Data Classification of Early-Stage Diabetes Risk Prediction Datasets and Analysis of Algorithm Performance Using Feature Extraction Methods and Machine Learning Techniques

Abstract

Diabetes is one of the more common diseases in the world today and is one which also plays a role in the development of many other critical or terminal illnesses such as heart diseases, coronary diseases, eye diseases, kidney diseases, and even nerve damage. Thus, early diagnosis is of great importance. With the development of machine learning techniques and artificial intelligence, the estimation of disease risks has started to be widely accepted and applied by researchers and medical doctors. In this study, a machine learning technique was proposed for the prognosis of early onset diabetes. An interface was designed using the MATLAB graphical user interface (GUI). The wrapper-based Particle Swarm Optimization (PSO), Tree Seed Algorithm (TSA), Crow Search Algorithm (CSA), Slime Mould Algorithm (SMA), and Artificial Bee Colony (ABC) algorithms were used to reduce and select the required input attributes. The results obtained with these algorithms were compared by using conventional machine learning algorithms such as Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), K Nearest Neighbor (kNN) and Feed Forward Neural Networks (FFNN). 16 features used in the diagnosis of diabetes, od the wrapper-based feature selection and feature reduction methods 10 features with PSO method, 9 features with TSA method, 13 features with CSA method, 6 features with SMA method and 8 features with ABC method has been determined. The features determined by each respective method were then classified using machine learning algorithms. All combinations have been tried and these are the results of the best five combinations on the results, methods displayed the best classification performances with success rates of PSO + SVM = 97.5, TSA + SVM = 96.15, CSA + FFNN = 99.04, SMA + FFNN = 94.23, and ABC + SVM = 96.73 respectively.