Evaluation of the Prediction Algorithms for the Diagnosis of Hepatic Dysfunction

Abstract

Acute liver failure develops due to liver dysfunction. Early diagnosis is crucial for acute liver failure, which develops in a short time and causes serious damage to the body. Prediction processes based on machine learning methods can provide assistance to the physician in the decision-making process in order for the physician to make a diagnosis earlier. This study aims to evaluate three recently presented algorithms with high predictive capabilities that can assist the doctor in determining the existence of acute liver failure. In this study, the prediction performances of the XGBoost, LightGBM, and NGBoost methods are examined on publicly available data sets. In this research, two datasets are used; the first dataset was gathered in the “JPAC Health Diagnostic and Control Center” during the periods 2008–2009 and 2014–2015. The dataset includes a total of 8785 patients' information, and it mostly does not contain patients' information that "acute liver failure" was developing. Furthermore, a dataset collected by Iesu et al., containing information on patients who developed or did not develop "acute liver dysfunction," is used for the second evaluation. According to the information obtained from the data set, "acute liver dysfunction" developed in 208 patients, while this situation did not develop in 166 patients. It is observed within the scope of the evaluations that all three algorithms give high estimation results during the training and testing stages, and moreover, the LightGBM method achieves results in a shorter time while the NGBoost method provides results in a longer time compared to other algorithms.