Predicting Risk of End-Stage Liver Disease in Antiretroviral-Treated Human Immunodeficiency Virus/Hepatitis C Virus-Coinfected Patients

Study Design and Data Source

We conducted a retrospective cohort study among ART-treated HIV/HCV-coinfected patients in the Veterans Aging Cohort Study (VACS) between January 1, 1997 and September 30, 2010 [9]. The Veterans Aging Cohort Study consists of electronic medical record data from HIV-infected patients receiving care at Veterans Affairs (VA) medical facilities across the United States. Data include demographic characteristics, hospital and outpatient diagnoses (recorded using International Classification of Diseases, Ninth Revision [ICD-9] codes), procedures, laboratory results, and dispensed medications. Clinically confirmed cancer diagnoses are identified using the VA Central Cancer Registry. Date of death was ascertained from the VA Vital Status file [10]. The underlying cause of death was determined by linkage with the National Death Index (NDI). [11] This study was approved by the Institutional Review Boards of the University of Pennsylvania and Philadelphia VA Medical Center.

Study Patients

Coinfected patients were included if they had: (1) HIV (identified using a validated algorithm [9]), (2) detectable HCV RNA, and (3) received ART, defined as use of at least 3 antiretrovirals from 2 different classes [12], for at least 12 months. We included only coinfected patients on ART because current guidelines recommend ART in all HIV/HCV patients regardless of CD4 count [12]. The majority of Veterans receive their medications from VA pharmacies [13, 14], and prior studies have shown the validity of VA dispensing data as a surrogate of actual adherence [15–18]. Patients were excluded if prior to start of follow-up (defined below), they had ESLD or received interferon-based therapy (because such treatment can affect noninvasive markers of hepatic fibrosis and reduces risk of ESLD [19]).

Follow-up began after 12 months on ART. The 12 months before the start of follow-up represented the baseline period, during which predictors were ascertained. Follow-up continued until study endpoint, death, initiation of HCV therapy, or last visit before September 30, 2010.

Main Study Outcome

The primary endpoint was ESLD, defined as hepatic decompensation, hepatocellular carcinoma, or liver-related death. Hepatic decompensation was determined by 1 hospital discharge ICD-9 diagnosis or 2 or more outpatient ICD-9 diagnoses of ascites, spontaneous bacterial peritonitis, or esophageal variceal hemorrhage (Supplementary Appendix 1). The requirement of 2 outpatient diagnoses aimed to exclude events that were suspected, but not subsequently confirmed, at follow-up visits. We validated this definition in a prior survey, with 91% of events confirmed by medical records [20]. Based on the results of the prior validation study [20], we did not include ICD-9 diagnoses for hepatic encephalopathy or jaundice, which could indicate decompensation, because these diagnoses were often linked to unrelated conditions (e.g., narcotic overuse, stroke recorded as encephalopathy; biliary obstruction, atazanavir-associated hyperbilirubinemia recorded as jaundice). The date of decompensation was defined as the hospital discharge date (if identified by hospital diagnosis) or initial outpatient diagnosis date (if identified by outpatient diagnoses). Hepatocellular carcinoma was determined using the VA Central Cancer Registry, which confirms diagnoses by histology, cytology, or consistent radiography. Deaths were considered liver-related if the underlying cause of death in the NDI was hepatic decompensation, liver cancer, alcoholic liver disease, viral hepatitis, or nonalcoholic liver disease [3]. The date of ESLD was defined as the initial date of hepatic decompensation, hepatocellular carcinoma, or death from liver disease, whichever occurred first.

Data Collection

Baseline variables evaluated as predictors of ESLD included the following: FIB-4 (<1.45; 1.45–3.25; >3.25), APRI (<1.0; 1.0–2.0; >2.0), age, body mass index ([BMI] <18.5; 18.5–24.9; 25.0–30.0; >30.0 kg/m²), severe anemia (hemoglobin <10 g/dL), CD4 cell count <200 cells/mm³, HIV RNA >400 copies/mL, alcohol dependence/abuse, diabetes mellitus, hepatitis B coinfection, and nonblack race, which have been shown to be risk factors for advanced liver disease in coinfected patients [3, 19, 21]. FIB-4 is a measure of liver fibrosis stage that was developed and validated compared with liver biopsy in HIV/HCV-coinfected patients [22]. It is calculated from age, alanine aminotransferase (ALT), aspartate aminotransferase (AST), and platelet count: FIB-4 = (age [years] × AST [U/L])/((platelet count [10⁹/L]) × (ALT [U/L])^1/2). FIB-4 <1.45 identifies no/minimal FIB, whereas results >3.25 indicate advanced hepatic FIB/cirrhosis [22]. APRI, another noninvasive measure of hepatic fibrosis, uses AST and platelet count: APRI = ([AST [U/L]/upper limit of normal [considered as 40 U/L])/platelet count (10⁹/L)] × 100 [23]. An APRI >2.0 identifies cirrhosis in chronic HCV, and APRI <0.5 identifies no or minimal FIB [23]. We collected ALT, AST, platelet count, body weight, hemoglobin, CD4 count, and HIV RNA results measured closest to, but prior to, the start of follow-up. Alcohol dependence/abuse was defined by previously validated ICD-9 diagnoses (Supplementary Appendix 1) [24]. Diabetes was defined by random glucose ≥200 mg/dL, ICD-9 diagnosis, and/or antidiabetic medication use [25]. Hepatitis B coinfection was defined by a positive hepatitis B surface antigen. Additional variables collected at baseline for descriptive purposes included sex, HCV genotype, and HCV RNA.

Statistical Analysis

We used Cox regression to estimate the relative hazards of ESLD (with 95% confidence intervals) for the predictors of interest [26]. We also evaluated FIB-4, APRI, hemoglobin, CD4 count, and HIV RNA level as continuous variables.

Next, because the stage of hepatic fibrosis is a strong determinant of ESLD [3, 19] and an important factor in the decision to initiate HCV therapy [8], we initially created separate models comprising FIB-4 and APRI alone. We then developed 2 separate lines of prognostic models with FIB-4 and APRI as the base models, respectively. Various combinations of the other predictors were added to the base models, including an interaction term for race. Age is a component of FIB-4 and so was not included in FIB-4 models.

Each model's performance was assessed by statistical significance and discrimination, using the concordance statistic (c-statistic) as an index of the model's ability to separate persons who will develop ESLD from those who will not [27]. We favored simpler models with fewer predictors that had appreciably different c-statistics from those of FIB-4 or APRI-based models alone. Performance of promising models was further compared by decision curves [28, 29], a graphical approach for evaluating and comparing prognostic models that considers the inherent trade-offs of false-positive versus false-negative classifications arising out of varying risk thresholds. Details appear in Supplementary Appendix 2. From the final model selected, we plotted Kaplan-Meier failure curves. Differences in ESLD by FIB-4 categories were estimated using log-rank tests.

To ensure that missing data did not affect the validity of our results, in a sensitivity analysis, we generated 10 imputed datasets, each with completed data on all predictors [30]. Data were analyzed using SAS 9.4 (SAS Institute Inc., Cary, NC) and Stata 13.1 (Stata Corporation, College Station, TX).

Patient Characteristics

Between 1997 and 2010, a total of 6016 ART-treated HIV/HCV-coinfected patients met our inclusion criteria (Figure 1). Absence of HCV RNA assessment was a common reason for exclusion (18%), but there were no clinically relevant differences in the characteristics between included patients and those excluded due to missing HCV RNA (data not shown).

The cohort (Table 1) was predominantly male and black, and one quarter of the patients had a history of alcohol dependence/abuse. Most patients were infected with HCV genotype 1 and had a high baseline HCV RNA level. The patients were followed for a median of 6.6 (interquartile range, 3.4–9.8) years. During follow-up, 482 (8.0%) patients initiated HCV therapy during and were censored. End-stage liver disease occurred in 532 (8.8%) patients during follow-up. At the time of the first ESLD event, a diagnosis indicative of hepatic decompensation was recorded in 397 (6.6%) patients, hepatocellular carcinoma was recorded in 63 (1.0%) patients, and liver-related death was recorded in 72 (1.2%) patients. Among the 397 patients with an incident hepatic decompensation event, 335 (84.4%) had ascites, 71 (17.9%) had spontaneous bacterial peritonitis, and 110 (27.7%) had variceal hemorrhage. Overall, 1936 (32.2%) patients died during follow-up, of which 164 (8.5%) were liver-related deaths (Supplementary Appendix 3).

Performance and Comparison of Prognostic Models for End-Stage Liver Disease

Baseline FIB-4, baseline APRI, and nonblack race were the strongest factors associated with ESLD among the variables evaluated in Cox regression analyses (Supplementary Appendix 4). Interactions with race were not observed for FIB-4 or APRI. Discrimination was similar for models with FIB-4 (c-statistic, 0.70) and APRI (c-statistic, 0.71) alone (Table 2). Alternative models comprising FIB-4 and race (c-statistic, 0.73) and APRI, age, and race (c-statistic, 0.74) had higher discrimination, but further addition of other combinations of predictors did not result in appreciably increased model discrimination (Table 2). Results of the decision curve analysis are shown in Supplementary Appendix 5. Because the model with FIB-4 + race had comparable discriminative ability and was simpler than the one with APRI + age + race, the model with FIB-4 + race was selected as the final model.

The risk of ESLD among non-blacks and blacks by category of FIB-4 is shown in Figure 2A and B, respectively (log-rank test among non-blacks, P < 0.001; log-rank test among blacks, P < 0.001). The risk at 5 years ranged from 2.6% to 26% among non-blacks (Figure 2A) and 2.0% to 14.0% among blacks (Figure 2B) across FIB-4 levels. Analyses using imputed data produced similar results.