WHAT IS ALREADY KNOWN ON THIS TOPICWHAT THIS STUDY ADDS

This longitudinal phase III biomarker validation study measured the performance of 39 potential HCC biomarkers in a large, well-defined large cohort of patients with cirrhosis. Our study has identified a panel of nine biomarkers which significantly improves the predictive ability of previously established risk stratification models containing demographic and clinical variables.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICYBackground

The relative risk of hepatocellular cancer (HCC) in patients with cirrhosis is considerably higher than in the general population. However, the absolute risk of HCC as expressed by annual incidence is only 1%–3%, and most patients with cirrhosis do not develop HCC.1–3 Furthermore, cirrhosis related to alcoholic and metabolic dysfunction-associated steatotic liver disease (MASLD) is becoming more common and has lower annual HCC rate than cirrhosis related to hepatitis C or B virus infections.1 4 Therefore, predicting the small subgroup of patients with cirrhosis who will develop HCC has become more difficult in contemporary clinical practice, and one that is not addressed by existing biomarkers and risk stratification tools.5

Risk stratification of HCC among patients with cirrhosis can allow for a more precise prognostication, targeted prevention and early detection that is different from the one-size-fits-all approach of current practices. For example, instead of HCC surveillance offered to all patients with cirrhosis, surveillance could be intensified or limited to those with high absolute HCC risk, while patients with low risk can be spared from the repetitive, costly and occasionally harmful routine testing. High-risk groups could also be targeted for enrolment in chemoprevention trials facilitating the feasibility and cost-effectiveness of these initiatives.6

To address the gap in HCC risk prediction in cirrhosis, we previously developed and validated an HCC risk stratification model, which contains demographic (age and sex), clinical (liver disease aetiology (hepatitis C virus [HCV] vs other aetiologies), alanine aminotransferase, platelet count, albumin, alpha-fetoprotein) and lifestyle features (alcohol use, smoking, body mass index (BMI)) (see online supplemental materials for more details). This predictive model demonstrated moderate discriminatory ability with area under the receiver operating characteristic curve (AUROC) of 0.75 (95% CI 0.65 to 0.85) and 0.77 (95% CI 0.71 to 0.83) for predicting HCC at 1-year and 2-year follow-up, respectively.7 Similarly, the aMAP score (age, male, albumin-bilirubin and platelets) was reported to achieve an overall C-index of 0.81 (95% CI 0.79 to 0.82)8 9 in patients with chronic hepatitis. These basic models are good foundations for HCC risk prediction in cirrhosis, but their current modest performance may not be sufficient for clinical use. We hypothesise that we can improve the predictive performance of the model by adding objective and reliable variables from multiple domains that reflect distinct aspects of cirrhosis progression to HCC.

Multiple biological pathways are affected in patients with cirrhosis who develop HCC.10 Several human studies have examined the role of biomarkers of metabolic dysfunction including insulin resistance (eg, proinsulin, insulin, insulin growth factor (IGF-1), insulin growth factor binding protein 1 (IGF-BP1)) and adipocytokines (eg, leptin, adiponectin, resistin and adipsin); inflammation including proinflammatory cytokines (eg, tumour necrosis factor (TNF), latent ttransforming growth factor beta 1 (TGF β1), interleukin-1 beta (IL-1β)), acute phase reactants (eg, C reactive protein) and anti-inflammatory cytokines (eg, several other interleukins). Several of these biomarkers were reported to be significantly associated with HCC,11–16 and therefore, can potentially be useful for HCC risk stratification or early detection in patients with cirrhosis. However, most of these studies were phase I or II biomarker studies (cross-sectional or case–control in design), examined one or few biomarkers and did not adjust for demographic and clinical confounders. A key next step in biomarker validation for HCC is to examine their performance in a phase III biomarker study including a well-defined cirrhosis population that is eligible for surveillance.17

In this study, we used samples and data from our longitudinal multicentre prospective cohort of patients with cirrhosis, the Texas HCC Consortium (THCCC) to examine the association of several circulating serum biomarkers that were previously linked with the risk of developing HCC. We also sought to develop a robust metabolic biomarker panel predictive of HCC and test the incremental value of this panel compared with our base HCC prediction model of demographic, clinical and lifestyle factors and the aMAP score.

MethodsStudy cohort

We used data from the THCCC. Adult patients with cirrhosis were prospectively recruited at eight liver clinics in four cities (Michael E. DeBakey VA Medical Center and Baylor St. Luke’s Medical Center in Houston; University of Texas Southwestern, Parkland Health & Hospital System and two clinical sites within Baylor Scott & White Research Institute in Dallas/Fort Worth; Doctor’s Hospital at Renaissance in McAllen and Texas Liver Institute in San Antonio). THCCC recruitment and follow-up began in December 2016 and are still ongoing.4 18 Blood samples, collected at each study visit, were allowed to clot and then centrifuged at the collecting site according to a standardised protocol. Serum aliquots were frozen locally at −80°C and batch shipped to the central laboratory for long-term storage. We analysed samples from 2266 unique subjects recruited between December 2016 and August 2021 who were followed through 31 December 2022. More details on recruitment, follow-up and data/specimen collection are provided in online supplemental file 1.

We followed the PRoBE (Prospective-specimen collection, retrospective-blinded-evaluation) guidance for conducting a phase III biomarker study,19 and the Transport Reporting of a Multivariate Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines for reporting our results.20 21 Cirrhosis diagnosis was based on predefined criteria for liver histology, radiology, or elastography, or serum biomarkers.7 Patients with uncontrolled hepatic decompensation, history of HCC or non-hepatic cancers were excluded. All participants had a negative liver imaging at baseline. Our primary outcome was incident HCC, defined as tumours occurring at least 1 month after index visit to minimise risk of prevalent HCC but we also conducted a sensitivity analysis excluding patients diagnosed with HCC within the first 6 months of follow-up. HCC was defined according to American Association for the Study of Liver Diseases (AASLD) criteria22 including histological or radiological diagnosis using characteristic appearance (arterial enhancement and delayed washout of contrast) on triple-phase CT or MRI (LI-RAD 5 or 4) that were reviewed in multidisciplinary tumour boards and treated as HCC.

Biomarkers and custom multiplex immunoassays

We reviewed the published literature through 2020 to identify biomarkers of inflammation, oncogeneosis or metabolic dysfunction reported to be significant or near significant predictors of HCC risk in human studies. Serum samples collected from patients with cirrhosis at the time of enrollment in THCCC and before HCC diagnosis were evaluated using the custom Fireplex platform for circulating levels of our analytes of interest (online supplemental table 1). The FirePlex platform uses bioinert cross-linked polyethylene glycol hydrogel particles. These hydrogels are composed of two regions that encode the identity of the target and one that determines the amount of target present. By quantifying the fluorescent intensity in the two encoding regions, the identity of a target was determined using an integrated FirePlex Analysis Workbench software http://www.abcam.com/FireflyAnalysisSoftware. We used commercial serum to initially screen and identify antibody pairs for each target of interest. We performed bioinformatics of cross-reactivity by screening analyte sequences for high homology against the human proteome. The specificity of the assays was confirmed within subpanels using commercial proteins. Based on our literature review, we initially identified 41 potential targets (online supplemental table 1) but excluded chemerin and Agouti-related protein (AgRP) because we were unable to identify antibody pair to confirm the specificity against these analytes. We then used 36 human serum samples (from patients with cirrhosis in THCCC) to evaluate assay performance including intra-assay and interassay coefficient variances, assay sensitivity and dynamic range for each analyte within our multiplex panels. Each sample was evaluated in duplicates and at 1:4, 1:200 and 1:200 000 dilutions. Separately, we assayed AFP L-3 and des-γ-carboxy prothrombin (DCP) using microchip capillary electrophoresis and liquid-phase binding assay on an uTASWako i30 Immunoanalyzer System (FujiFilm IVD HCU Lab in Richmond, Virginia, USA).

Statistical methods

We used two methods to select biomarker panels that are predictive of HCC occurrence: the stepwise regression and the least absolute shrinkage and selection operator (Lasso). The stepwise regression selects a panel from individual biomarkers based on their statistical significance with HCC in the model. For this analysis, we selected metabolic panels based on two p values cutoffs of <0.05 and <0.2. Lasso imposes penalty terms during model fitting to shrink the less contributive biomarkers with smaller coefficients, resulting in a reduced panel with biomarkers having non-zero coefficients; this approach not only enhances prediction accuracy but also mitigates the risk of overfitting. In alignment with the framework proposed by Riley et al,20 considering a targeted shrinkage of less than 10% as recommended, and working with a set of 36 variables, the sample size of our cohort was adequate to achieve an anticipated Cox-Snell R2 statistic of at least 0.15.

We constructed Cox proportional hazard models to calculate the C-index and the AUROC across various evaluation windows (ie, time-dependent AUROC). The C-index allowed us to evaluate the model discriminatory performance for HCC overall and time-dependent AUROC to evaluate the model performance for HCC at 1 year and 2 years of follow-up. We used the risk Regression package to estimate the time-dependent AUROC.23 The internal validation was conducted by bootstrapping, which provides nearly unbiased estimates of model performance compared with split-sample (eg, cross-validation) methods.24 Specifically, we constructed 200 bootstrap datasets by resampling the original data with replacement, thus effectively getting bias-corrected C-index and time-dependent AUROC to approximate how the panel will perform on future data. The calibration of the model was assessed visually using calibration curves. To estimate the association of the metabolic panels with HCC risk, using the original data, we calculated HRs with their 95% CIs from the Fine-Gray competing risk model that accounted for the competing risks of liver transplantation and death.25 The discriminatory ability was further examined by comparing the cumulative incidence function (CIF) curves between high-risk and low-risk groups, based on being above or below the median of derived risk scores, respectively.

The incremental predictive value of biomarker panel score was evaluated by comparing predictive performance to our previously validated base model. This model was developed using two prospective cirrhosis cohort studies including THCCC7 and externally validated in a separate prospective VA cohort. The computation of the risk index by the base model is: 0.0399×age+0.5617×gender(male=1)+0.9023log10 AFP+(−1.2631)×log10 platelets+(–0.3357)×log10 ALT+(–0.5859)×albumin+0.0252×BMI+0.2816×smoking(past/current=1)+0.2446alcohol(current heavy=1)+(0.3596)×alcohol(other=1)+0.4611aetiology(HCV active/cured=1). To do this, we used the bootstrap method to assess whether differences in model performance (C index, time-dependent AUROC) with and without metabolic panel scores were statistically significant.26

Last, we examined the clinical utility by estimating the number of cases that would be correctly reclassified from false positives using the base model to true negatives when adding the biomarker signature to the model.23

We performed several secondary analyses to estimate the C index and time-dependent AUROC in (1) two separate groups, stratified by underlying cirrhosis aetiology (viral hepatitis vs non-viral hepatitis); (2) after excluding patients with active HCV and (3) adding DCP, AFP L-3 or both to the models containing base model combined with the biomarker panels. We also examined the incremental predictive value of adding biomarker panel score to another previously developed predictive model, the aMAP risk score, using the published equation and parameter estimates (aMAP risk score = [+7.4]/14.77×100 only, the average of 200 bootstrapping C-index is 0.651 (95% CI 0.647 to 0.655)).8

Results

We analysed data from 2266 patients with cirrhosis, of whom 126 developed HCC during follow-up of 39.9 months. The HCC annual incidence rate was 16.6 per 1000 person (95% CI 13.8 to 19.8). During follow-up, there were 257 deaths (11.3%), 59 (2.6%) liver transplant and 450 (19.9%) dropped out (diagnosis of non-HCC primary cancer, withdrawal of consent or a loss to follow-up). The proportion censored by 31 December 2022 was 60.6% (n=1374).

The median duration between the date of serum metabolic biomarker measurement and the HCC date was 24.9 months (IQR 26.6 months). The mean age of the cohort was 59.6 years (SD 10.3), and 37.6% were women. The racial/ethnic distribution was 1158 (51.1%) non-Hispanic white, 700 (30.9%) Hispanic, 360 (15.9%) African American and 48 (2.1%) belonged to other groups. At the time of enrolment, the underlying risk factors for cirrhosis were alcohol-related liver disease in 394 (17.4%), MASLD in 733 (32.3%), active HCV in 301 (13.3%), cured HCV in 534 (23.6%), HBV in 30 (1.3%) and miscellaneous causes in the rest. Most patients had diabetes (46.7%), overweight (82.9%) or obesity (51.2%). All 39 biomarkers were measured in all patients; however, we excluded free TGF-β1 from the analysis because of missing values in more than 75% of cases.

The panel selected by the stepwise selection regression with a p<0.05 included three biomarkers (IGF-1, latent TGF-β1 and IL-10), while the panel selected with a p<0.2 had nine biomarkers that were inclusive of the previous three (IGF 1, latent TGF β1, macrophage stimulating protein alpha chain (MSPa), TNF-α, adipsin, IL-1β, IL-10, serum amyloid A (SAA) and fetuin-A). The panel selected by the Lasso consisted of five-biomarker (IGF-1, latent TGF-β1, IL-10, IL-1β, TNF α), which included the three-biomarker panel described above and were all contained within the 9-biomarker panel identified based on the p<0.2 (online supplemental table 2). The nine-biomarker panel had the highest C-index of 0.736 for predicting HCC in Cox regression model compared with the three-biomarker and 5-biomarker panels, which were not significantly different from each other (C-index=0.700 and 0.703, respectively; data not shown). Therefore, we evaluated the three-biomarkers (P3) and nine-biomarker (P9) panels further.

There were significant differences in baseline serum levels of the biomarkers according to underlying aetiology of liver disease. Fetuin-A, latent TGF-β1, TNF-α were higher in active HCV, MSPa in cured HCV and IGF-1 and DCP in non-viral hepatitis than the other two groups (table 1). There were also significant differences in baseline levels of several biomarkers (TGF-α, IGF-1, MSP-α, AFP L-3) between women and men. Online supplemental table 2 shows the baseline serum levels of the nine biomarkers in P9 as well as DCP and AFP L-3 in patients with cirrhosis who subsequently developed HCC and those who did not as well as the univariate risk estimates, HRs, for the association between baseline levels and future risk of developing HCC. Compared with patients who did not progress, serum levels of IL-10 and TNF-α were higher whereas the rest of the seven biomarkers were lower in patients who progressed to HCC. Both AFP L-3 and DCP were higher at baseline in those who developed HCC.

Table 1

Baseline serum levels of P9 biomarkers as well as AFP L-3 and DCP in patients with cirrhosis stratified by underlying aetiology and gender

Using the internal validation dataset, we evaluated five models for predicting HCC risk in our primary analysis. These were (1) model using P3; (2) model using P9; (3) a model using demographic and clinical variables (base model); (4) model using demographic and clinical variables in the base model and P3 and (5) a model using demographic and clinical variables in the base model and P9. The C-index values of these five models were 0.655, 0.668, 0.715 0.731 and 0.732, respectively. Adding P3 or P9 significantly (both p<0.01) improved the C-index by 2.2% and 2.4%, respectively, over the base model. Adding P3 or P9 also significantly (both p<0.01) improved the 1-year and 2-year AUROCs (p<0.01) over the base model (table 2 and figure 1). The improvement in AUROC with addition of P3 and especially P9 to the base model was more prominent for the 1-year HCC risk.

Table 2

Performance characteristics of predictive models for HCC risk stratification in patients with cirrhosis

Figure 1

Area under receiver operating curve for HCC at follow-up of 1 year (A) and 2 years (B) shown for base HCC predictive model, base model plus P3 and base model plus P9. AUC, area under the curve; HCC, hepatocellular cancer.

Both P-9 and P3 were independently predictive of HCC risk. The nine-biomarker panel was associated with HCC risk independent of the base model with an adjusted HR of 2.06 (95% CI 1.63 to 2.61) as compared with 1.75 (95% CI 1.16 to 2.64) for the three-biomarker panel; p<0.05.

We obtained serum levels of AFP L3 and DCP in 1945 (85.83%) and 1943 (85.75%) patients, respectively (both markers were available in 1929 (86.6%) of the cohort). We added each as well as both biomarkers to our base model as well as the model with P9, and found no statistically significant change in C value or AUROC (table 3) after adding AFP-L3 and/or DCP.

Table 3

Performance characteristics of aMAP with and without biomarker panels

We also tested the predictive value of aMAP with and without P3 and P9. aMAP had a lower predictive value than our base model (0.651 vs 0.715). However, adding the P9 to aMAP increased the C index from 0.651 with aMAP alone to 0.693 with aMAP and P9 (p<0.01); a similar magnitude of increase to that observed in our primary analysis (tables 2 and 3).

The calibration of the models is shown in figure 2. The base model combined with P9 had better calibration (ie, the predicted HCC risk has higher similarity to the actual HCC risk) at all ranges of HCC risk than the base model alone or that of the base model combined with P3. The CIF curves of the base model combined with P9 for the low-risk and high-risk groups were compared in figure 3. The high-risk group consistently exhibited a significantly higher risk of HCC (p<0.01) right from the outset, highlighting the strong discriminatory capability of the model.

Figure 2

Calibration plots for HCC predictive models (predicted vs actual HCC) at 2 years of follow-up (A) base model with P3 and (B) base model with P9. HCC, hepatocellular cancer.

Figure 3

The cumulative incidence function of HCC stratified by the baseline average risk scores derived from a predictive model that combines clinical and demographic variables with the P9 biomarker panel. The groups are stratified as high (above the median score) and low risk (below median score) for HCC. HCC, hepatocellular cancer.

We examined the number of HCC cases and non-cases that would be correctly or incorrectly identified per 1000 patients in a 2-year screening programme under scenarios of desired ≥90% sensitivity or ≥90% specificity (table 4). Of the 70 HCC cases per 1000 patients in the study cohort, 63 are captured by a cut-point with ≥0.90 sensitivity. At this cut-point, there was a specificity of 31.7% (with 636 false positive cases) using the base model. The addition of P9, while maintaining a cut-point with ≥0.90 sensitivity, resulted in a specificity of 34.2% with 612 false positives (ie, correct reclassification of 4% of false positive cases). The base model at the ≥0.90 specificity cut-point achieves a sensitivity of 25.4% with 52 false negatives. Incorporating P9 into the model increased the specificity to 29.4% with 49 false negatives, while maintaining ≥0.90 specificity. This resulted in a correct reclassification of 6% of false negatives as true positive.

Table 4

Predictive value of HCC occurring within 2 years of follow-up

The stratified analyses by aetiology in patients with viral hepatitis (301 active HCV, 534 cured HCV, 30 with HBV) and 1401 patients with non-viral hepatitis showed comparable results in both groups. However, the predictive values of the base model as well as base combined with P3 or P9 models were 3.6%–4.1% higher in the non-viral hepatitis group compared with the viral hepatitis group (table 4). The base model performed slightly worse in women than men (c static 0.667 vs 0.716) but the base model combined with P9 improved the predictive value disproportionately for women; (0.733 vs 0.728) (table 5). Lastly, the results of sensitivity analyses excluding 301 patients with active HCV and 12 patients who developed HCC within the first 6 months of follow-up showed no differences from the primary analyses (online supplemental table 3).

Table 5

C-index for HCC predictive models

Discussion

In this longitudinal, we concomitantly evaluated 39 biomarkers that were previously linked to HCC and derived and internally validated a 9- biomarker signature that was predictive of HCC risk among patients with cirrhosis. These biomarkers were predictive of short as well as intermediate-term HCC risk both independent of as well as in addition to our previously developed base HCC prediction model with demographic, lifestyle and clinical variables. This study is a major step in developing an adaptive multidimensional risk model that contains variables from several domains reflecting distinct aspects of the pathogenetic processes that drive or promote oncogenesis in cirrhosis.

Our approach represents a valuable framework for biomarker evaluation in HCC risk stratification. We followed a structured two-step approach that entailed (1) using a previously validated base model and (2) measuring the marginal predictive value by adding the new biomarkers to the base model. For the first step, we used two separate models—a model developed and externally validated by our group as well as previously developed aMAP.8 It would have been easier and perhaps ‘more exciting’ to present the results from the nine-biomarker panel in isolation; however, it would have been inconsistent with clinical practice where new biomarkers are unlikely to replace simple, routinely available information. Our approach represents a more critical and pragmatic appraisal of the value of new biomarkers by comparing the gain in predictive ability relative to that offered by simple tests. We also demonstrate a robust and almost identical incremental predictive benefit of the P9 biomarker panel over aMAP, similar to the additional value conferred to our base risk predictive model.

Our study highlights that risk prediction and stratification is difficult, and just like other disorders (eg, predicting cancer in Barrett’s oesophagus),27 one expects relatively small increments in predictive power with the addition of new biomarkers to robust clinical models. For example, addition of AFP L-3 and DCP did not have any effect on the predictive ability of the model. The incremental gain was larger in the subgroup with non-viral hepatitis related cirrhosis, which may have clinical implications, especially if our results are confirmed in other cohorts of patient with non-viral hepatitis related cirrhosis. We expect that further progress in risk stratification may occur in relatively small steps, and likely come from adding biomarkers reflecting different biological domains (eg, genomics, radiomics, transcriptomics) in progression to HCC.28 A guiding principle is that variables that require additional cost and testing, need to significantly add to the base model to maintain the eventual applicability and cost-effectiveness.

We conducted the largest PRoBE study, phase 3 HCC biomarker testing design, which is ideal to move the findings of many phases 1 or 2 studies further down the validation path.17 We also conducted comprehensive and efficient testing of available biomarkers including AFP L-3 and DCP. Placing these biomarkers on FirePlex immunoassay has greatly facilitated the efficiency of simultaneous testing. We selected markers based on biological relevance, availability of suggestive evidence from biomarker studies in humans, as well as eventual clinical applicability criteria. Some of the selected biomarkers represented similar pathways or underlying metabolic dysfunction and they were highly correlated. Therefore, we used special statistical techniques that examined both the correlation as well as the predictive value for HCC. We created three different biomarker panels based on three modelling approaches (Lasso and stepwise regression using two different cutoffs). The findings were robust as well as biologically plausible: all three biomarkers (IGF-1, Latent TGF-β1 and IL-10)29–32 identified by stepwise regression with a p cut-off of 0.05 and are parts of known pathways of oncogenesis, tumour suppression and inflammation, respectively. The Lasso method contained in the five biomarkers selected and the latter were all contained in the nine biomarkers selected from stepwise regression with a p cut-off of 0.20 (IGF-1, latent TGF-β1, IL-10, MSPa, TNF-α, adipsin,33 34 IL-1β, SAA and fetuin-A).11 30 35–37 The nine-biomarker panel was associated with the highest discrimination model and calibration and the highest correct reclassification value. Therefore, we believe that the nine-biomarker panel is the most appropriate panel to be used as an independent HCC predictor or to be included in our adaptive modelling approach. Our results were robust in several subgroup and sensitivity analyses.

Our study has some limitations. While this is one of the larger prospective cohort studies of patients with cirrhosis, the number of outcome HCC events is small (n=126). This limitation also prohibited modelling each biomarker individually in the final HCC predictive model. It is possible that with larger sample size the composition or the number of the additional biomarkers will change. The incremental discriminatory value of the biomarker panel was modest, and therefore, it is not clear whether this enhanced model is cost-effective across different clinical scenarios and healthcare systems. It is possible that the modest performance could be boosted by incorporating changes over time, and while we have longitudinal samples, we have not performed the assays yet. Last, while we performed internal validation, external validation in other datasets is required.

In summary, in a phase 3 PRoBE study, we have developed a panel of nine circulating serum biomarkers that is significantly associated with HCC risk among patients with cirrhosis. This biomarker panel was independently predictive of HCC but also led to small although significant improvements in the predictive value compared with readily available base models of demographic, clinical and lifestyle variables. Our approach also represents a valuable framework for biomarker evaluation in HCC risk stratification.

Data availability statement

No data are available.

Ethics statementsPatient consent for publication

Not applicable.

Ethics approval

This study involves human participants and this research was approved by the Baylor College of Medicine IRB under ID number H-40241. The BCM IRB is organised, operates and is registered with the US Office for Human Research Protections according to the regulations codified in the United States Code of Federal Regulations at 45 CFR 46 and 21 CFR 56. The BCM IRB operates under the BCM Federal Wide Assurance No. 00000286, as well as those of hospitals and institutions affiliated with the College. Participants gave informed consent to participate in the study before taking part.