Head-to-head comparison of FibroTest and ELF for diagnosing significant fibrosis (F2-F4) in outpatients with type-2 diabetes using Bayesian latent class models

In patients with metabolic dysfunction-associated steatotic liver disease, two serum fibrosis panels with patented algorithms are among the most cited: the Nash-FibroTest combination (FibroTest), which assesses fibrosis, steatosis and inflammatory activity in one sample, and the Enhanced Liver Fibrosis (ELF) score, which evaluates fibrosis alone. No head-to-head one-step comparison has been published in type-2 diabetes. Our aim was to compare FibroTest and ELF for diagnosing significant fibrosis stages (F2-F3-F4) in 398 type-2 diabetes outpatients, 59 years old median and 40% females, of the prospective ongoing QUID-NASH study. The standard reference, liver biopsy, is misclassified in about 25% of cases, inducing bias in sensitivity, specificity and area under the receiver operating characteristic curve (AUC), in both magnitude and direction, and distorting head-to-head differences. To relax the perfect-gold-standard assumption, we used Bayesian latent class models, which privilege no test, histology included. Posterior mean AUC was 0.94 for FibroTest and 0.76 for ELF, with a mean difference (FibroTest-ELF) of 0.18 (95% credible interval, 0.12–0.25); with no prior information imposed, the model estimated biopsy sensitivity at 0.69 and specificity at 0.74. These results highlight the importance of accounting for reference standard imperfection and support the use of head-to-head comparisons combined with latent class modeling. Trial registration: ClinicalTrials.gov identifier: NCT03634098.