Abstract
Differentiating dengue patients from other acute febrile illness patients is a great challenge among physicians. Several dengue diagnosis methods are recommended by WHO. The application of specific laboratory tests is still limited due to high cost, lack of equipment, and uncertain validity. Therefore, clinical diagnosis remains a common practice especially in resource limited settings. Bayesian networks have been shown to be a useful tool for diagnostic decision support. This study aimed to construct Bayesian network models using basic demographic, clinical, and laboratory profiles of acute febrile illness patients to diagnose dengue. Data of 397 acute undifferentiated febrile illness patients who visited the fever clinic of the Bangkok Hospital for Tropical Diseases, Thailand, were used for model construction and validation. The two best final models were selected: one with and one without NS1 rapid test result. The diagnostic accuracy of the models was compared with that of physicians on the same set of patients. The Bayesian network models provided good diagnostic accuracy of dengue infection, with ROC AUC of 0.80 and 0.75 for models with and without NS1 rapid test result, respectively. The models had approximately 80% specificity and 70% sensitivity, similar to the diagnostic accuracy of the hospital’s fellows in infectious disease. Including information on NS1 rapid test improved the specificity, but reduced the sensitivity, both in model and physician diagnoses. The Bayesian network model developed in this study could be useful to assist physicians in diagnosing dengue, particularly in regions where experienced physicians and laboratory confirmation tests are limited. In many parts of the world, dengue diagnosis still relies on clinical manifestation and basic laboratory tests. However, clinical manifestation of dengue patients is usually unspecific especially during the early stage, which makes accurate dengue diagnosis a challenge for physicians. Recently, the World Health Organization has recommended the use of the NS1 antigen rapid test to assist in dengue diagnosis. Unfortunately, the rapid test is still only in limited use due to its cost and relatively low sensitivity. In this study we constructed Bayesian Network (BN) models from available data and used them to predict the probability of dengue infection. Our BN models provided good performance in making dengue diagnosis, comparable with physician’s diagnosis. Information on NS1 rapid test could improve the diagnostic performance by improving specificity of dengue diagnosis. This study suggested that BN models can be useful to assist physicians in dengue diagnosis, particularly in areas where experienced physicians and laboratory confirmation tests are limited.

This publication has 36 references indexed in Scilit: