A study using the approach of machine learning algorithm in detection of diabetes

The signs and symptoms used in the early detection and diagnosis of diabetes with classical methods may not always be sufficient. In this thesis, it has been tried to develop a different perspective other than classical methods in the detection and diagnosis of diabetes by using machine learning algorithms. The Early-stage diabetes risk prediction dataset obtained from UCI Machine Learning Repository was used in the study. The dataset consists of 520 patient records, 320 of whom are diabetic and 200 are not. The patient profile information consists of 16 attributes and class information indicating diabetic and non-diabetic. The data set is divided into two parts as training (70%) and test (30%) dataset. Data preprocessing steps were applied to get better performance from the dataset. Nine features (Polydipsia, Polyuria, Sudden weight loss, Partial paresis, Gender, Irritability, Polyphagia, Alopecia, Age) were selected using the chi-square method, one of the feature selection techniques. To improve the accuracy performance, 10-fold cross-validation was performed. Accuracy, precision, sensitivity, F1-score, and ROC curve validation metrics were used to improve the classification performance quality of our model. Our aim in the thesis study is to compare the accuracy performances of single-based machine learning algorithms (Logistic Regression, Support Vector Machines, K-Nearest Neighbor, Naive Bayes, and Artificial Neural Networks) and ensemble machine learning algorithms (Decision Trees, Random Forest, AdaBoost, Gradient Boosting, XGBoost Voting, Super Learner) which are used to predict diabetes. The highest accuracy estimate was obtained with the Super Learner ensemble learning algorithm, with a value of 99,4%. This study means that ensemble machine learning algorithms can be considered better than single-based machine learning algorithms in predicting diabetes by using the Early-stage diabetes risk prediction dataset. This method is aimed at early detection of diabetes to assist the clinics.