Background and aim: Significant hepatic fibrosis is prognostic of liver morbidity and mortality in non-alcoholic fatty liver disease (NAFLD); however, it remains unclear whether non-invasive fibrosis models can determine this end-point. We therefore compared the accuracy of simple bedside versus complex fibrosis models across a range of fibrosis in a multi-centre NAFLD cohort.

Methods: Simple (APRI, BARD) and complex (Hepascore, Fibrotest, FIB4) fibrosis models were calculated in 242 NAFLD subjects undergoing liver biopsy. Significant (F2-4) and advanced fibrosis (F3,4) were defined using Kleiner criteria. Models were compared using area under the receiver operator characteristic curves (AUC). Cut-offs were determined by Youden Index or 90% predictive values.

Results: For significant fibrosis, non-invasive fibrosis models had modest accuracy (AUC 0.707-0.743) with BARD being least accurate (AUC 0.609, P < 0.05 vs others). Using single cut-offs, sensitivities and predictive values were < 80%; using two cut-offs, > 75% of subjects fell within indeterminate ranges. Simple models had significantly more subjects within indeterminate ranges than complex models (99.1-100% vs 82.1-84.4% respectively, P < 0.05 for all). For advanced fibrosis, complex models were more accurate than BARD (AUC 0.802-0.858 vs 0.701, P < 0.05). Using two cut-offs, complex models had fewer individuals within indeterminate ranges than BARD (11.1-32.3% vs 70.7%, P < 0.01 for all). For cirrhosis, complex models had higher AUC values than simple models.

Conclusions: In NAFLD subjects, non-invasive models have modest accuracy for determining significant fibrosis and have predictive values less than 90% in the majority of subjects. Complex models are more accurate than simple bedside models across a range of fibrosis.