Background: The precise estimation of cases with significant fibrosis (SF) is an unmet goal in non-alcoholic fatty liver disease (NAFLD/MASLD).
Aims: We evaluated the performance of machine learning (ML) and non-patented scores for ruling out SF among NAFLD/MASLD patients.
Methods: Twenty-one ML models were trained (N = 1153), tested (N = 283), and validated (N = 220) on clinical and biochemical parameters of histologically-proven NAFLD/MASLD patients (N = 1656) collected across 14 centres in 8 Asian countries. Their performance for detecting histological-SF (≥F2fibrosis) were evaluated with APRI, FIB4, NFS, BARD, and SAFE (NPV/F1-score as model-selection criteria).
Results: Patients aged 47 years (median), 54.6% males, 73.7% with metabolic syndrome, and 32.9% with histological-SF were included in the study. Patients with SFvs.no-SF had higher age, aminotransferases, fasting plasma glucose, metabolic syndrome, uncontrolled diabetes, and NAFLD activity score (p < 0.001, each). ML models showed 7%-12% better discrimination than FIB-4 to detect SF. Optimised random forest (RF) yielded best NPV/F1 in overall set (0.947/0.754), test set (0.798/0.588) and validation set (0.852/0.559), as compared to FIB4 in overall set (0.744/0.499), test set (0.722/0.456), and validation set (0.806/0.507). Compared to FIB-4, RF could pick 10 times more patients with SF, reduce unnecessary referrals by 28%, and prevent missed referrals by 78%. Age, AST, ALT fasting plasma glucose, and platelet count were top features determining the SF. Sequential use of SAFE < 140 and FIB4 < 1.2 (when SAFE > 140) was next best in ruling out SF (NPV of 0.757, 0.724 and 0.827 in overall, test and validation set).
Conclusions: ML with clinical, anthropometric data and simple blood investigations perform better than FIB-4 for ruling out SF in biopsy-proven Asian NAFLD/MASLD patients.
Keywords: NASH; artificial intelligence; fatty liver; liver fibrosis; mortality.
© 2024 John Wiley & Sons Ltd.