Rule extraction for fatty liver detection using neural networks

Non-alcoholic fatty liver disease (NAFLD) is one of the most common diseases in the world. Recently the FibroScan device is used as a noninvasive, yet costly method to measure the liver’s elasticity as a NAFLD indicator. Other than the cost, the diagnosis is not widely accessible to all patients. On the other hand, early detection of the disease can prevent later risks. In this study, we aim to use learning methods to infer the NAFLD severity level, only based on clinical tests. A dataset was constructed from clinical and ultrasonography data of 726 patients who were diagnosed with different NAFLD severity levels. Artificial neural networks (ANN) were used to model the relationship between NAFLD and the clinical tests. Next, a method was used to analyze the ANN and extract compact and human understandable rules. The derived rules can detect the fatty liver disease with an accuracy above 80%.