Prediction of dengue infection severity using classic and robust discriminant approaches

Abstract

Listen

Translate article

Dengue infection is one of feared diseases in the public because it often results in death in sufferers. Patients suspected of dengue infection are usually routinely drawn their blood to be checked in the laboratory examination. Unfortunately, death can be caused by a lack of speed and proper handling according to the severity of the patient. Refer to this problem, it is necessary to predict dengue infection severity based on blood diagnose results. This is important to prepare the precise treatment according to the severity of patients in order to reduce the number of death from this disease. Because the patient's blood examination result is a multivariate dataset then in this paper the prediction was solved using multivariate method, namely discriminant analysis. In this method, the parameter estimation was carried out using Maximum Likelihood (ML) method. This leads to classic discriminant analysis. Unfortunately, the ML method is heavily influenced by outlier so the estimator becomes less precise when data has been contaminated by outliers. To overcome this problem, a robust estimation method using Minimum Covariance Determinant (MCD) was used. This leads to the robust discriminant analysis. This study used a sample of dengue infection patient medical record data from Surabaya Hajj Hospital. The result of this study showed that the appropriate analysis for sample data was the quadratic discriminant analysis. Furthermore, the robust quadratic model with MCD estimator produced better prediction than the classic quadratic model with ML estimator. The robust quadratic model produced percentage of classification accuracy of 87.2% in the male patient training data which is greater than the classic quadratic model accuracy of 85.7%. In the female patient training data, the robust quadratic model produced percentage of classification accuracy of 88.7% which is greater than the classic quadratic model accuracy of 80.7%. In addition, the MCD estimator was able to detect more outlier data than the ML estimator.