A Kernel k-Means-based Method for Diabetes Diagnosis

Abstract

Diabetes diagnosis is important due to the death and complication consequences caused by the disease. It thus has attracted much research attention and effort in Artificial Intelligence to support human decisions. Our work proposes a kernel k-means-based predictive method and explores attribute selections for effective and robust diabetes diagnosis. This method uses homogeneous subclusters in the high dimensional kernelized feature space to compute the distance of a new instance to those subclusters and classify it accordingly. The PIMA and MIMIC data sets are respectively used for training and testing. Our experimental results could identify the best effective attribute groups and show that the proposed method outperforms existing ones for the task.