Sage Journals: Discover world-class research

Abstract

Objectives

The aim of this study was to develop and internally validate a hepatocellular carcinoma (HCC) risk prediction model incorporating repeated-measures data (longitudinal model), and compare with baseline predictions.

Methods

A total of 1097 participants with chronic hepatitis C after direct-acting antivirals (DAA) treatment were included in this prospective cohort study. The framework of joint models for longitudinal and survival data was used to construct the longitudinal prediction model. For comparison, a baseline model incorporating the same predictors was constructed through the multivariate Cox regression models. Model performance was evaluated using dynamic discrimination index (DDI), areas under the receiver-operating characteristics curves (AUROC), and Brier scores.

Results

Over a median follow-up of 7.25 years, 60 patients (5.5%) developed HCC. Key risk factors identified were aspartate aminotransferase (AST), cholinesterase, gamma-glutamyl transferase (GGT), albumin, hemoglobin (Hb), platelet count, alpha-fetoprotein (AFP), antigen-125 (CA-125), and carcinoembryonic antigen (CEA). The final joint model, with GGT and CEA removed, showed superior average predictive performance (DDI = .871) compared to models with all predictors included. Validation showed high predictive accuracy for HCC, with AUROCs above .9 for 1-, 3-, 4-, and 5-year predictions. In comparison, the baseline Cox model only achieved mediocre AUROCs of .7 (.75, .67, .69, and .67, respectively).

Conclusion

Compared to static models, our dynamic prediction model can predict the risk of HCC in patients after DAA treatment more accurately, providing better information to distinguish high-risk populations.

Keywords

hepatocellular carcinoma direct-acting antivirals prediction joint model hepatitis

Introduction

Hepatitis C virus (HCV) infection is a significant public health concern globally, with approximately 56.8 million viraemic HCV infections at the beginning of 2020 worldwide.¹ HCV infection can demonstrate both acute and chronic disease courses, with approximately 70.00% of individuals developing chronic HCV infection.² Hepatocellular carcinoma (HCC) is a critical clinical outcome for patients with chronic HCV infection, causing poor outcomes and heavy health care burden.³ The development of direct-acting antivirals (DAAs) has altered the scenario of HCV-related HCC and extensive independent studies confirmed that a high proportion of the patients achieved sustained virologic response (SVR) after DAA regime and exhibited a significantly decreased risk of HCC.⁴

Although DAA treatment regimens make it possible to achieve HCV elimination, the residual risk of HCC still exists in cases of HCV eradication.⁵ Identifying the subsets of post-DAA treatment patients with high HCC risk would be of great help for the clinicians to promote proactive and intensified monitoring and management for them. Several previous studies have identified the risk factors of HCC among individuals with chronic hepatitis C infection (eg, age, liver fibrosis, platelet count, and type 2 diabetes), and developed risk prediction models accordingly.⁶ Despite prior endeavors, most of them only include baseline predictors collected at one time-point, which may neglect the dynamic nature of HCC risk among these patients due to the modifications of certain risk factors over time. Longitudinal models with repeated measurements of the potential predictors may be capable of accurately capturing the fluctuated risk of HCC risk among patients who received DAAs.⁷ Therefore, the current study aimed to construct a longitudinal predictive model to predict the long-term risk of HCC onset among post-DAA treatment patients with chronic hepatitis C infection and compare the performance of the longitudinal model with a conventional baseline model.

Material and Methods

Patients and Follow-Up

The current prospective cohort study is a part of the “Chronic Hepatitis C Research Program of Jiangsu” (CHCRPJ) project. Participants with chronic hepatitis C were recruited in Jurong People’s Hospital between 1/1/2012 and 31/12/2023. Chronic Hepatitis C can be diagnosed using serology tests (antibody) or molecular (presence of viral RNA or DNA) diagnostics. Participants with missing data on required serum biomarkers were omitted. Participants with prior or baseline (within 6 months from the DAA treatment initiation) HCC events were further excluded. Eventually, a total of 1097 participants were included in this study. The participant inclusion process is briefly demonstrated in Figure 1.

Figure 1.

Flow chart. HCC, hepatocellular carcinoma.

The index date of the study was defined as the date of receiving DAA regimen. Patients were followed until the study outcome, death or 31/12/2023, whichever came first. The study outcome was new onset of HCC after the index date. Information on HCC occurrence before and after treatment was obtained from hospital inpatient and outpatient diagnoses. Laboratory tests were carried out at the date of receiving DAA regimen and annually after that when patients returned for their follow-up visit.

Chronic hepatitis C was defined as being positive for both anti-HCV and HCV RNA.⁸ HCC was diagnosed according to the American Association for the Study of Liver Diseases (AASLD) guidelines.⁹

This study was conducted in line with the Declaration of Helsinki. The study protocol was approved by the Institutional Review Board of Nanjing Medical University (Approval No. 95/2014, 445/2017, 528/2019, and 939/2022), and written informed consent was obtained from all the study participants.

Selection of Predictors

The potential predictors utilized in model development were selected based on their availability and associations with HCC as described in previous literature.^10-12 The predictors were further classified into 2 categories: baseline predictors and longitudinal predictors.

Baseline factors, including gender, age, use of ribavirin, smoking status, alcohol use, and hypertension, were collected at enrollment and did not change over time. Univariate Cox regression analysis was used to estimate the effects of baseline factors on the risk of HCC occurrence. Only those baseline factors that exhibited statistical significance in the univariate analysis were included in the final survival predictors.

The longitudinal factors, which might change over time, were collected at enrollment and repeatedly measured during follow-up visits. These factors included serum biomarkers such as total bilirubin (TBIL), direct bilirubin (DBIL), alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), cholinesterase, gamma-glutamyl transferase (GGT), albumin, hemoglobin (Hb), platelet count, alpha-fetoprotein (AFP), urea, creatinine, total bile acid (TBA), triglycerides (TG), total cholesterol (T-Chol), high-density lipoprotein-cholesterol (HDL-C), low-density lipoprotein-cholesterol (LDL-C), apolipoprotein A1 (ApoA1), apolipoprotein B (ApoB), carbohydrate antigen-199 (CA-199), carbohydrate antigen-125 (CA-125), and carcinoembryonic antigen (CEA). Joint models were run separately for each longitudinal factor to estimate the impact of their dynamically changing on the risk of HCC occurrence. Only those longitudinal factors found to be statistically significant in the individual joint modelling analysis were included in the final longitudinal predictors.

Patients attended follow-up visits at variable time intervals. At each visit, their serum biomarkers, including the aforementioned longitudinal predictors, were measured. If any measurement from a follow-up visit was missing one of the longitudinal predictors, the entire data from that visit was excluded. Consequently, the time intervals between the repeated measurements for each patient were irregular.

Model Development

Two models, namely longitudinal and baseline models, were constructed to predict the occurrence of HCC among post-DAA patients with chronic hepatitis C. The performance of the above 2 models was further compared. The longitudinal model was constructed by the joint model framework, which consist of 2 linked sub-models: a survival sub-model and a longitudinal sub-model.

Survival sub-model was fitted for survival predictors. To be specific, we started by fitting a multivariate Cox proportional hazards regression model for the baseline predictors, which were found statistically significant in univariate Cox regression analysis. For the longitudinal sub-model, a non-linear mixed-effects model was fitted using the nlme R package, in which AST, GGT, Hb, AFP, CA-125 and CEA were involved (all variables were natural log-transformed). We included the main effect of time (time points that the corresponding longitudinal response were recorded) for the fixed-effects part, and we included an intercept and a time term for the random-effects part.

After having separate sub-models, we jointly modeled the longitudinal responses and time-to-event data under a maximum likelihood approach by using the R package JMbayes2, which fits joint models under a Bayesian approach using Markov chain Monte Carlo algorithms.

Joint modelling supports the inclusion of multiple longitudinal predictors to construct the model, but the model performance may not be optimal when all the longitudinal predictors are included in the joint model.¹³ Therefore, we additionally fitted the joint model with one or 2 longitudinal predictors removed and subsequently compared the model performance to screen for the best longitudinal model. A total of 1 + 8 + 28 different longitudinal models were fitted and compared. In addition, AFP has been previously identified as a biomarker of HCC, thus, we also fitted a joint model only containing AFP for comparison. Finally, we developed a baseline model using multivariate Cox proportional hazards regression models based on the same predictors as the best longitudinal model, but utilizing only a single measurement taken at baseline.

Model Performance and Internal Validation

The study participants were randomly split into a training set (70%) and a validation set (30%). Baseline and longitudinal models were developed on the training set and assessed on the validation set. Across our main model, the predictive performance of the models was evaluated in terms of both discrimination and calibration. Discrimination at 1, 2, 3, 4, and 5 years of the baseline and longitudinal models was assessed using time-dependent areas under the receiver-operating characteristics curves (AUROC).^14,15 Additionally, we used the dynamic discrimination index (DDI) to compare longitudinal model performance, which summarizes discrimination throughout the follow-up period.¹⁶ We calculated the DDIs with prognostic windows of 1, 2, and 3 years because most HCC occur within this period. Brier scores, which capture both discrimination and calibration, were used as a metric for overall calibration.¹⁴ Brier scores range between 0 and 1, with scores closer to 0 representing higher calibration and better model performance.

The joint model with the highest discrimination (as determined by the highest average DDI across all 3 prognostic windows) was chosen as the final longitudinal model. The prognostic performance of the final joint longitudinal model was then compared to a Cox model with the same predictors and joint model only with AFP using time-dependent AUC values.

Statistical Analysis

The reporting of this study conforms to TRIPOD guidelines.¹⁷

Continuous variables were presented as median (interquartile range, [IQR]), and categorical variables were presented as frequency (percentage). The follow-up time of patients was presented as median (range). All continuous variables were natural log-transformed and included in the model. The averaged trajectory of each longitudinal predictor was estimated using mixed-effects models with random and fixed effects for measurement time. All statistical analyses were two-sided, and statistical significance was set at P < .05.

All data analysis was performed using the statistical packages in R software (version: 4.3.3).¹⁸ The main statistical packages applied in the current study are as follows. Statistical significance was set at P < .05. The joint model was performed using the JMbayes2 package,¹⁹ and the Cox proportional hazards regression models was constructed using the Survival package.²⁰ The time-dependent AUROC was calculated using the RisksetROC package.²¹

Results

Baseline Characteristics of Patients

Since January 2012, a total of 1329 patients with hepatitis C underwent treatment at Jurong People’s Hospital. We excluded patients who were less than 6 months from treatment (n = 158), as well as patients with missing serum biomarkers data (n = 65) and patients who had HCC before treatment (including HCC within 6 months of the start of treatment, n = 9). The remaining 1097 patients who achieved SVR constituted the study population (Figure 1). Baseline characteristics of included participants are summarized in Table 1. The mean age of the study population was 57 years (IQR, 51-63), and of those, 23.3% were cirrhosis. In total 8 demographic characteristics and 24 laboratory results were collected.

Table 1.

Baseline Characteristics of Patients by Outcome Grouping.

Characteristics	All Patients (N = 1097)	Patients Without HCC (N = 1037)	Patients With HCC (N = 60)
Gender, female	807 (73.6)	767 (74.0)	40 (66.7)
Age (years)	57.00 (51.00, 63.00)	57.00 (51.00, 63.00)	60.00 (51.75, 66.25)
Cirrhosis	256 (23.3)	217 (20.9)	39 (65.0)
Use of ribavirin	433 (39.5)	408 (39.3)	25 (41.7)
Smoke	62 (5.7)	61 (5.9)	1 (1.7)
Use of alcohol	45 (4.1)	43 (4.1)	2 (3.3)
Hypertension	344 (31.4)	319 (30.8)	25 (41.7)
Total bilirubin (μmol/L)	15.70 (11.90, 20.70)	15.50 (11.80, 20.60)	17.55 (14.75, 23.25)
Direct bilirubin (μmol/L)	4.30 (2.80, 6.40)	4.20 (2.80, 6.30)	5.05 (3.40, 7.82)
ALT (U/L)	43.00 (25.00, 75.00)	42.00 (25.00, 73.00)	58.00 (28.75, 105.25)
AST (U/L)	38.00 (25.00, 62.00)	38.00 (25.00, 60.00)	61.50 (32.00, 84.50)
Cholinesterase (U/L)	6837.00 (5633.00, 7953.00)	6876.00 (5713.00, 8013.00)	5671.50 (4920.00, 7031.75)
ALP(U/L)	78.00 (62.00, 97.00)	78.00 (62.00, 97.00)	84.00 (69.75, 102.25)
GGT (U/L)	36.30 (22.70, 63.80)	35.60 (22.20, 62.10)	58.20 (33.10, 107.10)
Albumin (g/L)	45.10 (42.00, 47.50)	45.10 (42.10, 47.50)	44.35 (40.93, 46.73)
Hb (g/L)	131.00 (120.00, 141.00)	131.00 (121.00, 141.00)	127.00 (114.00, 142.00)
Platelet count (10⁹/L)	127.00 (90.00, 167.00)	128.00 (93.00, 169.00)	105.50 (67.75, 140.50)
AFP (ng/ml)	4.70 (2.76, 8.00)	4.60 (2.70, 7.84)	7.15 (3.60, 14.95)
Urea (mmol/L)	4.45 (3.59, 5.39)	4.45 (3.60, 5.38)	4.37 (3.42, 5.59)
Creatinine (μmol/L)	65.20 (56.90, 77.20)	64.70 (56.70, 77.00)	68.65 (60.40, 79.08)
TBA (mmol/L)	6.80 (3.40, 12.70)	6.70 (3.40, 12.40)	10.70 (4.32, 17.92)
TG (mmol/L)	1.20 (.91, 1.52)	1.20 (.90, 1.52)	1.24 (.95, 1.54)
T.Chol (mmol/L)	4.25 (3.74, 4.80)	4.25 (3.74, 4.81)	4.30 (3.79, 4.79)
HDL-C (mmol/L)	1.55 (1.33, 1.81)	1.55 (1.34, 1.81)	1.49 (1.31, 1.78)
LDL-C (mmol/L)	2.45 (2.09, 2.83)	2.46 (2.09, 2.83)	2.33 (2.05, 2.73)
ApoA1 (g/L)	1.63 (1.43, 1.82)	1.63 (1.43, 1.82)	1.57 (1.39, 1.75)
ApoB (g/L)	.73 (.59, .88)	.73 (.60, .88)	.72 (.59, .87)
CA-199 ( U/mL)	19.61 (13.59, 27.02)	19.41 (13.50, 26.89)	21.93 (15.73, 32.24)
CA-125 (U/ml)	11.71 (8.08, 16.37)	11.46 (8.03, 16.10)	14.39 (11.06, 24.74)
CEA (ng/mL)	2.24 (1.58, 3.04)	2.22 (1.57, 3.00)	2.55 (1.86, 3.60)
Cystatin c (mg/L)	.70 (.56, .87)	.70 (.56, .86)	.74 (.58, .91)
Median follow-up time (years) (range)	7.25 (.51, 12.45)	7.18 (.51, 12.45)	3.08 (.64, 9.9)

Continuous variables were presented as median (IQR), and categorical variables were presented as count (percentage). The follow-up time of patients was presented as median (range).

Abbreviations: ALT, alanine aminotransferase; AST, aspartate aminotransferase; ALP, alkaline phosphatase; GGT, gramma-glutamyl transferase; Hb, hemoglobin; AFP, alpha-fetoprotein; TBA, total bile acid; TG, triglycerides; T. Chol, total cholesterol; HDL-C, high-density lipoprotein cholesterol; LDL-C, low density lipoprotein cholesterol; ApoA1, apolipoprotein A1; ApoB, apolipoprotein B; CA-199, carbohydrate antigen-199; CA-125, carbohydrate antigen-125; CEA, carcinoembryonic antigen.

Incidence of HCC after DAA Treatment in HCV Patients

During a median follow-up of 7.25 years (range, 0.51-12.45), 60 patients (5.5%) developed HCC. HCC incidence rate was 7.22 per 1000 PY during the first year, gradually increasing to 10.94 per 1000 PY in the third year and then decreasing afterward. The cumulative incidence at 1, 2, 3, 4, 5, and 6 years after SVR were 0.3%, 1.4%, 2.5%, 3.5%, 4.0% and 4.5%, respectively. Throughout the follow-up period, HCC occurred as early as .64 years after treatment and as late as 9.9 years after treatment. Figure 2 shows the HCC occurred time and its distribution during the follow-up period.

Figure 2.

HCC occurred time and its distribution. HCC, hepatocellular carcinoma.

Trajectories of Longitudinal Factors Over Time

A total of 6361 sets of laboratory results were obtained at different time points from the 1097 patients. The longest time span between laboratory results was 12.45 years, during which time the average number of measurements was 5.8 (ranging from 1-40). All the laboratory result values were natural log-transformed to be analyzed as longitudinal factors. To demonstrate evolution of these factors over time, we plotted the individual and mean trajectories of the 6 longitudinal factors for the entire cohort of patients with and without HCC in Figure 3.

Figure 3.

Trajectories of 6 longitudinal predictors in patients with HCC and without from the entire cohort. The longitudinal predictors were AFP, AST, Platelet Count, Cholinesterase, HGB, and CA125. The grey lines represent individual trajectories of each patient, the blue lines are the averaged trajectories estimated using linear mixed-effects models The values of all predictor variables are on a log scale. Abbreviations: HCC, hepatocellular carcinoma; AFP, alpha-fetoprotein; AST, aspartate aminotransferase; HGB, hemoglobin; CA125, Carbohydrate antigen 125.

Dynamics of Longitudinal Factors Associated With the Risk of HCC

Joint models were run separately for each longitudinal factor to evaluate the dynamics of longitudinal predictors that might contribute to the development of HCC. The risk factors identified in the joint model analysis, including AST (HR = 5.71; 95% CI = 2.23 - 14.40) , Cholinesterase (HR = 0.18; 95% CI = 0.08 - 0.47) , GGT (HR = 3.35; 95% CI = 1.89 - 5.83) , Hb (HR = 0.025; 95% CI = 0.003 - 0.298) , Platelet Count (HR = 0.37; 95% CI = 0.20 - 0.73) , AFP (HR = 2.59; 95% CI = 1.94 - 3.55) , CA-125 (HR = 2.99; 95% CI = 1.66 - 5.24) , CEA (HR = 1.84; 95% CI = 1.29 - 2.62) were associated with a higher risk of HCC in patients after DAA treatment, as depicted in Table 2.

Table 2.

Factors Associated With HCC Risk in Patients in Joint Model.

Variables	Univariate
Variables	HR (95% CI)	P Value
AST (U/L)	5.71 (2.23-14.40)	<.001
Cholinesterase (U/L)	.18 (.08-.47)	<.001
GGT (U/L)	3.35 (1.89-5.83)	<.001
Hb (g/L)	.025 (.003-.298)	.002
Platelet count (10⁹/L)	.37 (.20-.73)	.002
AFP (ng/ml)	2.59 (1.94-3.55)	<.001
CA-125 (U/ml)	2.99 (1.66-5.24)	.001
CEA (ng/mL)	1.84 (1.29-2.62)	.002

Continuous variables were log transformed.

HR, hazard ratio; CI, confidence interval of HR; AST, aspartate aminotransferase; GGT, gramma-glutamyl transferase; Hb, hemoglobin; AFP, alpha-fetoprotein; CA-125, carbohydrate antigen-125; CEA, carcinoembryonic antigen.

Feature Reduction

HCC risk factors identified by separate joint model analyses were used as longitudinal predictors to construct longitudinal sub-models. The age and sex as baseline predictors and fitted survival sub-models through multivariate Cox regression models (Table 3). The DDI of the joint models incorporating different longitudinal predictors at years 1, 2 and 3 of the prediction window are shown in Figure 4. We found that in models with one longitudinal predictor removed, the best average discriminant performance was achieved when GGT was removed (DDI = .844). Whereas the performance was worst when AFP was removed (DDI = .727). For the model with both longitudinal predictors removed, the best discrimination performance was achieved when GGT and CEA were removed (DDI = .871), which was better than the model with all longitudinal predictors included (DDI = .808), and may be easier to use in clinical practice (ie, only 6 items instead of 8 need to be repeatedly assessed). We therefore chose the model containing 2 baseline predictors and 6 longitudinal predictors as the final model (Table 4).

Table 3.

Baseline Factors Associated With HCC Risk in Patients.

Variables	Univariate
Variables	HR (95% CI)	P Value
Gender, female	.53 (.29-.99)	.046
Age (years)	1.06 (1.02-1.10)	.002
Use of ribavirin	.76 (.41-1.42)	.390
Smoke	.39 (.05-2.82)	.349
Use of alcohol	.54 (.07-3.95)	.547
Hypertension	1.52 (.83-2.77)	.174

HR, hazard ratio; CI, confidence interval of HR.

Figure 4.

DDI for 1-year, 2-year and 3-year prognostic windows for longitudinal models with 1 predictor censored or 2 predictors censored. DDI, dynamic discrimination index; GGT, gramma-glutamyl transferase; AFP, alpha-fetoprotein; CEA, carcinoembryonic antigen.

Table 4.

Summary of Final Joint Model.

	Coef	P Value	HR	CI
Survival sub-model
Age	.0256	.5211	1.03	(.98, 1.07)
Gender	−.3087	.2671	.73	(.28, 1.88)
Longitudinal sub-model
AST (log transformed)	.7367	.1787	2.09	(.74, 6.14)
Cholinesterase (log transformed)	1.4755	.1293	4.37	(.67, 36.91)
Hb (log transformed)	−4.6368	.0067	.01	(.0002, .28)
Platelet count (log transformed)	−.4931	.2082	.61	(.27, 1.31)
AFP (log transformed)	.9603	.0000	2.61	(1.82, 3.78)
CA-125 (log transformed)	.6902	.0851	1.99	(.93, 4.24)

AST, aspartate aminotransferase; Hb, hemoglobin; AFP, alpha-fetoprotein; CA-125, carbohydrate antigen-125; CEA, carcinoembryonic antigen; Coef, regression coefficient; HR: hazard ratio; CI, 95% confidence interval of HR.

Performance of Prediction Models in the Validation Set

Validation of the models was performed on a random 30% split of the entire study cohort. The characteristics of patients in the 2 groups are displayed in Table 1. The validation set was not included in model development. Three years after enrollment, 255 out of 330 patients in the validation set were still at risk of HCC. In this subset of patients, the longitudinal model showed excellent performance in predicting HCC events that occurred 1 year, 3 year and 4 year, with AUROCs both above .9 (.90, .92, .95 and .93, respectively). For 2-year prediction, the performance of the longitudinal model was good as well, with AUROC .83. Furthermore, the longitudinal model achieved remarkably low brier scores in the 1-, 2-, 3-, 4- and 5- year predictions of HCC (.0138, .0203, .0144, .0099, and .011, respectively).

In comparison, the baseline Cox model was also constructed with multivariate cox proportional hazards regression models which only achieved mediocre AUROCs in predicting HCC events 3, 4, and 5 years (.67, .69, and .67, respectively) (Table 5). The longitudinal model outperformed the baseline Cox model with better discrimination performance. Additionally, given the significant contribution of AFP to the longitudinal model’s performance, we explored a longitudinal model solely based on AFP levels. Compared to the baseline model, this “AFP-only” model showed better results, especially in terms of 4- and 5-year predicted HCC, with an AUROC of .89. However, its 1-year, 2-year, 3-year, 4-year, and 5-year predictive performance was not as good as that of the longitudinal model. The comparison of the AUCs of the 3 models is demonstrated in Figure 5.

Table 5.

Comparison of the Performance Characteristics of the Longitudinal and Baseline Cox Models to Predict the Development of HCC.

	Longitudinal	AFP Only	Baseline Only
1-year prediction
AUROC	.90	.83	.75
Brier score	.0138	.0139	.009
2-year prediction
AUROC	.83	.73	.78
Brier score	.0203	.0205	.0120
3-year prediction
AUROC	.92	.72	.67
Brier score	.0144	.0146	.0233
4-year prediction
AUROC	.95	.89	.69
Brier score	.0099	.0103	.0329
5-year prediction
AUROC	.93	.81	.67
Brier score	.0110	.0115	.0396

AUROC, area under the receiver-operating characteristic curve.

Figure 5.

Area under the receiver operating characteristic curves value of the baseline model, AFP-only model and longitudinal model for predictions made 1, 2, 3, 4, and 5 years from year 1, year 2, year 3, year 4. Predictions were made at year 1 for HCC occurrence 1, 2, 3,4, and 5 years from year 1, which equals 2, 3, 4, 5, and 6 years from baseline. Predictions were made at year 2 for HCC occurrence 1, 2, 3,4, and 5 years from year 2, which equals 3, 4, 5, 6, and 7 years from baseline. Predictions were made at year 3 for HCC occurrence 1, 2, 3,4, and 5 years from year 3, which equals 4, 5, 6, 7, and 8 years from baseline. Predictions were made at year 4 for HCC occurrence 1, 2, 3,4, and 5 years from year 4, which equals 5, 6, 7, 8, and 9 years from baseline.

Discussion

This study was conducted to build a longitudinal predictive model to evaluate the long-term HCC risk among post-DAA treatment individuals with chronic hepatitis C infection. Three major findings were reported in the current study. Firstly, over the study follow-up period, different trajectories of AFP, AST, platelet count, cholinesterase, Hb, and CA-125 were observed between patients stratified by HCC development status. Secondly, several longitudinal factors were found to be associated with an elevated risk of HCC among post-DAA treatment patients with chronic hepatitis C, encompassing AST, cholinesterase, GGT, Hb, platelet count, AFP, CA-125, and CEA. Lastly, the longitudinal model with 6 factors exhibited better discrimination ability and calibration for identifying long-term HCC risk than the baseline and AFP-only models.

The wide use of the DAA regimen dramatically alters chronic hepatitis C management for its excellent HCV eradication ability, making a large proportion of patients achieving SVR foreseeable.^22-24 Despite this, residual risk of HCC occurrence and recurrence still existed among patients with chronic hepatitis C who have achieved SVR.^5,24 For example, a recent large cohort study conducted in the U.S. veterans recruited 22,500 patients with hepatitis C (19,518 of them achieved SVR) in which 271 HCC cases still occurred after DAA treatment completion, and HCC developed in 183 patients with SVR during 20,415 person per year follow-up.²⁵ HCC exhibited the worst 5-year survival probabilities of any cancer, detecting it at an early stage would greatly enhance the survival probabilities.²⁶ Thus, identifying post-DAA patients with chronic hepatitis C who needed intensified monitoring and management would be of great significance in reducing the risk of HCC onset and the subsequent poor outcome related to HCC. Although several HCC risk prediction models were constructed in previous studies, most of them only select predictive features at baseline alone, neglecting the dynamic modifications of the potential features that may cause the reduction of predictive accuracy and loss of information.^6,27,28 Therefore, constructing a risk prediction model with longitudinal features may benefit the prediction accuracy and add more information.

In this study, a total of 8 longitudinal features were identified as risk factors for HCC among post-DAA patients with chronic hepatitis C. Their trajectories were distinct among participants with HCC development and those without, indicating heterogeneous development patterns of these features may be associated with the onset of HCC.²⁹ Six of them were selected to construct the longitudinal prediction model. The final model exhibited excellent discrimination power for HCC risk as indicated by the high time-dependent AUROC values ranging from .83 to .95 and low Brier scores ranging from .0099 to .0203 within 5-year predictions in the validation set, which outperformed the baseline model. This result further highlights the necessity and importance of including the dynamic change of predictors (ie, predictors with longitudinal information) in the prediction of HCC risk among the target patient population. Of note, among the included longitudinal features, AFP had the most significant influence on the discrimination power of the model as indicated by the alternations of DII in the feature reduction process. AFP is a widely used HCC-screening biomarker whose expression is induced by liver stem/progenitor cells, and an increasing serum AFP level usually indicates the hepatocellular regeneration initiated by severe liver damage.^30-32 Given the significance of AFP to HCC risk, we also investigate the discrimination power of AFP alone, and compare it with the baseline and final longitudinal model with all features. Although the AFP-only model exhibited better discrimination ability when compared to the baseline model, its discrimination power was much lower than the final model.

The current study had several strengths. Firstly, the prospective nature of the study design ensures that data on predictors and outcomes were collected systematically and reduces recall bias. Secondly, the study incorporates longitudinal predictors, which allows for a more dynamic and accurate assessment of HCC risk by considering how predictor variables change over time.

However, several inevitable limitations of this study should be noted. Firstly, this study is conducted at a single center in Jiangsu, China, which may limit the generalizability of the findings to other populations or settings. Thus, external validations of the current model are necessary when the results are extrapolated to other populations or settings. Secondly, due to the small number of events, but many variables were considered. Therefore, there is a risk of overfitting.

Conclusion

This study developed and validated a longitudinal predictive model for HCC risk in chronic hepatitis C patients treated with DAAs, demonstrating superior accuracy by incorporating dynamic biomarkers. The model’s enhanced performance highlights the importance of longitudinal data for precise risk prediction. Future research should validate these findings in larger cohorts and explore clinical implementation to improve patient outcomes.

Supplemental Material

Supplemental Material - Dynamic Prediction of the Risk of Hepatocellular Carcinoma After DAA Treatment for Hepatitis C Patients

Supplemental Material for Dynamic Prediction of the Risk of Hepatocellular Carcinoma After DAA Treatment for Hepatitis C Patients by Xinyan Ma, Lili Huang, Meijie Yu, Rui Dong, Yifan Wang, Hongbo Chen, Rongbin Yu, Peng Huang and Jie Wang in Journal of Cancer Control.

Footnotes

Author Contributions

Conceptualization: Xinyan Ma, Jie Wang and Peng Huang; Data curation: Xinyan Ma, Lili Huang, and Meijie Yu; Formal analysis: Xinyan Ma and Lili Huang; Writing – original draft: Xinyan Ma and Rui Dong; Writing – review and editing: Yifan Wang, Rongbin Yu, Hongbo Chen, Peng Huang, and Jie Wang.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was sponsored by the Open Project of Jiangsu Health Development Research Center [grant number JSHD2022046].

Ethical Statement

ORCID iD

Jie Wang

Supplemental Material

Supplemental material for this article is available online.

Appendix

References

Polaris Observatory HCV Collaborators . Global change in hepatitis C virus prevalence and cascade of care between 2015 and 2020: a modelling study. Lancet Gastroenterol Hepatol. 2022;7(5):396-415.

Lingala

Ghany

. Natural history of hepatitis C. Gastroenterol Clin N Am. 2015;44(4):717-734.

Veracruz

Gish

Cheung

Chitnis

Wong

. Global incidence and mortality of hepatitis B and hepatitis C acute infections, cirrhosis and hepatocellular carcinoma from 2010 to 2019. J Viral Hepat. 2022;29(5):352-365.

Russo

Zanetto

Pinto

, et al.

Hepatocellular carcinoma in chronic viral hepatitis: where do we stand?

Int J Mol Sci. 2022;23(1):500.

Kanwal

Kramer

Asch

Chayanupatkul

Cao

El-Serag

. Risk of hepatocellular cancer in HCV patients treated with direct-acting antiviral agents. Gastroenterology. 2017;153(4):996-1005.

Semmler

Meyer

Kozbial

, et al. HCC risk stratification after cure of hepatitis C in patients with compensated advanced chronic liver disease. J Hepatol. 2022;76(4):812-821.

Zou

Yue

Jia

, et al. Accurate prediction of HCC risk after SVR in patients with hepatitis C cirrhosis based on longitudinal data. BMC Cancer. 2023;23(1):1147.

Maness

Riley

Studebaker

. Hepatitis C: diagnosis and management. Am Fam Physician. 2021;104(6):626-635.

Marrero

Kulik

Sirlin

, et al. Diagnosis, staging, and management of hepatocellular carcinoma: 2018 practice guidance by the American association for the study of liver diseases. Hepatology. 2018;68(2):723-750.

10.

Chen

Ding

, et al. Bile acids promote the development of HCC by activating inflammasome. Hepatol Commun. 2023;7(9):e0217.

11.

Fattovich

Stroffolini

Zagni

Donato

. Hepatocellular carcinoma in cirrhosis: incidence and risk factors. Gastroenterology. 2004;127(5 Suppl 1):S35-S50.

12.

Kaewdech

Sripongpun

Assawasuwannakit

, et al. FAIL-T (AFP, AST, tumor sIze, ALT, and Tumor number): a model to predict intermediate-stage HCC patients who are not good candidates for TACE. Front Med. 2023;10:1077842.

13.

Sahle

Owen

Chin

Reid

. Risk prediction models for incident heart failure: a systematic review of methodology and model performance. J Card Fail. 2017;23(9):680-687.

14.

Blanche

Proust-Lima

Loubère

Berr

Dartigues

Jacqmin-Gadda

. Quantifying and comparing dynamic predictive accuracy of joint models for longitudinal marker and time-to-event in presence of censoring and competing risks. Biometrics. 2015;71(1):102-113.

15.

. Quantifying and estimating the predictive accuracy for censored time-to-event data with competing risks. Stat Med. 2018;37(21):3106-3124.

16.

Rizopoulos

. Dynamic predictions and prospective accuracy in joint models for longitudinal and time-to-event data. Biometrics. 2011;67(3):819-829.

17.

Collins

Reitsma

Altman

Moons

. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMC Med. 2015;13:1. https://core.ac.uk/reader/81583807?utm_source=linkout. Accessed November 10, 2024.

18.

R Core Team . R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing; 2024.

19.

Rizopoulos

Papageorgiou

Miranda Afonso

. JMbayes2: extended joint models for longitudinal and time-to-event data. R package version 0.4-5. 2023.

20.

Therneau

. A package for survival analysis in R. R package version 3.5-8. 2024.

21.

Heagerty

Saha-Chaudhuri

. RisksetROC: riskset ROC curve estimation from censored survival data. R package version 1.0.4.1. 2022.

22.

Jacobson

Lawitz

Kwo

, et al. Safety and efficacy of elbasvir/grazoprevir in patients with hepatitis C virus infection and compensated cirrhosis: an integrated analysis. Gastroenterology. 2017;152(6):1372-1382.

23.

Llaneras

Riveiro-Barciela

Lens

, et al. Effectiveness and safety of sofosbuvir/velpatasvir/voxilaprevir in patients with chronic hepatitis C previously treated with DAAs. J Hepatol. 2019;71(4):666-672.

24.

Luo

Zhang

Wang

Shen

Che

. Eradication of hepatitis C virus (HCV) improves survival of hepatocellular carcinoma patients with active HCV infection - a real-world cohort study. Cancer Manag Res. 2020;12:5323-5330.

25.

Dash

Aydin

Widmer

Nayak

. Hepatocellular carcinoma mechanisms associated with chronic HCV infection and the impact of direct-acting antiviral treatment. J Hepatocell Carcinoma. 2020;7:45-76.

26.

Innes

Jepsen

McDonald

, et al. Performance of models to predict hepatocellular carcinoma risk among UK patients with cirrhosis and cured HCV infection. JHEP Rep. 2021;3(6):100384.

27.

Ioannou

Green

Beste

Mun

Kerr

Berry

. Development of models estimating the risk of hepatocellular carcinoma after antiviral treatment for hepatitis C. J Hepatol. 2018;69(5):1088-1098. doi:10.1016/j.jhep.2018.07.024

28.

Steinebrunner

Stein

Sandig

Bruckner

Stremmel

Pathil

. Predictors of functional benefit of hepatitis C therapy in a ‘real-life’ cohort. World J Gastroenterol. 2018;24(7):852-861. doi:10.3748/wjg.v24.i7.852

29.

Yang

Guo

, et al. A longitudinal study of AFP trajectories and clinical outcomes in intermediate-stage hepatocellular carcinoma after hepatectomy. J Hepatocell Carcinoma. 2024;11:219-228.

30.

Fausto

Campbell

Riehle

. Liver regeneration. Hepatology. 2006;43(2 Suppl 1):S45-S53.

31.

Katoonizadeh

Nevens

Verslype

Pirenne

Roskams

. Liver regeneration in acute severe liver impairment: a clinicopathological correlation study. Liver Int. 2006;26(10):1225-1233.

32.

Wang

Liu

, et al. Incidence, predictors and prognosis of genotype 4 hepatitis E related liver failure: a tertiary nested case-control study. Liver Int. 2019;39(12):2291-2300.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.28 MB