Nomogram for individualized prediction of hepatocellular carcinoma occurrence in hepatitis C virus cirrhosis (ANRS CO12 CirVir)

PATIENT SELECTION

Among patients with biopsy-proven compensated (Child-Pugh A) and uncomplicated virus-related cirrhosis who were enrolled in the French multicenter ANRS CO12 CirVir cohort between March 2006 and July 2012 and prospectively followed up for development of HCC,4 those with chronic infection by HCV (positive HCV serology whatever the level of viral replication) and without HBV coinfection (negative serum hepatitis B surface antigen [HBsAg]) were selected. If data on well-known predictive factors for HCC occurrence in patients with cirrhosis were not available, patients were secondarily excluded.

Preinclusion assessment included usual clinical and biological parameters (sex, age, body mass index [BMI], diabetes, arterial hypertension, dyslipidemia, metabolic syndrome [MetS], prothrombin time, bilirubin serum level, albumin serum level, platelet count, liver enzymes [aspartate aminotransferase {AST}, alanine aminotransferase {ALT}, and gamma-glutamyl transpeptidase {GGT}], alpha-fetoprotein (AFP) serum level, esophageal varices, HCV genotype, HCV viral load, human immunodeficiency virus [HIV] coinfection, and histological lesions of steatohepatitis [SH]) and a Doppler ultrasonography (US) to rule out complications at baseline, including HCC.

Patient information was recorded in a computerized database by a clinical research associate specifically dedicated to the ANRS CO12 CirVir cohort at each center. For all patients, past and ongoing alcohol or tobacco consumption was quantified and recorded at inclusion.

Excessive alcohol consumption was defined according to World Health Organization criteria (more than 2 glasses per day for females and more than 3 glasses per day for males); an overall minimal duration of 5 years was required.

Past medical history was also noted. In particular, the senior hepatologist in charge of a given patient at each center determined whether MetS was present or not, based on clinical guidelines14 and/or histological examination of liver biopsy.15

FOLLOW-UP

Follow-up was scheduled according to French guidelines (Haute Autorité de Santé). Examination by Doppler US was performed every 6 months. For a given patient, it was recommended that US be formed at the same center by an experienced operator. A report was completed by each operator, mentioning the presence or not of focal liver lesions.

All events occurring during follow-up, liver related or not, were recorded based on information obtained from medical files of patients at each center and monitored by a panel of three dedicated clinical research associates. All treatments, including antiviral therapy, were recorded at inclusion, and any modification during follow-up was notified.

HCC DIAGNOSIS AND TREATMENT

In case of focal liver lesions detected by US, the following procedures were carried out: (1) echogenicity, number and diameter of nodules (classified as <10, 11-20, 21-30, 31-50, or >50 mm), and anatomic localization according to the Couinaud classification were reported; (2) portal vasculature (main trunk and branches), hepatic veins, and vena cava were systematically examined; and (3) a diagnostic procedure using contrast-enhanced imaging (computed tomography scan or magnetic resonance imaging), serum AFP assay, and/or guided biopsy was performed according to 2005 American Association for the Study of Liver Diseases (AASLD) guidelines6 updated in 2011.16 HCC diagnosis was thus established either by histological examination by an experienced pathologist or using probabilistic noninvasive criteria (mainly dynamic imaging showing early arterial hypervascularization and portal washout) according to the different periods of time (before and after 2011). When the HCC diagnosis was established, treatment was decided using a multidisciplinary approach according to AASLD or European Association for the Study of the Liver guidelines for HCC.5, 6, 16 Reports of imaging techniques showing liver focal lesions were secondarily reviewed by the two senior hepatologists from institution 1 (authors V.B. and P.N.).

ANTIVIRAL TREATMENT AND VIRAL REPLICATION

From 2006 to January 2014 (before the introduction of second-generation direct antiviral agents [DAAs] in France), all antiviral therapies conducted were interferon (IFN) based according to international guidelines. Since January 2014, patients could also receive second-generation DAAs with or without Peg-IFN and RBV for 12 to 24 weeks as part of early access programs, or after market approval according to national and/or international guidelines. The primary efficacy outcome was a sustained virological response (SVR), defined as HCV RNA undetectable by qualitative polymerase chain reaction assay (<50 IU/mL) at the end of either a 24-week untreated follow-up period for IFN-based antiviral therapies without second-generation DAAs, or a 12-week untreated follow-up period for antiviral therapy, including second-generation DAAs.

STATISTICAL ANALYSES

To internally validate the final adjusted HCC algorithm, we used a split-sample approach to assess discrimination, calibration, and predictive ability of the model. We randomly divided the studied population of HCV-related Child-Pugh A cirrhosis into a development subset (two thirds, on which model parameter estimates and the corresponding model-based HCC probabilities were derived) and a validation subset (one third, to which model results were applied) and used our final model in these two samples.

Descriptive statistics are presented for the entire cohort and for the two subsets, as median (interquartile range; IQR) for continuous variables and as numbers (percentages) for categorical data. Comparisons of patient subsets were performed using the Student t test or Wilcoxon rank-sum test for continuous variables, and the chi-square or Fisher's exact test for categorical variables.

The cumulative incidence of HCC was estimated in the training cohort in a competing risks framework, where deaths free of HCC were considered as competing events. Because no difference was observed with Kaplan-Meier estimates, univariate and multivariate Cox proportional hazards models were used in the training cohort to assess factors associated with occurrence of HCC. In particular, the effect of an SVR was assessed in Cox regression as a time-dependent covariate, given that patients without SVR at inclusion could be (re)treated and achieve SVR during follow-up. For these patients, end of treatment was defined as time T0, given that patients with undetectable HCV RNA at that time were considered as having an SVR. All variables tested in univariate Cox analysis with a P value less than 0.20 were tested in multivariate Cox regression.

A predictive score was then constructed as a weighted sum of all independent predictive factors for the occurrence of HCC, identified in Cox multivariate analysis. The weight of each factor was given by a linear transformation (multiplied by 3) of the estimated coefficient in Cox regression analysis and rounded off to the nearest integer. Baseline HCC-free probabilities were also estimated from Cox multivariate analysis at 1, 3, and 5 years. Predicted risks for HCC were estimated by the following equation: , where was baseline HCC-free probability and was the inverse multiplicative coefficient used to build the HCC score. Predicted probabilities of HCC at 1, 3, and 5 years are presented for each level of the score as a nomogram.

Cumulative incidence curves according to the different risk groups of HCC determined from the score were built using a clock reset approach. Patients who switched from non-SVR to SVR status were censored at the time of SVR, which was the date of the end of treatment. The time of SVR was reset at time 0 for patients having SVR status during follow-up.

Comparison of cumulative incidence curves was assessed with univariate Cox regression analyses.

Discrimination was assessed by receiver operating curve (ROC) analysis. Overall predictive accuracy of the derived score for predicting the development of HCC at 1 and 3 years after inclusion was tested by estimating the area under the ROC curve (AUROC).

Calibration of the model was assessed by Harrell's C statistics and the correlation coefficient between predicted and observed probabilities of HCC.

The score was secondarily tested in the validation cohort.

Last, the score was reassessed after 12 months of follow-up.

All statistical analyses were performed using Stata software (version 13.0; StataCorp LP, College Station, TX). A P value of 0.05 or less was considered statistically significant.

TRAINING COHORT

Based on multivariate Cox regression analysis, 5 of 23 variables listed in Table 3 were independently associated with occurrence of HCC in the training cohort: age >50 years (hazard ratio [HR], 1.94; 95% confidence interval [CI], 1.16; 3.25; P = 0.012); past excessive alcohol intake (HR, 1.55; 95% CI, 1.02; 2.36; p = 0.041); low platelet count (<100 Giga/mm³: HR, 2.70; 95% CI, 1.62; 4.51; P < 0.001; [100; 150] Giga/mm³: HR, 1.87; 95% CI, 1.10; 3.18; P = 0.021); GGT above the upper limit of normal (ULN; (HR, 1.96; 95% CI, 1.11; 3.47; P = 0.021); and no occurrence of SVR during the study period (HR, 3.02; 95% CI, 1.67; 5.48; P < 0.001; Table 3). An 11-point risk score was derived from the estimated β-regression coefficients of these five variables, by linear transformation (multiplied by 3) of the coefficients and rounding off (Table 4). The population was stratified into three groups according to this scoring system, with HCC occurrence gradually increasing from 2% to 30.1% at 5 years for patients with the lowest (≤3; n = 117; 13.8%) and highest (≥8; n = 349; 40.6%) HCC risk scores, respectively (P < 0.001; Fig. 2A). AUROC for 1- and 3-year predictions after baseline was 0.678 [0.555-0.801] and 0.717 [0.661-0.773], respectively, in the training group. Implementation of this score after 12 months of follow-up, taking into account increasing age, new values of platelets and GGT, and potential recent SVR, gave an AUROC for the 1- and 3-year predictions of 0.678 [0.574-0.783] and 0.727 [0.655- 0.800], respectively, in this subset of patients still at risk of HCC.

In order to develop an individualized prediction of HCC, a nomogram derived from this score was also built, given the 1-, 3-, and 5-year probabilities of HCC for each level of the score (Fig. 3). The fit of the model was assessed by Harrell's C statistics at 0.72.

Calibration was also assessed for predicted HCC risk and observed risk, and showed a good correlation (r = 0.0.91; P < 0.001).

VALIDATION SET

The cumulative incidence of HCC risk was estimated by the Kaplan-Meier method according to the three groups of the score built from the training cohort. HCC occurrence gradually increased from 0% to 20.3% at 5 years, for patients with the lowest (≤3; n = 64; 14.7%) and highest (≥8; n = 165; 37.9%) HCC risk scores (P < 0.001; Fig. 2B).

In this group, the AUROC for the 3-year prediction of HCC was 0.736 [0.639-0.833], similar to that of the training group (P = 0.75 between AUCs). Correlation between observed and predicted HCC risk was estimated with a coefficient at 0.86 (P < 0.001).