Investigating Some Biological Parameters in Patients with Diabetes to Diagnose the Disease Using a Machine Learning Approach

Abstract

Background: Diabetes has several complications and late diagnosis of this disease leads to an increase in the complications. The present study aimed to investigate the possibility of predicting diabetes using machine learning techniques. Methods: This study was a cross-sectional descriptive-analytical study. The population included the people referred to Falavarjan Social Security Center in Isfahan province in Iran in 2020 for diabetes screening. Blood samples were collected from 250 diabetic patients and 100 healthy non-diabetic samples. Then, glucose, cholesterol, triglyceride, high-density lipoprotein (HDL), low-density lipoprotein (LDL), and very low-density lipoprotein (VLDL) were measured and some characteristics such as height, weight, age and gender were collected from patients’ records. Finally, the data were analyzed and compared using the k-nearest neighbor (KNN) algorithm, artificial neural networks (ANNs), support vector machine (SVM), Naive Bayes, and decision tree (DT). All analyses and modeling were performed in Python programming environment. Results: In all criteria, the best results were obtained by SVM with an accuracy of 0.98, followed by ANNs with an accuracy of 0.96, respectively. Then, the K-NN algorithm with an accuracy of 0.87, Naive Bayes with an accuracy of 0.87, and DT with an accuracy of 0.76 were considered. Conclusion: Both ANNs and linear SVMs are recommended as superior final models for the diagnosis of diabetes due to their higher performance (accuracy) in final decision-making.