SUPPORT VECTOR MACHINES COMBINED WITH FEATURE SELECTION FOR DIABETES DIAGNOSIS

Abstract

Clinical Decision Support Systems (CDSS) are used as a service software which provides huge support to clinical decision making process where the main properties of a patient are matched to a tangible clinical knowledge. Within this gathered important information about patients, the medical decisions can be made more accurately. In this paper we present a CDSS that uses four physiological parameters of patients such as Pre-prandial Blood Glucose, Postprandial Blood Glucose, Hemoglobin A1C (HbA(1c)) and Glucose in Urine to produce a prediction about the possibility of being diabetic. According to collected reference data provided from hospitals, the disease can be predicted by comparing the input data of patients. If the system cannot procure a prediction about patients' status with these parameters, then the second phase which uses soft computing techniques is put into process with requesting additional data about patients. Our conducted experiments show that the diagnosis can be established in a breeze by getting the patients information with % 80 accuracy. Support Vector Machines were applied to achieve maximum success rate with nine different physiological parameters such as; Pregnancy, glucose, blood pressure, skin fold, insulin, Hemoglobin A1C, body mass index, family tree and age. Four different Kernel Functions are implemented in case studies and classification process is optimized by reducing attributes with feature selection algorithms. This represents an improvement in classification of CDSS, while reducing computational complexity.