FEATURE SELECTION TECHNIQUES IN PREDICTING EARLY STAGE DIABETES MELLITUS

Abstract

Diabetes is noted to be one of the chronic diseases that affect the ability of the body to produce insulin. Patients of this disease experience different health complications such as stroke, and foot ulcers. Our study provides a comprehensive model to predict the disease at an early stage based on some features collected on the patient. The model would help physicians to diagnose diabetes mellitus at an early stage before its health complication worsens. The dataset used was the SHD dataset which consists of 521 instances and 17 attributes. The methods used are supervised machine learning algorithms, including logistic regression, support vector machines-linear and nonlinear kernel, random forest, decision tree, K-nearest neighbor, and Naïve Bayes. The variance feature selection technique was used to select the most important features for both the training dataset and the testing dataset. These features include age, polyuria, polydipsia, sudden weight loss, polyphagia, visual blurring, itching, delayed healing and partial paresis The results showed that Random Forest outperformed the other comparative algorithms with an accuracy of 100%.