Machine learning models are superior to noninvasive tests in identifying clinically significant stages of NAFLD and NAFLD-related cirrhosis

Background and Aims We assessed the performance of machine learning (ML) models in identifying clinically significant NAFLD-associated liver fibrosis and cirrhosis. Approach and Results We implemented ML models including logistic regression (LR), random forest (RF), and artificial neural network to predict histological stages of fibrosis using 17 demographic/clinical features in 1370 patients with NAFLD who underwent liver biopsy, FibroScan, and labs within a 6-month period at multiple U.S. centers. Histological stages of fibrosis (>= F2, >= F3, and F4) were predicted using ML, FibroScan liver stiffness measurements, and Fibrosis-4 index (FIB-4). NASH with significant fibrosis (NAS >= 4 + >= F2) was assessed using ML, FibroScan-AST (FAST) score, FIB-4, and NAFLD fibrosis score (NFS). We used 80% of the cohort to train and 20% to test the ML models. For >= F2, >= F3, F4, and NASH + NAS >= 4 + >= F2, all ML models, especially RF, had primarily higher accuracy and AUC compared with FibroScan, FIB-4, FAST, and NFS. AUC for RF versus FibroScan and FIB-4 for >= F2, >= F3, and F4 were (0.86 vs. 0.81, 0.78), (0.89 vs. 0.83, 0.82), and (0.89 vs. 0.86, 0.85), respectively. AUC for RF versus FAST, FIB-4, and NFS for NASH + NAS >= 4 + >= F2 were (0.80 vs. 0.77, 0.66, 0.63). For NASH + NAS >= 4 + >= F2, all ML models had lower/similar percentages within the indeterminate zone compared with FIB-4 and NFS. Overall, ML models performed better in sensitivity, specificity, positive predictive value, and negative predictive value compared with traditional noninvasive tests. Conclusions ML models performed better overall than FibroScan, FIB-4, FAST, and NFS. ML could be an effective tool for identifying clinically significant liver fibrosis and cirrhosis in patients with NAFLD.