Background:  Liver fibrosis progression is a significant health concern in those with HCV/HIV co-infection. Biopsy is costly and invasive, however, the utility of existing fibrosis prediction models developed with non-invasive biomarkers is suboptimal in co-infected patients. We aimed to assess predictive capacity of current models in the AIDS Clinical Trials Group 5178 and improve discriminatory capacity with novel statistical methods.

Methods:  Pre-treatment data from 210 of the first 218 study entries were available, including demographics, laboratory values, and METAVIR stage from biopsy. Current models tested were the aspartate aminotransferase (AST)/alanine aminotransferase (ALT) ratio (AAR), age-platelet index (API), AST/platelet ratio index (APRI), ordinally scored platelet count/AAR/INR, and FIB-4. Area under the ROC curve (AUROC) was assessed using pre-defined cut-points. Individual covariates were assessed using ordinal logistic regression, then entered into a classification and regression tree (CART) model to predict specific METAVIR stages. The CART model was “boosted” to decrease classification error. Boosting weights incorrectly classified samples with iterative resampling so that they have a greater chance of being resampled on the next iteration.

Results:  The cohort consisted of subjects with well-compensated HIV (median CD4 503, range 135 to 2162; 85% on ART). Of these, 31.2% had significant fibrosis (METAVIR F3/F4) and 14% had cirrhosis (F4). Differences between those with and without F3/F4 were seen for CD4⁺ count, platelet count, and INR. For current models, FIB-4 performed best. At a cutoff ≥3.25 it had 88% specificity for F4 and >86% negative predictive values for F3/F4, but AUROCs were low (0.58±0.05 and 0.56±0.03) as were sensitivities. Similar trends were seen for predicting absence of F3/F4. After univariable ordinal regression, age, gender, CD4⁺, HCV/HIV viral loads, INR, platelets, ALT/AST, bilirubin, and ART or PI alone were chosen for inclusion in CART models. Training sets included 75% of the data; 25% was for cross-validation. Overall misclassification rate was 45%. After boosting, it was 1.31%. AUROC for predicting each individual fibrosis stage (F1-F4) for training/validation sets were 93/89, 93/90, 94/84, and 97/91%.

Conclusions:  Current models for fibrosis assessment have poor discriminatory capacity in HCV/HIV-co-infected patients. Statistical methods such as boosting have potential to substantially improve stage-wise predictive ability, thus discriminating mild, moderate, and severe fibrosis.