Classification of Chronic Kidney Disease Using Hybrid Models of Neural Networks and Evolutionary Algorithms رسالة ماجستير

Abstract

Chronic Kidney Disease (CKD) is a risk factor for cardiovascular disease and has a significant economic impact on healthcare systems. Early detection of chronic kidney disease can save a person's life from a heart attack. Artificial Intelligence (AI) has emerged as a new tool that helps in the early detection of disease and predicting its occurrence. AI has a good impact on saving lives, providing a treatment plan for the disease, and conducting more developmental research. Various artificial intelligence techniques can be used to classify and predict chronic kidney diseases by applying them to medical data. This thesis presents hybrid models combining evolutionary algorithms, neural networks, and machine learning techniques to classify chronic kidney disease. In this thesis, global and local datasets were used as well. In the first stage, several different machine learning models were applied to datasets, including Decision Tree (DT), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Multi-Layer Neural Networks (MLPNNs). In the second stage, several hybrid models of evolutionary algorithms, including biogeography-based optimization (BBO), particle swarm optimization (PSO), and genetic algorithms (GAs) were trained on multi-layer neural networks (MLPNNs) to obtain the best results in CKD classification. The DT, SVM, KNN, and MLPNNs models were applied to the global dataset for chronic kidney disease classification, and revealed accuracy results of 97%, 99.5%, 98.2%, and 99.8%, respectively. Furthermore, the MLPNNs and SVM models showed the highest accuracy and the best models in classification with close accuracy rates. These models were also applied to the Palestinian local dataset to classify chronic kidney diseases, and the accuracy results obtained were: 96.4%, 96.2%, 93.6%, and 98.1%, respectively, in the same order mentioned previously. The MLPNNs model revealed the highest accuracy and was the best model in classification followed by the DT model. VI The experimental results in applying the hybrid models to the global dataset showed that both GAs-MLPNNs and BBO-MLPNNs were almost similar in performance, with the results for GAs-MLPNNs being: 99.5%, 99.6%, 99.3%, 99.6%, and 99.6% accurate, sensitive, specific, precise, and F-scored, respectively. In addition, the results in applying the hybrid models to the local dataset of both the GAs-MLPNNs and BBO-MLPNNs were also similar in performance, as they were: 99%, 99%, 99.1%, 99.2%, and 99.1% accurate, sensitive, specific, precise, and F-scored, respectively