Artificial Neural Networks Based Dengue Diagnosis Prediction Model

Abstract

Dengue incidents have increased 30 times within the last 50 years. Annually, in 100 endemic countries, 50 – 100 million infections occur. The spread of the Dengue virus increases the need for immediate diagnosis of Dengue and begins the treatment as soon as possible in the earliest stages to avoid medical complications. However, diagnosis of Dengue on time and treatment in the early stages are the biggest challenges to hospitals. Specifically, depending heavily on human resources affects the prompt identification of Dengue patients and treating them. The delay in diagnosis of Dengue and human errors could lead to crucial medical situations. It is, therefore, important to reducing the time taken for Dengue diagnosis and this study tries to address this problem. In line with that, there is a significant need to automate the process of diagnosing Dengue as early as possible with higher efficiency along with utilizing fewer human resources, which could reduce human errors and delays. This study aims to identify the best performing Dengue diagnosis prediction model using Artificial Neural Network (ANN) to automate the process of Dengue diagnosis. In developing the prediction model, feature selection is a crucial phase as it has been used to identify key attributes that directly impact the accuracy of the Dengue prediction model. In this study two approaches were used for the feature selection; Principal Component Analysis (PCA) and Wrapper feature selection methods with Naïve Bayes, KNN, & J48 as classifiers. The PCA made a huge impact in developing higher accuracy Dengue diagnosis prediction model. Further, with PCA the initial 22-dimensional system was reduced to the 8-dimensional system with a cumulative variance of 59 %. A total of 360 Dengue diagnosis prediction models have developed using four different sets of feature combinations and with different hyperparameters. In conclusion, ANN with PCA, learning rate of 0.01, batch size of 32 and 200 epochs shows the highest accuracy of 73.41% and it is most suitable for the dengue diagnosis prediction based on current dataset.