Diagnosis of Hepatitis Disease with Logistic Regression and Artificial Neural Networks

Abstract

Hepatitis C refers to the inflammatory state of the liver caused by viruses, bacteria, fungi, and exposure to toxins such as alcohol and self-immunity. The diagnosis requires investigating many laboratory tests and comparing the results to those of the former patients with the same conditions. This study presents the results of our experiments to build a hybrid system that combines both neural networks and logistic regression for the diagnosing of the hepatitis dataset using clinical and laboratory test results. The first experiment compared the performances of Multilayer Perceptual Neural Networks (MLPNN) and Radial Basis Function Neural Network (RBFNN) versus the conventional and stepwise Logistic Regression (LR) algorithms, where the results demonstrated the ability of neural networks to deliver better performance than LR models. In the second experiment, the features selected by backward and forward LR models have been evaluated for the improvement of the performances of MLPNN and RBFNN models. The hepatitis dataset was downloaded from the machine-learning repository by the University of California at Ervine. Missing values have been imputed with a separate Classification and Regression Tree (C&RT) for each attribute. Classification models have been evaluated in terms of statistical accuracy, specificity, sensitivity, F1-score and the Area Under the Receiver Operating Characteristic Curve (AUCROC). Experimental results showed that the performances of neural network models have been improved when employing stepwise LR models to select only the predictive attributes. The hybrid system which combined both backward stepwise LR for attribute selection and MLPNN for classification has outperformed other systems in the diagnosis of the hepatitis dataset with 0.973 AUCROC for the training subset and 0.886 for the test one.