Using Hybrid Model of Particle Swarm Optimization and Multi-Layer Perceptron Neural Networks for Classification of Diabetes

Abstract

Diabetes mellitus is one of the deadliest and chronic diseases that affect persons who have an increase in their blood glucose levels. Type 1 Diabetes Mellitus “T1DM” is considered one of the most dangerous types of diabetes as it is the reason why diabetes is called the silent killer. Due to the common symptoms of type 1 and 2 diabetes, diabetes specialists face doubts about their diagnosis of the type of diabetes in the patient where the uncertainty about the diagnosis of the disease may lead to delays in controlling the potential complications, especially if they have T1DM. The prime motto of this work is to classify the diabetes types accurately. In this work, we have collected a local Palestinian dataset “DataPal” With the assistance of the Palestinian Diabetes Institute. The "DataPal" dataset was applied using machine learning algorithms to predict diabetes types. The “DataPal” consists of 9 predictors used to diagnose diabetes types. The dataset consists of 314 instances of diabetic females. Thus, our samples were for females with both types 1 and 2 diabetes, aged between 5 and 89 years. The local dataset “DataPal” was preprocessed using the K Nearest Neighbor (KNN) algorithm to fill their missing values, wherein medical diagnosis there is no room for error. The Support Vector Machine (SVM) algorithm was applied to the dataset to select the most optimal features to predict diabetes types. Both the two-fold and four-fold cross-validation methods were applied to the datasets to evaluate the applied models fairly. A hybrid Model Particle Swarm Optimization with Multi-Layer Perceptron Neural Networks (PSO-MLPNNs) uses the PSO evolutionary algorithm to train an MLPNNs and find the optimal weight values of the trained network. The PSO-MLPNNs model was applied to the preprocessed dataset. Then the performance of the model was evaluated using different metrics such as the overall accuracy, recall, specificity, and others. The obtained results show that the proposed PSO-MLPNNs model outperformed all models applied in this work in the classification of diabetes types with an overall accuracy (ACC= 98.73%).