CHeCS is a retrospective/prospective, observational multicenter study that includes hepatitis patients from 4 large US health systems; the study design and execution have been described previously.^14,15 CHeCS follows US Department of Health and Human Services guidelines for the protection of human subjects; protocols were approved by the institutional review boards at each site. Briefly, electronic administrative claims data and electronic health records (EHRs) for HCV patients age 18 years and older who received health services at any study site between January 1, 2006, and December 31, 2013, were used to identify study candidates; eligibility was confirmed during medical chart abstraction. To reflect a real-world setting, patients were excluded if they had participated in a clinical trial of HCV antiviral therapy or were co-infected with hepatitis B; HIV co-infected patients remained eligible for inclusion.

For each patient, observation commenced at the index date. For treated patients, this was defined as the date of first treatment initiation. For untreated patients, an automated computer algorithm using a frequency-matched approach was used to select index dates at or after the HCV diagnosis date (the earliest date of an HCV-associated diagnosis code and/or a positive laboratory test in the patient’s medical records). We also performed a sensitivity analysis by substituting the last treatment initiation date for the index date to include prior treatment experience at baseline. Patients receiving ongoing HCV treatment without sufficient follow-up evaluation or those who received a liver transplant before their index date were excluded.

We used the FIB4 score as the outcome of interest and APRI for the sensitivity analysis, calculated using the following formulas: FIB4=Age(y)×aspartateaminotransferase(U/L)Plateletcount(109/L)×[alanineaminotransferas(U/L)]1/2 APRI=Aspartateaminotransferase(UL)UpperlimitofnormalPlateletcount(109/L)×100

We used laboratory tests collected within 7 days of one another and patient age at the time of laboratory assessment. Aminotransferase levels and platelet counts measured during antiviral therapy were excluded because therapy can influence these parameters.¹⁶ The FIB4 score and APRI were summarized using a median smoother for every 90-day interval. Because of non-normality, data were log-transformed. Patients with at least 2 FIB4 intervals after the index date were included in the analysis.

Demographic information included age, sex, race/ ethnicity, estimated median annual household income, and insurance status at index date. Racial/ethnic classifications included Native American, Asian, black, white, or Pacific Islander/Hawai’ian. For analytic purposes, Asian and Pacific Islander/Hawai’ian was considered a single group. Insurance status (insured/not insured) was classified by the encounter nearest the index date. Consistent with our published results,¹⁴ income was categorized using US Census variable P053001.¹⁷ Index year was used as a baseline control variable.

A subset of commonly known risk factors (age, sex, race, HCV genotype [GT], Charlson/Deyo index [calculated from inpatient, outpatient, and claims data for 12 months before the index date], diabetes diagnosis, and recent drug/alcohol abuse) were selected as covariates to study the effect of HCV treatment on progression of liver fibrosis. Additional variables were included in the analysis to adjust for possible confounding owing to treatment selection bias. Clinical data captured before and at index included comorbid conditions,¹⁸ HIV co-infection, and laboratory testing. We also used International Classification of Diseases, 9th revision diagnostic and procedure codes as well as Current Procedural Terminology-version 4 procedure codes to assess possible contraindications to antiviral therapy (Supplementary Table 1).

Detailed antiviral medication data (drug name, start/ stop dates) were collected via chart abstraction at or after the index date. Combination therapy was identified when multiple hepatitis drugs were administered concomitantly. Data on routine HCV RNA quantification tests were obtained via the EHR. Patients were classified as having achieved SVR if laboratory tests collected 12 weeks or more after therapy showed undetectable viral RNA loads; otherwise they were classified as having treatment failure (TF). Treatment status (SVR, TF, or untreated) was considered a time-varying covariate.

We hypothesized that liver fibrosis would change over time, and that changes would be associated with antiviral HCV therapy as well as known risk factors. Before our analysis, however, we assessed possible selection bias for receipt of antiviral therapy. If bias was present, we used a weighted propensity-score method to adjust for differences between patients who had or had not received antiviral therapy. This method included logistic regression modeling of the controlling baseline covariates. Patient data were weighted based on the inverse propensity to receive/not receive antiviral therapy (the IPTW approach).¹⁹ Covariate balance was checked after adjustment.

We then examined the effects of treatment and time on FIB4 progression (in a natural log scale) using mixed models with specific covariance patterns for longitudinal data, adjusted by IPTW. Specifically, the trajectory of FIB4 for each individual was modeled as a function of time and of time-varying patient treatment group variables. Time was treated as a nominal class variable. We considered a variety of covariance structures, including directly-specified compound symmetry, autoregressive with/without heterogeneous variances, and indirectly specified structures, based on random coefficients in time. Any variable with individual (univariate) effects on the change in FIB4 was a candidate for the initial model. We used the Bayesian Information Criterion (BIC) to assess the models’ goodness-of-fit. The final model retained variables that significantly contributed to differentiation of FIB4 trajectories with the lowest (ie, best-fitting) BIC scores.

We also tested treatment-by-risk factor (covariate) interactions, followed by further evaluation of quantitative interactions (magnitude but not direction of response varies by the presence/absence of the risk factor) or qualitative interaction (change in direction of response; eg, a treatment benefits the young but harms the old).²⁰ All interactions were quantitative; because they did not contribute to goodness-of-fit and could potentially complicate the models, they were dropped from the final analysis.

Four sensitivity analyses were conducted, as follows: (1) APRI as the outcome of interest, owing to the inclusion of age in the formula for calculating FIB4 and thus high correlation between these; (2) substitution of last treatment initiation date for index date, owing to patients receiving more than one course of treatment; (3) a 1-to-1 (treated and untreated) matched cohort to match controls at baseline; and (4) omission of several important treatment selection or prognostic variables (eg, baseline fibrosis, covariates related to treatment contraindications) for propensity score calculation. In the last sensitivity analysis, we tested the treatment group effect to determine whether it had changed; a robust treatment effect indicates that unobserved confounding can be disregarded.²¹

Table 1 describes our sample. Asian patients were over-represented (6%) because of the participation of a large health care system in Hawai’i. Notably, at baseline roughly half (2290; 48%) of all patients had advanced fibrosis; patients with and without advanced fibrosis received antiviral treatment at the same rate (35%). The median time from initial indication of HCV infection/ diagnosis to initiation of antiviral treatment was approximately 2 years; the interquartile range (25%–75%) was 1 to 7 years.

In unadjusted comparisons, untreated patients were older, more likely to have a history of substance abuse, and more likely to be black than untreated patients (Table 1). Untreated patients also had higher comorbidity scores, were less likely to have insurance, and had lower median household incomes. A larger proportion of treated patients had an abnormal alanine aminotransferase result at baseline. After propensity-score adjustment, demographic and clinical characteristics were balanced between treatment groups (Table 1).

The final model with an autoregressive covariance pattern and time as a nominal variable showed the best data fit. Descriptive analyses indicated the following: (1) change in FIB4 was nonlinear and varied by individual; and (2) treatment impact was time-dependent (treatment-by-time interaction with P value < .0001) (Table 2). The basic model with an autoregressive covariance pattern was selected for further examination, given evidence of superior fit. Race, sex, baseline FIB4 score, and hepatitis genotype were retained in the final model, resulting in improved fit (BIC 18,442) compared with the unadjusted model (BIC 21,125). Sensitivity analysis showed a similar trend in treatment effects and post-treatment group-by-time interaction (detailed later).

Figure 1 shows 4 representative trajectories from our sample. Patient A did not achieve SVR until completion of a third course of therapy, approximately 5 years after the first treatment initiation. Patient B achieved SVR after the first course of treatment. Notably, FIB4 decreased after SVR in both patients, consistent with the findings of the French national prospective cohort of patients co-infected with HIV and HCV (ANRS CO13 HEPAVIH) Cohort investigators.¹² Furthermore, this reduction continued long term, suggesting possible regression of fibrosis. Conversely, patients C (untreated) and D (treatment failure) showed upward trends in FIB4 score.

Figure 2A presents the observed mean trajectories of FIB4 (in log scale) by treatment group. A cut-off value of 1.81 (0.59 after log transformation) is presented as a reference; this cut-off value was validated previously in this cohort to predict advanced fibrosis.²² At baseline, 1657 patients had received treatment and 3074 patients were untreated. The first SVR occurred at interval 3 (24 weeks after treatment initiation) in 68 patients. At the end of the follow-up period, 755 treated patients had achieved SVR.

Figure 2B presents the final model estimates of FIB4 trajectories by treatment group (Table 2), which showed patterns similar to those of the observed data. As shown in Figure 2B, predicted FIB4 in patients who achieved SVR (Figure 2B, uneven dashed line) can be broken up into 3 phases: (1) decrease; (2) plateau; and (3) increase. FIB4 started higher than 1.81 and sharply decreased, remaining lower than 1.81 from years 2 to 5, when a gradual increase began. At 77 weeks after treatment initiation, this decrease became significantly different from baseline (0.25; 22% on original scale; P < .0001). We noted that a few patients showed a late increase, likely because missing FIB4 values resulted in a wide confidence interval for the predicted results. FIB4 in the SVR group was consistently lower over time than in the untreated or TF groups (P < .05). In untreated patients, FIB4 gradually increased; in year 2, the increase became statistically significant (increased by 11%; P = .013). By year 5, FIB4 had increased by 44% (P < .0001) compared with baseline.

Sensitivity analyses showed similar effects of treatment on APRI trajectories (Figure 3). APRI remained significantly lower in SVR patients than in untreated or TF patients. Similar treatment trajectories were observed using the last treatment initiation date as the index date, with follow-up evaluation limited to 6 years (data not shown). We also observed a consistent treatment effect in our 1:1 matched cohort (Supplementary Figure 1), as well as in analyses that omitted several key covariates for the calculation of propensity scores (Supplementary Figure 2).

Across time (Table 2), men showed consistently higher FIB4 levels than women (coefficient = 0.05; P < .001), and black patients presented with lower FIB4 than white patients (coefficient = −0.11; P < .001). GT 1 and 3 patients had higher FIB4 scores than GT 2 patients; although GT 3 patients showed lower FIB4 than GT 1 patients, the difference was not significant (coefficient = 0.04; P = .19). We noted that patients with lower baseline FIB4 scores (<1.81) tended to remain low over time (P < .001). Age remained an important risk factor; older age at baseline was associated with a higher FIB4 over time. No qualitative risk factor-by-treatment interaction was detected, indicating that risk factor effects were consistent among the treatment groups.

In Table 2, patients’ FIB4 trajectories were classified into subgroups based on combinations of risk factors and treatment status, which is shown in Figure 4. FIB4 growth patterns among treatment groups were similar to those described in Figure 2B. Older patients (>60 y) showed consistently higher FIB4 than younger patients (≤40 y). Interestingly, white male patients with HCV GTs 1 or 3 showed the highest FIB4 levels in each of the 3 treatment groups, whereas black female patients with HCV GT 2 had the lowest FIB4 levels.