option, which allows using the classification approach in practice. The accurate prediction of drug failure at the moment of prescription can assist clinical decisions for patients. In addition, in-depth analysis can be considered for those prescriptions that are predicted as failure or rejected.