A Machine Learning Approach to Differentiating Bacterial From Viral Meningitis

Abstract

Clinical reports indicate that differentiating bacterial from viral (aseptic) meningitis is still a difficult issue, compounded by factors such as age and time of presentation. Clinicians routinely rely on the results from blood and cerebrospinal fluid (CSF) to discriminate bacterial from viral meningitis. Tests such as the CSF Gram stain performed prior to broad-spectrum antibiotic treatment yield sensitivities between 60 and 92%. Sensitivity can be increased by performing additional laboratory testing, but the results are never completely accurate and are not cost effective in many cases. In this study, we wished to determine if a machine learning approach, based on rough sets and a probabilistic neural network could be used to differentiate between viral and bacterial meningitis. We analysed a clinical dataset containing records for 581 cases of acute bacterial or viral meningitis. The rough sets approach was used to perform dimensionality reduction in addition to classification. The results were validated using a probabilistic neural network. With an overall accuracy of 98%, these results indicate rough sets is a useful approach to differentiating bacterial from viral meningitis.