α-Fetoprotein model versus Milan criteria in predicting outcomes after hepatic resection for hepatocellular carcinoma: multicentre study

Abstract

Background

The Milan criteria and the French α-fetoprotein (AFP) model have both been validated for predicting outcomes after liver transplantation for hepatocellular carcinoma, with the Milan criteria also used for predicting outcomes after hepatic resection. The aim of this study was to evaluate the AFP model's predictive value for recurrence and survival following hepatocellular carcinoma resection and compare its performance with that of the Milan criteria.

Methods

Data for patients who underwent hepatocellular carcinoma resection between 2002 and 2021 were analysed. For both the AFP model and Milan criteria, patients were divided into two groups: those with hepatocellular carcinoma within and beyond the AFP model (scores ≤ 2 and > 2 points, respectively) and the Milan criteria. Cumulative recurrence and overall survival rates were compared between patients within and beyond the AFP model. Predictions of recurrence and overall survival by the AFP model and Milan criteria were compared using net reclassification improvement and area under the receiver operating characteristic curve analyses.

Results

Among 1968 patients evaluated, 1058 (53.8%) and 940 (47.8%) were classified as beyond on the AFP model and Milan criteria, respectively. After controlling for competing factors on multivariable analyses, being beyond the AFP model was independently associated with recurrence and worse overall survival after resection of hepatocellular carcinoma. Time-dependent net reclassification improvement and area under the receiver operating characteristic curve analyses demonstrated that the AFP model was superior to the Milan criteria in predicting recurrence. Of note, patients who were classified as beyond both the Milan criteria and AFP model had an even higher risk of postoperative recurrence and mortality (hazard ratios 1.51 and 1.47, respectively).

Conclusion

The French AFP model demonstrated superior prognostic accuracy to the Milan criteria in predicting recurrence and survival after hepatocellular carcinoma resection. The AFP model not only effectively stratified patient risk but also identified a subgroup of high-risk patients among those beyond the Milan criteria.

Introduction

Hepatocellular carcinoma (HCC) is a leading cause of cancer-related deaths worldwide^1–4, with hepatic resection often serving as the primary curative treatment due to limited donor organs^5,6. However, recurrence rates remain high (50–70%) within 5 years after surgery^6–10. The Milan criteria have been widely accepted for the selection of patients for liver transplantation in HCC¹¹ because they predict postoperative recurrence following liver transplantation or hepatic resection for HCC^12–19. However, in real-world clinical practice, using the Milan criteria could be challenging when treating patients with HCC^20,21, not least because of a lack of pretransplant α-fetoprotein (AFP) values^22–24. Thus, other criteria to select HCC patients for liver transplantation have emerged, including the AFP model, the Hangzhou criteria, and the Metroticket 2.0 model^{12–14,20,25,26}.

AFP is the most commonly used HCC biomarker. AFP is used for diagnosis, surveillance, measuring the efficacy of treatment, and prognostic evaluation^27–29. Among patients with HCC, high serum AFP concentrations have been associated with both an increased risk of recurrence and worse survival after liver transplantation^12–15 and hepatic resection^16–18. In 2012, a French study group²⁰ proposed the French AFP model for HCC transplant candidates. This model incorporates tumour burden (the number and size of nodules) and preoperative AFP levels (Table S1), with scores over 2 points considered ‘beyond’ the model and associated with a greater risk of recurrence after liver transplantation^20,30–32. The performance of the AFP model in predicting recurrence and survival after hepatic resection for HCC has not been thoroughly investigated.

Thus, the aim of the present study was to characterize the utility of the French AFP model in predicting recurrence and survival after hepatic resection for HCC, and to compare it to the Milan criteria.

Methods

Patient selection

This was a multicentre retrospective study conducted on patients who had undergone hepatic resection with curative intent for HCC between 2002 and 2021 at 11 Chinese hospitals. The study was approved by the institutional review board at each participating centre. The inclusion criteria were HCC diagnosis confirmed on resected specimens, curative hepatic resection (complete resection of all macroscopic tumours with microscopically clear resection margins (R0 resection) and no residual tumours on postoperative imaging within 4–6 weeks), no gross/macroscopic vascular invasion, and no previous anti-HCC treatment before resection. The exclusion criteria were age ≤ 18 years, recurrent HCC, ruptured HCC before surgery, palliative hepatic resection (R1) or R2 resection margins, residual tumours on imaging, and missing data for prognostic variables and/or follow-up data. Patients who died within 90 days after surgery or were lost to follow-up within 6 months were also excluded.

Informed consent was obtained from all patients. The study was performed according to the Declaration of Helsinki and the Ethical Guidelines for Clinical Studies of the participating centres.

Data collection

Data were collected on patient age, sex, American Society of Anesthesiologists grade, hepatitis B virus (HBV) status, preoperative Child–Pugh score, cirrhosis, preoperative serum AFP level within 1 week before surgery, largest tumour size, number of tumours, microvascular invasion, satellite nodules, tumour encapsulation, tumour differentiation, intraoperative blood loss, intraoperative blood transfusion, type of resection (anatomical or non-anatomical), extent of hepatectomy (major or minor), and resection margin status. A narrow resection margin was defined as a distance from the resection margin to the tumour edge of less than 1 cm. Tumour differentiation was defined according to the modified Edmondson criteria³³. Anatomical resections were classified based on the Brisbane 2000 Nomenclature of Liver Anatomy and Resections, whereas non-anatomical resections included wedge resection or limited resection³⁴. Major hepatectomy was defined as resection of three or more Couinaud liver segments, whereas minor hepatectomy was defined as resection of fewer than three liver segments³⁴.

Follow-up

Patients attended follow-up visits at 2-monthly intervals for the first 6 months, at 3-monthly intervals for the next 1.5 years, and then once every 6 months thereafter. Information regarding the patient’s history, physical examination, serum AFP level, and imaging results (ultrasonography, contrast-enhanced computed tomography (CT), or magnetic resonance imaging (MRI)) was collected at each follow-up visit. Enhanced CT or MRI, hepatic angiography, bone scans, or positron emission tomography were performed when recurrence or distant metastasis was suspected. Tumour recurrence was defined as the new appearance of intrahepatic or extrahepatic tumour nodule(s) with radiological characteristics similar to those of the primary tumour. Treatment of recurrence included hepatic resection, liver transplantation, local ablation, transcatheter arterial chemoembolization, radiotherapy, systemic therapy, or palliative care, as appropriate. The dates of recurrence, last follow-up, and death were recorded.

Division of patients into groups beyond and within the AFP model and Milan criteria

AFP scores were calculated using the simplified version of the AFP model²⁰. Patients with an AFP score of ≤ 2 points were classified as being ‘within’ the AFP model, and those with a score of > 2 points were classified as being ‘beyond’ the model. Patients were also classified as being ‘within’ and ‘beyond’ the Milan criteria.

Statistical analysis

The primary outcome of this study was cumulative recurrence, calculated from the date of surgery to the date when recurrence was diagnosed, or the date of the last follow-up for patients without recurrence. The secondary outcome was the overall survival (OS) rate, which was calculated from the date of surgery to a patient’s death, or the date of the last follow-up for surviving patients.

Clinical characteristics and operative variables are reported as frequencies and percentages for categorical variables and as either mean(standard deviation) or median (interquartile range [IQR]) for continuous variables. Categorical and continuous variables were compared using Chi-square test and the Wilcoxon rank-sum test, respectively. The Kaplan–Meier method was used to estimate cumulative recurrence and OS rates, which were compared between groups using the log rank test. Hazard ratios (HRs) with 95% confidence intervals for recurrence and OS were calculated using univariate and Cox regression multivariable analyses.

The prediction of recurrence was assessed and compared between the AFP model and Milan criteria using the time-dependent net reclassification improvement (NRI) method³⁵. The time-dependent area under the receiver operating characteristic curve (AUROC) was used to predict recurrence, as well as to compare the AFP model and Milan criteria³⁶. Two-sided P < 0.050 was considered statistically significant. Statistical analyses were performed using SPSS^® version 25.0 (IBM, Armonk, NY, USA) or R version 4.0.5 (R Foundation for Statistical Computing, Vienna, Austria).

Results

A flow chart of the study is shown in Fig. S1. Of the 1968 patients undergoing HCC resection, 1684 (85.6%) were men and 284 (14.4%) were women. The median age at the time of surgery was 52 years (range 19–86 years). Most patients had chronic HBV infection (1760, 89.4%), and cirrhosis (1345, 68.3%). Clinical and operative variables differed significantly between patients beyond (score > 2 points) and within (score ≤ 2 points) the AFP model (Table 1). Patients beyond the AFP model were younger and had less cirrhosis, higher AFP levels, larger tumours, and more aggressive tumour characteristics, including multiple tumours, satellite nodules, poor differentiation, and incomplete encapsulation, than those within the AFP model (all P < 0.001). The group beyond the AFP model had greater blood loss, higher transfusion rates, and a greater percentage of patients undergoing major hepatectomy and with narrow resection margins (all P < 0.001).

At a median follow-up of 54.3 months, 964 patients (49.0%) had developed HCC recurrence (intrahepatic: 703, 35.7%; extrahepatic: 84, 4.3%; both: 177, 9.0%). Among the patients with recurrence, 902 died (45.8%), with 676 cancer-related deaths (34.3%). For the overall cohort, the 1-, 3-, and 5-year recurrence rates were 18.9, 37.5, and 46.3%, respectively, and OS rates at 1, 3, and 5 years were 93.4, 75.9, and 64.0%, respectively.

Recurrence and OS in patients beyond and within the AFP model

As indicated in Table 2, the overall incidence of recurrence and death was higher among patients beyond (score > 2 points) than within (score ≤ 2 points) the AFP model (57.8 versus 38.8% for recurrence; and 55.0 versus 35.2% for death; both P < 0.001). There was no significant difference in the incidence of non-cancer-specific deaths between patients beyond and within the AFP model (11.6 versus 11.3%, respectively; P = 0.831). The incidence of cancer-specific deaths was markedly higher in the group beyond than within the AFP model (43.4 versus 23.8%; P < 0.001).

Comparisons of cumulative recurrence and OS between patients beyond and within the AFP model are shown in Fig. 1. The 1-, 3-, and 5-year recurrence rates among patients beyond the AFP model scores were 26.1, 47.9, and 57.4%, respectively, much higher than corresponding rates for patients within the AFP model (10.6, 25.8, and 33.8%, respectively; HR 2.01, 95% confidence interval (c.i.) 1.77 to 2.30; P < 0.001). OS rates at 1, 3, and 5 years among patients beyond the AFP model were 90.3, 66.7, and 52.7%, respectively, which were worse than for patients within the AFP model (96.9, 86.4, and 76.9%, respectively; HR 2.02, 1.77 to 2.32; P < 0.001).

Recurrence and OS in patients beyond and within the Milan criteria

Clinical characteristics, operative variables, and long-term recurrence and survival between patients within and beyond the Milan criteria were compared (Tables S2 and S3). The differences between the two groups were similar those seen for patients within and beyond the AFP model. The 1-, 3-, and 5-year recurrence rates among patients beyond the Milan criteria were 27.5, 49.6, and 58.7%, respectively, which were higher than those for patients within the Milan criteria (11.1, 26.7, and 35.2%, respectively; HR 2.00, 95% c.i. 1.75 to 2.27; P < 0.001). OS rates at 1, 3, and 5 years among patients beyond the Milan criteria were 89.1, 65.4, and 51.9%, respectively, and worse than corresponding values among patients within the Milan criteria (97.3, 85.4, and 75.0%, respectively; HR 1.97, 1.72 to 2.25; P < 0.001) (Fig. 2).

Univariable and multivariable analyses in predicting recurrence and OS

Univariable and multivariable Cox regression analysis identified risk factors associated with postoperative recurrence after hepatic resection for HCC (Table 3). Multivariable regression models examining the AFP model and Milan criteria separately that included variables comprising each assessment tool revealed that both the AFP model and Milan criteria were independently associated with increased risks of recurrence (AFP model: HR 1.62, 95% c.i. 1.39 to 1.88; Milan criteria: HR 1.43, 1.23 to 1.67). Similar results were found regarding OS (AFP model: HR 1.54, 1.30 to 1.83; Milan criteria: HR 1.52, 1.28 to 1.81) (Table 4).

NRI and AUROC values between the AFP model and Milan criteria

Compared with the Milan criteria, the AFP model had improved predictive power to risk classify 1-, 2-, 3-, 4-, and 5-year recurrence (NRI 0.023–0.071; 1- and 2-year recurrence, P < 0.050; and 3-, 4-, and 5-year recurrence, P > 0.050). The plot for time-dependent NRI showed that the AFP model maintained better predictive power than the Milan criteria for the first 5 years of follow-up. Of note, the AFP model was particularly better than the Milan criteria in predicting HCC recurrence with the first 2 years after surgery (Fig. 3a). Time-dependent AUROC curves for the AFP model and Milan criteria showed that the AFP model maintained higher AUROC values than the Milan criteria at 1, 2, 3, 4, and 5 years (0.669 versus 0.644, 0.655 versus 0.638, 0.639 versus 0.626, 0.631 versus 0.624, and 0.629 versus 0.622, respectively) (Fig. 3b).

Subgroup analysis of patients with HCC beyond the Milan criteria

Of 940 patients who were beyond the Milan criteria, 185 (19.7%) and 755 (80.3%) were within and beyond the AFP model, respectively. Figure S2 shows that patients who were beyond both the Milan criteria and the AFP model had significantly higher 1-, 3-, and 5-year recurrence rates than those who were beyond the Milan criteria but within the AFP model (29.7, 52.8, and 62.0% versus 18.7, 36.6, and 45.6%, respectively; HR 1.62, 95% c.i. 1.29 to 2.04; P < 0.001). In addition, 1-, 3-, and 5-year OS rates were worse in the former than latter group (88.9, 62.3, and 48.1% versus 90.3, 77.8, and 67.0%, respectively; HR 1.60, 1.26 to 2.03; P < 0.001).

The results of univariate and multivariable Cox regression analyses identifying risk factors associated with recurrence and OS after hepatic resection for HCC beyond the Milan criteria are presented in Tables S4 and S5. After hepatic resection, patients who were beyond the Milan criteria and beyond the AFP model exhibited increased recurrence (HR 1.51, 1.17 to 1.94; P = 0.002) and worse OS (HR 1.47, 1.03 to 1.93; P = 0.031) than those beyond the Milan criteria but within the AFP model.

Discussion

This multicentre study demonstrated that the AFP model outperforms the Milan criteria in predicting recurrence and survival after hepatic resection for HCC. Key findings include superior accuracy of the AFP model, especially for early recurrence within 2 years, and effective risk stratification, identifying high-risk subgroups among patients beyond the Milan criteria and poorer outcomes for patients exceeding both the AFP model and Milan criteria. The AFP model’s inclusion of serum levels of AFP, a biomarker of the aggressiveness of HCC, provides a more nuanced evaluation of tumour biology compared with the Milan criteria, which are based on tumour burden. The results of the present study validate the utility of the AFP model in hepatic resection and align with previous evidence from liver transplantation studies^20,30–32.

The AFP model has the advantage of being a simple, widely accepted scoring system. Although novel tools incorporating artificial intelligence or nomograms are promising³⁷, they often lack sufficient validation for clinical application. This study emphasizes the importance of validating established models like the AFP model, rather than developing new ones, to enhance their adoption in practice. Time-dependent AUROC analysis showed that both the AFP model and Milan criteria were more predictive of early recurrence within 2 years after surgery than late recurrence, in which tumour biology differs. Early recurrence is more likely the result of occult metastasis from the primary tumour and is more likely to be aggressive, whereas late recurrence more likely represents different clonal origins^38–40. Subgroup analysis highlighted the poor outcomes for patients exceeding both the AFP model and Milan criteria, with > 60% recurrence and < 50% survival rates at 5 years, compared with more favourable outcomes for patients within AFP model thresholds. These findings underscore the potential of the AFP model to refine risk stratification and guide adjuvant therapy trials for high-risk subgroups. Patients identified as being at high risk by both models may benefit from more intensive surveillance protocols and could be prioritized for adjuvant therapy trials^41–45. The AFP model has several clinical applications, including aiding preoperative decision-making, personalizing postoperative surveillance, and selecting candidates for adjuvant therapies^43,44,46,47.

The present study has several limitations that warrant discussion. Its retrospective nature introduced potential selection bias and limited the control for all confounding factors. The predominance of HBV-related HCC in this cohort, reflecting the epidemiology of HCC in China, may limit the generalizability of these findings (such as hepatitis C virus, alcohol, non-alcoholic steatohepatitis)^48–50. Further validation in diverse patient populations is necessary. Although the multicentre design of the enhances the generalizability of the results, the protocols across different centres could not be entirely standardized, which introduced potential variations in surgical techniques and perioperative management. Molecular markers or gene expression profiles were not included in this study, which could provide additional prognostic information beyond clinical and pathological factors¹⁰. Advancements in surgical techniques and adjuvant therapies may have influenced the results during the long follow-up in this study. In a subsequent study, comprehensive analysis will be conducted to explore temporal trends and potential variations in the clinical characteristics and management strategies for HCC. In addition to the Milan criteria, other predictive models and standards (such as the up-to-7 criteria) have significant clinical value. The Milan criteria remain the most widely recognized and used staging scoring system for liver transplantation worldwide. Although the present study demonstrated the prognostic value of the AFP model, it did not directly prove that its use improves patient outcomes; prospective studies evaluating the impact of AFP model-guided decision-making on long-term outcomes are needed.

Funding

This study was supported by grants from the National Natural Science Foundation of China (82425049, 82372813 and 82273074), Dawn Project Foundation of Shanghai (21SG36), Shanghai Health and Hygiene Discipline Leader Project (2022XD001), Shanghai Outstanding Academic Leader Programme (23XD1424900), the Natural Science Foundation of Shanghai (22ZR1477900), Shanghai Science and Technology Committee Rising-Star Programme (22QA1411600), Special Clinical Project of Shanghai Municipal Health Commission (20224Y0299), and National Natural Science Foundation Incubation Project (2022MS039 and 2022GZR005).

Author contributions

Chao Li (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Resources, Software, Writing—original draft), Yong-Kang Diao (Conceptualization, Data curation, Formal analysis, Methodology, Software), Yi-Fan Li (Conceptualization, Formal analysis, Investigation, Validation, Writing—original draft), Shao-Dong Lv (Conceptualization, Writing—review & editing), Xian-Ming Wang (Conceptualization, Formal analysis, Project administration, Resources, Visualization, Writing—original draft), Xue-Dong Wang (Data curation, Project administration, Resources, Visualization), Qi-Xuan Zheng (Data curation, Investigation, Resources, Validation), Hong Wang (Data curation, Project administration, Validation), Han Liu (Data curation, Methodology, Resources, Validation), Kong-Ying Lin (Data curation, Methodology, Resources, Supervision), Ying-Jian Liang (Data curation, Project administration, Resources, Visualization), Ya-Hao Zhou (Data curation, Investigation, Resources, Software, Validation), Wei-Min Gu (Data curation, Investigation, Resources, Validation), Ming-Da Wang (Conceptualization, Data curation, Formal analysis, Supervision, Visualization), Lan-Qing Yao (Data curation, Project administration, Software, Validation), Xin-Fei Xu (Data curation, Investigation, Software, Validation), Jia-Hao Xu (Data curation, Investigation, Resources, Software, Validation), Li-Hui Gu (Data curation, Investigation, Methodology, Validation), Timothy M. Pawlik (Formal analysis, Supervision, Writing—review & editing), Feng Shen (Conceptualization, Formal analysis, Supervision, Writing—review & editing), and Tian Yang (Conceptualization, Data curation, Funding acquisition, Project administration, Supervision, Writing—review & editing)

Disclosure

The authors declare no conflict of interest.

Supplementary material

Supplementary material is available at BJS Open online.

Data availability

The data set for this study is available upon request by contacting the corresponding author. The authors confirm that the data supporting the findings of this study are available within the article and its Supplementary material.

References

Author notes