Machine learning-based mortality prediction models for non-alcoholic fatty liver disease in the general United States population

Background & AimsNowadays, the global prevalence of non-alcoholic fatty liver disease (NAFLD) has reached about 25%, which is the most common chronic liver disease worldwide, and the mortality risk of NAFLD patients is higher. Our research created five machine learning (ML) models for predicting overall mortality in ultrasound-proven NAFLD patients and compared their performance with conventional non-invasive scoring systems, aiming to find a generalizable and valuable model for early mortality prediction in NAFLD patients. MethodsNational Health and Nutrition Examination Survey (NHANES)-III from 1988 to 1994 and NHANES-III related mortality data from 2019 were used. 70% of subjects were separated into the training set (N = 2262) for development, while 30% were in the testing set (N= 971) for validation. The outcome was all-cause death at the end of follow-up. Twenty-nine related variables were trained as predictor features for five ML-based models: Logistic regression (LR), K-nearest neighbors (KNN), Gradient-boosted decision tree (XGBoost), Random forest (RF) and Decision tree. Five typical evaluation indexes including area under the curve (AUC), F1 score, accuracy, sensitivity and specificity were used to measure the prediction performance. Results3233 patients with NAFLD in total were eligible for the inclusion criteria, with 1231 death during the average 25.3 years follow up time. AUC of the LR model in predicting the mortality of NAFLD was 0.888 (95% confidence interval [CI] 0.867-0.909), the accuracy was 0.808, the sensitivity was 0.819, the specificity was 0.802, and the F1 score was 0.765, which showed the best performance compared with other models (AUC were: RF, 0.876 [95%CI 0.852-0.897]; XGBoost, 0.875 [95%CI 0.853-0.898]; Decision tree, 0.793 [95%CI 0.766-0.819] and KNN, 0.787 [95%CI 0.759-0.816]) and conventional clinical scores (AUC were: Fibrosis-4 Score (FIB-4), 0.793 [95%CI 0.777-0.809]; NAFLD fibrosis score (NFS), 0.770 [95%CI 0.753-0.787] and aspartate aminotransferase-to-platelet ratio index (APRI), 0.522 [95%CI 0.502-0.543]). ConclusionsML-based models, especially LR model, had better discrimination performance in predicting all-cause mortality in patients with NAFLD compared to the conventional non-invasive scores, and an interpretable model like Decision tree, which only used three predictors: age, systolic pressure and glycated hemoglobin, is simple to use in clinical practice.