Optimal Feature Selection in Diabetes Classification Using the MLP Algorithm

Abstract

In 2021, approximately 531 million people worldwide were affected by diabetes, with 90% diagnosed as type 2. Diabetes often coexists as a comorbidity with other conditions such as kidney and heart disease. The research aims to employ machine learning for diabetes classification, with the Multilayer Perceptron (MLP) algorithm being a key component in the early detection process. The experiments utilized data from the UCI database of Sylhet hospitals, featuring 16 attributes and 2 classes indicating positive and negative diabetes cases. Performance testing using the MLP algorithm involved varying the number of neurons in the hidden layer. The research architecture is denoted as n:p:m, where n represents 16 neurons based on the attributes, m signifies 2 neurons based on the number of classes, and p undergoes variations. The machine learning tool employed in this research is Weka. Within the Weka tool, MLP offers types of hidden layer neuron configurations: 'a', 't', 'i', and 'o'. The test results, conducted with 520 training data and testing on the same dataset, yielded accuracies of 98.85%, 98.85%, 99.42%, and 98.46% for types 'a', 't', 'i', and 'o', respectively.

Keywords

• diabetes, multilayer perceptron, machine learning