IDENTIFICATION OF HEPATIC STEATOSIS IN LIVING LIVER DONOR BY MACHINE LEARNING MODELS

Abstract

Background: Living donor liver transplantation (LDLT) has been a promising alternative option for end-stage liver disease patients because of limited cadaveric donation in Asia. Selecting optimal donors is crucial for both donor and recipient’s safety and hepatic steatosis is one to be considered. We aimed to build a prediction model evaluating the macrovesicular steatosis in potential donors using machine learning algorithms with non-invasive variables. Methods: The study population consisted of potential living donors who underwent donation workup including percutaneous liver biopsy. We split the whole data into two sets; 1) train set for the learning process; 2) test set for evaluating model performance depending on the date of liver biopsy. We defined the cut off value of macrovesicular hepatic steatosis as 5 %. We employed the algorithms of the random forest (RF), support vector machine (SVM), regularized discriminant analysis (RDA), mixture discriminant analysis (MDA), flexible discriminant analysis (FDA), deep neural network (DNN) as well as traditional logistic regression. Results: A total of 1654 participants, 1098 (66.4%) were male and the mean age was 31.4 ± 9.4 years. The mean value of BMI was 24.2 ± 3.4 kg/m2. The average for hepatic Hounsfield units was 54.4 ± 10.0. The pathology results reported that 68.7% of people had macrovesicular steatosis less than 5%, while 44.5% had more than 5%. The train set consisted of 1166 (70.5%), and the test set of 488 people. The RF and SVM had the best prediction power and the prediction performances of each model are as followed; 1) logistic model of 0.812 and 0.818 for train and test sets; 2) RF of 1.000 and 0.799; 3) SVM of 0.868 and 0.797; 4) RDA of 0.840 and 0.805; MDA of 0.834 and 0.795; 5) FDA of 0.830 and 0.872; and 6) DNN of 0.931 and 0.791. Conclusion: Our algorithm to predict macrovesicular steatosis with generally assessed parameters would be beneficial for finding optimal potential living donors avoiding superfluous liver biopsy.