Comparative Performance of Random Forest and Multilayer Perceptron Models for Diabetes Classification using the Pima Indian Diabetes Dataset

Abstract

Diabetes has become one of the most severe public health issues around the globe and is especially severe in developing nations. Aging leads to an increase in population, as well as impact rate. If timely diagnosis is not achieved, diabetes can significantly increase the prevalence of microvascular and macrovascular complications that can lead to increased costs of care and mortality. Commonly believed, none of these tests, namely Hemoglobin A1c (HbA1c), the Oral Glucose Tolerance Test (OGTT) or the Fasting Plasma Glucose (FPG) is a predictive or predisposing test. Making a model that can foresee these potentially fatal diseases is crucial for identifying those at risk before they become very sick. Using the Pima Indians Diabetes Dataset of 768 with 9 variables, this research assesses diabetes prediction using MLP Neural Network (MLP-NN) and Random Forest (RF) models. Several methods for feature engineering and data preparation were used to improve the model's performance. We validated our model using a three-fold stratified cross validation technique. This study uses ROC-AUC, F1-Score, Accuracy, and Brier for evaluating the performance of the model. With an accuracy of 76.43%, F1-Score of 63.28%, Brier 15.62% and ROC-AUC of 83.86%. However, after the simulation of this model if design and implemented accordingly, it will predict accurately and be reliable based on the results.