Performance evaluation of classification algorithms by excluding the most relevant attributes for dipper/non-dipper pattern estimation in Type-2 DM patients

Abstract

Diabetes Mellitus (DM) is a high prevalence disease that causes cardiovascular morbidity and mortality. On the other hand, the absence of physiologic night-time blood pressure decrease can further lead to morbidity problems such as target organ damage both in diabetics and non-diabetics patients. However, the Non-dipping pattern can only be measured by the 24-hour ambulatory blood pressure monitoring (ABPM) device. ABPM has certain challenges such as insufficient devices to distribute to patients, lack of trained staff or high costs. Therefore, in this study, it is aimed to develop a classifier model that can achieve a sufficiently high accuracy percentage for Dipper/non-Dipper blood pressure pattern in patients by excluding ABPM data. The study was conducted with 56 Turkish patients in Marmara University Hypertension and Atherosclerosis Center and School of Medicine Department of Internal Medicine, Division of Endocrinology between the years 2010 and 2012. Our purpose was to find out if the proposed method would be able to detect non-dipping/dipping pattern through various data mining algorithms in WEKA platform such as J48, NaïveBayes, MLP, RBF. All algorithms were run to get accurate Dipper/non-Dipper pattern estimation excluding the attributes of ABPM data. The results show that Neural Network (MLP and RBF) algorithms mostly produced reasonably high classification accuracy, sensitivity and specificity percentages reaching up to 90.63% when the attributes were reduced. However in medical sciences, sensitivity is taken as a valid and reliable indication for diagnosis. Therefore, MLP had a higher sensitivity percentage (83.3%) than others. Also, ROC values, which had the closest values to 1, were achieved by RBF for each selection mode. ROC was 0.872 for 10 fold CV mode and 0.856 for percentage split mode. Finally, ANN MLP and RBF algorithms were used, and it was observed that RBF algorithm had the highest success rate regarding sensitivity that was 83.3%. In medical diagnosis, a higher sensitivity performance is regarded as a more valid indication of metric than a higher specificity. The proposed model could represent an innovative approach that might simplify and fasten the diagnosis process by skipping some steps in Dipper/non-Dipper diagnosis/prognosis.