Hepatocellular carcinoma prediction model performance decreases with long-term antiviral therapy in chronic hepatitis B patients

Abstract

Background/aims: Existing hepatocellular carcinoma (HCC) prediction models are derived mainly from pretreatment or early on-treatment parameters. We reassessed the dynamic changes in the performance of 17 HCC models in patients with chronic hepatitis B (CHB) during long-term antiviral therapy (AVT).

Methods: Among 987 CHB patients administered long-term entecavir therapy, 660 patients had 8 years of follow-up data. Model scores were calculated using on-treatment values at 2.5, 3, 3.5, 4, 4.5, and 5 years of AVT to predict threeyear HCC occurrence. Model performance was assessed with the area under the receiver operating curve (AUROC). The original model cutoffs to distinguish different levels of HCC risk were evaluated by the log-rank test.

Results: The AUROCs of the 17 HCC models varied from 0.51 to 0.78 when using on-treatment scores from years 2.5 to 5. Models with a cirrhosis variable showed numerically higher AUROCs (pooled at 0.65-0.73 for treated, untreated, or mixed treatment models) than models without (treated or mixed models: 0.61-0.68; untreated models: 0.51-0.59). Stratification into low, intermediate, and high-risk levels using the original cutoff values could no longer reflect the true HCC incidence using scores after 3.5 years of AVT for models without cirrhosis and after 4 years of AVT for models with cirrhosis.

Conclusion: The performance of existing HCC prediction models, especially models without the cirrhosis variable, decreased in CHB patients on long-term AVT. The optimization of existing models or the development of novel models for better HCC prediction during long-term AVT is warranted.

Keywords: Antiviral treatment; Carcinoma, hepatocellular; External validation; Hepatitis B, chronic; Prediction model.

Figure 1.

AUROCs of risk prediction modelscores at different on-treatment timepoints. The AUROCs demonstrated the predictability of model scores for the three-year development of HCC after each on-treatment timepoint. The dotted lines represent criteria generally accepted to judge the discrimination: less than 0.50 (red dotted line) indicates that the predictions are no better than chance; less than 0.60 (gray dotted line) reflects poor discrimination; 0.60 to 0.75 (green dotted line), indicates possibly helpful discrimination; and greater than 0.75, indicates clearly useful discrimination. During long-term AVT, the AUROCs were poor for untreated models without the cirrhosis variable (A), were possibly helpful for treated or mixed models without the cirrhosis variable (C), and were numerically higher for models with the cirrhosis variable derived from treated, mixed, or untreated CHB patients (B and D) compared with other models (A and C). AUROC, area under receiver operating curve; HCC, hepatocellular carcinoma; CHB, chronic hepatitis B; AVT, antiviral therapy.

Figure 2.

Hazard ratios of risk prediction modelscores at different on-treatment timepoints. The hazard ratio (HR) estimates demonstrate the magnitude of increase in three-year hepatocellular carcinoma (HCC) risks associated with every 10% increase in model scores at each on-treatment timepoint. The 95% confidence interval (CI) covering the value of 1.0 demonstrated a nonsignificant correlation of on-treatment scores with HCC incidence. The HRs were nonsignificant at either timepoint for untreated models without the cirrhosis variable (A), became nonsignificant after antiviral therapy (AVT) year 3.5 for treated or mixed models without the cirrhosis variable (C), and remained significant until AVT year 5 for most models with cirrhosis as a variable derived from treated, mixed, or untreated CHB patients (B and D). For all models, the HR estimateslessened over time.

Figure 3.

Cumulative three-year HCC incidence by risk group stratified by on-treatment model scores using original cutoffs. At each ontreatment timepoint, the subsequent three-year HCC incidence for high-risk (red bars), intermediate-risk (yellow bars), and low-risk categories (green bars) classified using on-treatment model scores were calculated for each model. The differences in HCC incidence between the high-risk and intermediate-risk groups gradually diminished with prolonged AVT. With the original cutoffs, the true HCC incidence across low-, intermediate-, and high-risk levels became non-significant using scores after AVT year 3.5 for untreated models with the cirrhosis variable (A) and treated or mixed models without cirrhosis (B), and became non-significant using scores after AVT year 4 for treated or mixed models with cirrhosis (C). HCC, hepatocellular carcinoma; AVT, antiviral therapy.