Development and validation of GADA model for diagnosing hepatocellular carcinoma in Chinese hepatitis B patients

Abstract

Background:

Chronic hepatitis B virus (HBV) infection accounts for more than 40% of hepatocellular carcinoma (HCC) cases worldwide. The GALAD (Gender, Age, alpha fetoprotein heterogeneity [AFP-L3], AFP, and des-gamma-carboxy prothrombin [DCP]) model has potential for enhancing HCC diagnosis. However, its diagnostic performance in HBV-associated chronic liver disease requires further optimization. This study aimed to establish a diagnostic model for HBV-associated HCC.

Methods:

The data from 2190 patients with HBV-related liver diseases, including chronic hepatitis B (CHB), cirrhosis, and HCC were analysed. The GADA model was developed by omitting AFP-L3 for improved accessibility, and its performance was assessed via the area under the receiver operating characteristic curve (AUC). Validation was performed on 20,908 patients across ten hospitals.

Results:

The simplified GADA model demonstrated exceptional diagnostic performance in identifying HBV-related HCC in the modeling group, achieving an AUC of 0.940 (95% CI, 0.927–0.952), with a sensitivity and specificity of 82.57% and 91.53%, respectively, surpassing the original GALAD model’s AUC of 0.907 (95% CI, 0.891–0.921). In the internal validation cohort, the GADA model had an AUC of 0.935 (95% CI, 0.914–0.953), sensitivity of 82.04%, and specificity of 92.65%. For early-stage HCC detection, the AUC in the modeling group was 0.858 (95% CI, 0.837–0.877), and in the internal validation group was 0.868 (95% CI, 0.836–0.895), significantly better than the original GALAD model. External validation with 20,908 participants yielded an AUC of 0.919 (95% CI, 0.916–0.923), with sensitivity of 86.00% and specificity 81.88%. For early-stage HCC diagnosis, the AUC was 0.903 (95% CI: 0.898–0.907), with sensitivity of 88.16% and specificity 74.50%.

Conclusions:

The GADA model shows exceptional promise for diagnosing HBV-related HCC, enhancing early detection and reducing misdiagnosis rates. Its clinical utility is supported by robust validation, indicating significant implications for patient prognosis and treatment options.

Introduction

Primary liver cancer ranks as the sixth most prevalent malignancy and the third leading cause of cancer-related mortality globally. According to the World Health Organization, hepatocellular carcinoma (HCC) is the most common form of liver cancer.^[1] Approximately 80% of HCC cases are associated with chronic hepatitis B virus (HBV) or hepatitis C virus (HCV) infections.^[2] In contrast to Western regions such as Europe and the United States, HBV infection is the primary cause of HCC in China, accounting for 86% of cases.^[3–5] Although effective antiviral therapies have reduced the incidence of end-stage liver diseases, such as cirrhosis and liver failure, the overall incidence of HCC has not declined.^[6] This discrepancy highlights the urgent need for early detection strategies, which are crucial for enhancing long-term survival in patients with HCC.

The ongoing challenge in managing HCC lies in the difficulty of early diagnosis. Patients with HCC are often asymptomatic in the early stages. Conventional diagnostic tools, including imaging and alpha-fetoprotein (AFP) testing, exhibit limited sensitivity and specificity, especially in detecting early-stage HCC.^[7,8] This highlights the need for more reliable diagnostic strategies, especially in high-risk populations where early intervention can significantly improve survival outcomes.^[9]

In recent years, advances in biomarker research have led to the development of novel diagnostic models aimed at enhancing the early detection of HCC. In 2020, the FDA-approved GALAD (Gender, Age, alpha-fetoprotein heterogeneity [AFP-L3], AFP, and des-gamma-carboxy prothrombin [DCP]) score, a breakthrough algorithm that integrates three tumor markers, AFP, AFP-L3, and DCP, with patient gender and age, was introduced as an important alternative for HCC diagnosis.^[10] This algorithm has demonstrated promising diagnostic accuracy in Western populations, where HCV is more prevalent than HBV.^[11] However, the applicability of the GALAD model in HBV-endemic regions, such as China, remains uncertain. Given the distinct pathophysiological mechanisms underlying HBV-related HCC, there is a need to evaluate and potentially adapt the GALAD model for populations with chronic HBV infection.

Furthermore, certain biomarkers included in the GALAD model, such as AFP-L3, are not readily accessible in many clinical settings, particularly in resource-limited environments.^[12] This presents a significant barrier to the widespread implementation of the GALAD model in regions where HBV is prevalent. To address these limitations, ongoing research is focused on refined and simplified diagnostic models for HCC, ensuring that they maintain high diagnostic accuracy while being practical for routine clinical use.

This study aims to develop a simplified and accessible diagnostic model specifically for HBV-related HCC by integrating key biomarkers, such as AFP and DCP. It aims to enhance early detection capabilities and facilitate timely, personalized treatment. The model is designed to identify HCC at an asymptomatic stage, thereby improving therapeutic outcomes, reducing healthcare costs, and enhancing the quality of life for patients. Furthermore, this research aims to advance scientific progress, promote the application of new technologies, and provide robust support for comprehensive HCC prevention and treatment strategies.

Methods

Study design and participants

A total of 2190 patients with HBV-related chronic liver disease were recruited from the Mengchao Hepatobiliary Hospital of Fujian Medical University between June 2015 and June 2020. The cohort was randomly divided into two datasets at a 7:3 ratio for model development and internal validation. Between July 2022 and April 2024, an external validation cohort of patients with HBV-related HCC, chronic hepatitis B (CHB), and HBV-related liver cirrhosis (HBV-related LC) was recruited from ten hospitals across China, including Shenzhen Third People’s Hospital, Peking University Shenzhen Hospital, The First Affiliated Hospital of Zhejiang University School of Medicine, Sanming First Hospital, Peking University People’s Hospital, Qinghai Provincial People’s Hospital, The Second Affiliated Hospital of Chongqing Medical University, The Second Hospital of Shanxi Medical University, The First Affiliated Hospital of Fujian Medical University, and Wuhan Union Hospital.

Demographic information, clinical characteristics, diagnoses, and laboratory results (excluding AFP and DCP) were extracted from clinical databases. The outliers were analyzed to ensure data integrity.

This study was approved by the Peking University and the Mengchao Hepatobiliary Hospital of Fujian Medical University Ethics Committee (Nos. IRB00001052-19081, 2017-014-01). All ten subcenters obtained the corresponding ethical approvals, with detailed documents provided in Supplementary Table 1, http://links.lww.com/CM9/C981, and human genetic resource collection was authorized under the National Science and Technology Resource Office (Approval No. [2023] CJ1465). Informed consent was obtained from all participants, who permitted the use of their data for research purposes while ensuring anonymity. The study adhered to the principles of the 2013 Declaration of Helsinki.

Inclusion criteria

To be included in the study, participants were required to meet the following diagnostic standards: (1) primary liver cancer was diagnosed following the Guidelines for the Diagnosis and Treatment of Primary Liver Cancer (2011, 2017 Edition)^[13,14] with confirmation through histological or cytological examination of liver biopsy samples. The diagnoses of liver cirrhosis and chronic hepatitis were based on the Guidelines for the Prevention and Treatment of Chronic Hepatitis B (2015, 2019 Edition).^[15,16] HCC can be further classified into early and late stages according to the Milan criteria.^[17] In addition, the following conditions must be met: (a) patients must be 18 years or older, with no gender restrictions; (b) positive for hepatitis B surface antigen (HBsAg) and/or HBV DNA for more than six months; (2) all patients with HCC were newly diagnosed; (3) participants must not have severe liver or kidney dysfunction; and (4) participants must be willing to participate in the study and be capable of understanding and signing the informed consent form.

Exclusion criteria

Participants will be excluded from the study if any of the following conditions are met: (1) diagnosis of other viral infections, such as human immunodeficiency virus (HIV), before enrollment; (2) history of major organ transplantation, ongoing warfarin therapy, or prior surgery, ablation, radiotherapy, or chemotherapy for liver cancer, liver metastasis, or intrahepatic cholangiocarcinoma; (3) history or current clinical/laboratory evidence of pancreatitis or symptoms and laboratory results indicative of pancreatitis; (4) lack of available frozen serum samples or insufficient serum volume for biomarker analysis; (5) other conditions deemed unsuitable for inclusion by the investigators.

Collection, storage, and measurement of serum samples

Peripheral blood samples were collected from patients who met the study criteria. Blood samples were centrifuged at 4000 ×g for 15 minutes to isolate the serum. The serum was aliquoted and immediately frozen at −80°C until testing. The storage facilities and conditions at each research center were standardized. The serum levels of AFP and DCP were measured via chemiluminescent immunoassays (MAGLUMI^® AFP [CLIA] and MAGLUMI^® PIVKA-Ⅱ [CLIA], Snibe, Shenzhen, China), with specific storage conditions and usage guidelines following the instructions provided by the assay kits. Regular maintenance and service of the instruments were conducted throughout the project period. No adverse events occurred during the collection of serum samples.

The GALAD score is calculated using the following formula: −10.08 + 0.09 × (Age) +1.67 × (Gender [1 for male, 0 for female]) + 2.34 × log10 (AFP) + 0.04 × AFP-L3 (%) + 1.33 × log10 (DCP).

Statistical methods

According to the distribution characteristics of each variable, the appropriate descriptive statistical methods were selected, and data normality was assessed via the Kolmogorov–Smirnov (K–S) test. For quantitative data with a normal distribution, the mean and standard deviation were used for statistical description. Analysis of variance (ANOVA) was used to compare multiple groups, and intergroup comparisons were performed via the Student–Newman–Keuls method. The median and upper and lower quartiles were used for variables with a nonnormal distribution, and the Mann–Whitney U test was used for intergroup comparisons. Categorical variables were expressed as percentages, and intergroup comparisons were tested using chi-squared tests. SPSS version 26.0 software (IBM/SPSS Inc., Chicago, IL, USA), GraphPad Prism version 9.0 (for Windows, GraphPad Software, San Diego, California, USA; www.graphpad.com), R language (version 4.3.2, R Foundation for Statistical Computing, Vienna, Austria ) were used to perform the analyses, draw graphs, or both.

Receiver operating characteristic (ROC) curve analysis was performed. The optimal cutoff value was determined using the Youden index. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and percentage of correctly classified results were calculated. When ROC curves for the models/markers were generated for the same individuals, between-area under the receiver operating characteristic curve (AUC) comparisons were performed via the method proposed by Delong.^[18] This nonparametric statistical test compared the areas under two or more correlated receiver operating curves (MAGLUMI^® X8; Snibe, Shenzhen, China). The roc.test function in the pROC package in R was used to test the ROC curves with the parameters paired = TRUE and method = “delong”. The nomogram model was plotted via the rms and regplot packages. The Hosmer–Lemeshow goodness-of-fit test was used to assess the model’s fit. Internal validation was performed via the bootstrap method, and the corrected C-index was calculated to evaluate the stability of the prediction model. In addition, calibration curves were drawn to assess the model’s calibration. All the statistical tests were two-sided. A P value of <0.05 was considered to indicate statistical significance.

Results

Study design and cohorts

This study involved both a modeling and an internal validation cohort and an external validation cohort. The modeling and internal validation cohort data were derived from Mengchao Hepatobiliary Hospital, Fujian Medical University, and included 2336 patients. After the exclusion of 72 patients due to incomplete information, confirmed HIV infection, pancreatitis (either history or current clinical/laboratory abnormalities), and patients under 18 years of age, 41 patients with a history of liver cancer surgery, and an additional 33 patients who did not meet the study criteria, 2190 patients were included in the modeling analysis.

The external validation cohort consisted of data from ten medical institutions across China, totaling 23,663 patients. A total of 410 patients were excluded because of incomplete information, confirmed HIV infection, a history of pancreatitis or current pancreatitis-related abnormalities, or being under 18 years of age; 1310 patients had a history of liver cancer surgery; and 1035 patients did not meet the study criteria, resulting in 20,908 patients being included in the external validation cohort [Figure 1].

Figure 1.

Flowchart of participant selection. HBV: Hepatitis B virus; HCC: Hepatocellular carcinoma; HIV: Human immunodeficiency virus.

In total, 23,098 patients were included in this study. The data from the modeling and internal validation cohort were first randomly divided into a modeling set and an internal validation set. The modeling set was used to develop a predictive model, whereas the internal validation set was used for an initial assessment of the model’s accuracy and performance. The new model was subsequently validated in an external validation cohort from ten multicenter institutions to assess its generalizability and stability across different clinical settings. This stepwise validation strategy ensures the robustness of the constructed model and its applicability in real-world clinical practice.

Patient characteristics

Modeling and internal validation cohort

This study included 2190 patients. There were 989 patients with CHB, 575 patients with CHB-related cirrhosis, and 626 patients with HCC associated with CHB infection. A comparative analysis was conducted on the serum levels of the biomarkers AFP, AFP-L3%, and DCP among patients with CHB, LC, and HCC. The results demonstrated that the levels of AFP (median: 118.62 ng/mL), AFP-L3% (median: 6.34%), and DCP (median: 1240.00 mAU/mL) in the HCC group were significantly greater than those in the CHB and LC groups (P <0.001) [Figure 2A and B], indicating their importance in the early diagnosis of HCC [Table 1]. To assess the stability and accuracy of the model for the diagnosis of HCC, the study population was randomly divided into a modeling group (70%, 1533 patients) and an internal verification group (30%, 657 patients) [Supplementary Tables 2 and 3, http://links.lww.com/CM9/C981].

Figure 2.

Serum AFP and DCP levels in CHB, LC, and HCC groups within the modeling/internal validation cohort (A, B) and external validation cohort (C, D). AFP: Alpha-fetoprotein; CHB: Chronic hepatitis B; DCP: des-gamma-carboxy prothrombin; HCC: Hepatocellular carcinoma; LC: Liver cirrhosis.

Table 1.

Basic characteristics of patients with CHB, LC, and HCC

Variables	CHB	LC	HCC	H/χ2	P
Modeling and internal validation
Gender	989	575	626	76.540	<0.001
Male	692 (70.00)	426 (74.00)	555 (89.00)
Female	297 (30.00)	149 (26.00)	71 (11.00)
Age (years)	41 ± 12.71	53 ± 11.48	55 ± 11.79*#	353.770	<0.001
AFP (ng/mL)	6.10 (3.40, 27.90)	4.70 (3.00,9.90)*	118.62 (8.00,7378.81)*#	385.310	<0.001
AFP-L3%	0.50 (0.50, 1.41)	0.50 (0.50,4.32)	6.34 (0.50,14.54)*#	416.270	<0.001
DCP (mAU/mL)	25.00 (20.00, 32.30)	23.00 (18.00,30.00)	1240.00 (71.00,18553.50)*#	826.010	<0.001
External validation	17,802	2201	905
Gender				233.283	<0.001
Male	10,860 (61.00)	1564 (71.00)	743 (82.00)
Female	6942 (39.00)	637 (29.00)	162 (18.00)
Age (years)	46 ± 12.48	54 ± 11.33*	59 ± 11.45*#	1566.950	<0.001
AFP (ng/mL)	2.37 (1.63, 3.53)	2.59 (1.68, 4.43)*	10.10 (2.44, 477.50)*#	688.383	<0.001
DCP (mAU/mL)	23.60 (17.10, 32.10)	22.00 (15.70, 30.30)*	189.00 (33.10, 4079.90)*#	1237.780	<0.001

Data are presented as n, n (%) or median (Q1, Q3). ^*Compared with the chronic hepatitis B group, P <0.05; ^#Compared with the liver cirrhosis group, P <0.05. AFP: Alpha-fetoprotein; AFP-L3: Alpha-fetoprotein heterogeneity; CHB: Chronic hepatitis B; DCP: Des-gamma-carboxy prothrombin; HCC: Hepatocellular carcinoma; LC: Liver cirrhosis.

External validation cohort

This study included a total of 20,908 patients from ten hospitals nationwide, comprising 20,003 non-HCC patients (including CHB: 17,802, LC: 2201) and 905 patients with HCC. A comparative analysis of the demographic characteristics and serum biomarker levels among the CHB, LC, and HCC groups was conducted. The results revealed that the average age of the HBV-HCC patients was 59 years, which was significantly greater than the 46 years and 54 years in the CHB and LC groups (P <0.001, Table 1) respectivly. Compared with patients with CLD, patients with HBV-HCC were predominantly male, with significantly elevated serum DCP and AFP levels (P <0.001; [Table 1, Figure 2C and D]; Supplementary Table 2, http://links.lww.com/CM9/C981). These findings were consistent across all participating centers [Supplementary Tables 4–13 and Supplementary Figures 1 and 2, http://links.lww.com/CM9/C981]. Furthermore, a weak correlation was observed between the AFP and DCP levels in both the CLD and the HCC groups (Spearman r = 0.1052, P <0.001; [Supplementary Figure 3, http://links.lww.com/CM9/C981]). These results reinforce the idea that the combined use of AFP and DCP biomarkers can improve the diagnostic accuracy for HCC. Supplementary data, http://links.lww.com/CM9/C981 also provided detailed information on the expression of AFP and DCP across different centers, as well as the levels of other biomarkers (such as ALT, AST, ALP, and GGT) in the CLD and HCC groups [Supplementary Tables 4–13 and Supplementary Figures 1 and 2, http://links.lww.com/CM9/C981].

Simplification of the GALAD model to the new GADA model specific for HBV-related HCC diagnosis

To facilitate broader application of the model in the diagnosis of HBV-HCC while enhancing economic efficiency, we attempted to simplify the GALAD model while maintaining diagnostic accuracy. After analysis, a simplified model called GADA was developed by removing AFP-L3 from the original GALAD formula. The GADA model only includes gender (G), age (A), log10(DCP) (D), and log10(AFP) (A) as variables. The new formula is as follows: Z = −11.111 + 1.122 × gender (male: 1, female: 0) + 0.073 × age + 2.660 × Log10(DCP)+0.578×Log10(AFP) [Supplementary Table 3, http://links.lww.com/CM9/C981]. We constructed a nomogram prediction model for diagnosing HBV-HCC using the variables from the model. The results of the nomogram indicate that the probability of HCC increases in male patients with CLD associated with HBV infection as their age, log10(AFP), and log10(DCP) levels increase [Supplementary Figure 4, http://links.lww.com/CM9/C981].

Value assessment of the GADA model for diagnosing HCC

To validate that GADA is an optimized model and to assess the value of GADA as an improved model for diagnosing HBV-HCC, we compared the diagnostic performance of the GADA and GALAD models in both the modeling and internal validation groups. Moreover, the diagnostic values of AFP and DCP for HCC were also assessed in parallel. In the modeling group, for the diagnosis of HBV-HCC, the AUC of the GADA model was 0.940 (95% CI: 0.927–0.952), with a sensitivity and specificity of 82.57% and 91.53%, respectively. The AUC for the original GALAD model was 0.907 (95% CI: 0.891–0.921). With respect to diagnostic performance, GADA outperformed the original GALAD model. Nevertheless, these models demonstrated superior diagnostic performance compared with AFP (0.767, 95% CI: 0.745–0.788) and DCP (0.892, 95% CI: 0.875–0.907) when assessed individually [Figure 3A and Table 2]. We obtained similar results in the internal validation cohort, with the AUC, sensitivity, specificity, PPV, NPV, and Youden’s index for GADA being 0.935 (95% CI: 0.914–0.953), 82.04%, 92.65%, 79.19%, 93.80%, and 0.75, respectively, all of which were higher than those of the GALAD model [Figure 3B and Table 2].

Figure 3.

Performance of the GADA and GALAD models for HBV-HCC detection. (A) Model group. (B) Internal validation group. (C) External validation group. GALAD: Gender, Age, alpha-fetoprotein heterogeneity, alpha-fetoprotein, and des-gamma-carboxy prothrombin; HBV: Hepatitis B virus; HCC: Hepatocellular carcinoma.

Table 2.

Values of AFP, AFP-L3, and DCP alone and in combination with the GADA model in the diagnosis of HCC.

Variables	AUC (95% CI)	Cutoff	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Youden index	P
Modeling group (n = 1074/459)
GADA	0.940 (0.927–0.952)	−2.00	82.57	91.53	80.64	92.47	0.74	<0.001*
GALAD	0.907 (0.891–0.921)	1.49	81.92	83.33	67.74	91.51	0.65	<0.001*
AFP	0.767 (0.745–0.788)	20.00	65.58	73.74	51.60	83.40	0.41	<0.001*
DCP	0.892 (0.875–0.907)	40.00	81.92	86.78	72.60	91.80	0.73	<0.001*
AFP-L3%	0.754 (0.731–0.775)	0.02	71.81	68.69	58.43	79.90	0.41	<0.001*
Internal validation (n = 490/167)
GADA	0.935 (0.914–0.953)	−2.00	82.04	92.65	79.19	93.80	0.75	<0.001*
GALAD	0.899 (0.873–0.921)	1.49	76.05	84.08	62.00	91.20	0.64	<0.001*
AFP	0.747 (0.712–0.780)	20.00	59.28	78.57	48.50	85.00	0.40	<0.001*
DCP	0.887 (0.860–0.910)	40.00	81.44	88.78	71.20	93.30	0.73	<0.001*
AFP-L3%	0.738 (0.702–0.771)	0.02	68.67	72.22	44.91	86.49	0.41	<0.001*
External validation (n = 20,003/905)
GADA	0.919 (0.916–0.923)	−2.00	86.00	81.88	15.96	99.32	0.68	<0.001*
AFP	0.748 (0.742–0.754)	20.00	44.57	97.94	46.42	97.78	0.41	<0.001*
DCP	0.841 (0.836–0.846)	40.00	68.55	85.79	16.19	98.55	0.53	<0.001*

^*P <0.05. AFP: Alpha-fetoprotein; AFP-L3: Alpha-fetoprotein heterogeneity; AUC: Area under the receiver operating characteristic curve; DCP: Des-gamma-carboxy prothrombin; PPV: Positive predictive value; NPV: Negative predictive value.

The Hosmer–Lemeshow goodness-of-fit test was used to evaluate the model’s fit, revealing no statistically significant prediction bias between the predicted and actual values in both the modeling (P = 0.396) and internal validation (P = 0.320) cohorts (both P values >0.05). Internal validation of the prediction model via 1000 bootstrap resamples indicated that the calibration of the prediction model was good in both the modeling and internal validation cohorts (in the modeling cohort: corrected C-index = 0.942, Brier score = 0.076 <0.250; in the internal validation cohort: corrected C-index = 0.835, Brier score = 0.069 <0.250) [Supplementary Figure 5, http://links.lww.com/CM9/C981].

Diagnostic value of the GADA model for HCC in external validation

In this study, we evaluated the diagnostic value of the GADA model, DCP, and AFP in patients with HBV-HCC. The results indicated that the GADA model exhibited superior diagnostic performance, with an AUC of 0.919 (95% CI: 0.916–0.923), which was significantly greater than that of DCP (AUC: 0.841, 95% CI: 0.836–0.846) and AFP (AUC: 0.748, 95% CI: 0.742–0.754). At a cutoff value of −2.00, the GADA model achieved an optimal accuracy of 0.68, with a sensitivity of 86.00%, which is a significant improvement over DCP (68.55%) and AFP (44.57%). Furthermore, the GADA model maintained a high specificity of 81.88%, highlighting its potential in clinical diagnosis [Figure 3C and Table 2].

To further validate the generalization and robustness of the GADA model, its diagnostic performance was evaluated across multiple centers, as presented in the Table 3. The AUC values consistently remained high, ranging from 0.895 to 0.977. The sensitivity ranged from 80.00% to 100%, and the specificity ranged from 73.96% to 97.11%. In addition, the NPVs were uniformly high across all centers, with some centers achieving a perfect score of 100%. In some centers, the PPVs were relatively low due to variations in disease prevalence; however, the highest PPV reached 50.33%. These findings demonstrate that the GADA model has been thoroughly validated in diverse clinical settings (P <0.001) [Table 3].

Table 3.

Individual analysis from each center further validated the diagnostic performance of the GADA model across diverse populations.

Institution	n	AUC (95% CI)	Cutoff	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Youden index	P
Multicenter (overall)	20,908	0.919 (0.916–0.923)	−2.00	86.00	81.88	15.96	99.32	0.68	<0.001*
Multicenter 1	2533	0.977 (0.971–0.983)	−1.08	93.15	97.11	47.50	99.80	0.90	<0.001*
Multicenter 2	2375	0.972 (0.965–0.979)	−2.48	100.00	86.70	1.99	100.00	0.87	<0.001*
Multicenter 3	2311	0.949 (0.939–0.958)	−1.53	84.47	91.53	37.61	98.98	0.76	<0.001*
Multicenter 4	656	0.930 (0.907–0.948)	−1.29	82.56	90.00	50.33	97.68	0.73	<0.001*
Multicenter 5	2074	0.935 (0.924–0.945)	−1.61	88.89	85.88	6.94	99.85	0.75	<0.001*
Multicenter 6	2649	0.910 (0.898–0.921)	−0.60	82.47	83.46	14.15	99.31	0.66	<0.001*
Multicenter 7	2635	0.903 (0.891–0.914)	−2.57	92.59	73.96	3.46	99.90	0.67	<0.001*
Multicenter 8	1364	0.913 (0.896–0.927)	−1.03	82.93	88.46	27.28	99.00	0.71	<0.001*
Multicenter 9	2151	0.901 (0.887–0.913)	−2.03	80.00	86.07	42.06	97.14	0.66	<0.001*
Multicenter 10	2160	0.895 (0.881–0.908)	−2.23	88.00	87.35	7.49	99.84	0.75	<0.001*

^*P <0.05. Multicenter1: Shenzhen Third People’s Hospital; Multicenter2: Peking University Shenzhen Hospital; Multicenter3: The First Affiliated Hospital of Zhejiang University School of Medicine; Multicentre4: Sanming First Hospital; Multicenter5: Peking University People’s Hospital; Multicentre6: Qinghai Provincial People’s Hospital; Multicentre7: The Second Affiliated Hospital of Chongqing Medical University; Multicentre8: The Second Hospital of Shanxi Medical University; Multicentre9: The First Affiliated Hospital of Fujian Medical University; Multicentre10: Wuhan Union Hospital. AUC: Area under the receiver operating characteristic curve; Cutoff: Cutoff value: PPV: Positive predictive value; NPV: Negative predictive value.

The supplementary data provided the results of the diagnostic performance of AFP and DCP in identifying HBV-HCC across different centers, comparing these results with those of the GADA model. The findings indicate that the overall performance of the GADA model surpasses that of AFP or DCP used independently. Moreover, compared with AFP, DCP demonstrated superior diagnostic efficacy, sensitivity, and specificity, further highlighting the advantage of the GADA model over single biomarkers [Supplementary Tables 14–23 and Supplementary Figure 6, http://links.lww.com/CM9/C981].

Diagnostic value of the GADA model for early HCC

Early detection of HCC continues to pose a clinical challenge, and in this study, the diagnostic value of GALAD and GADA for early HCC was further analyzed. The results revealed that the AUC of GADA for the diagnosis of early HCC in the modeling group was 0.858 (95% CI: 0.837–0.877), which was significantly better than that of GALAD or AFP and DCP (P <0.001) alone [Figure 4A and Table 4]. These results suggest that the modified model can significantly contribute to the screening and auxiliary diagnosis of early HCC and that missed diagnoses could be avoided, at least to some extent. The same results were obtained in the internal validation group, and the AUC of the GADA model for diagnosing early HCC was 0.868 (95% CI: 0.836–0.895) [Figure 4B and Table 4]. The sensitivity, specificity, and other performance evaluation indicators of GADA performed well in both groups. The GADA model demonstrated an AUC of 0.903 (95% CI: 0.898–0.907) in the external validation of the multicenter study, with a sensitivity of 88.16% and specificity of 74.50%. These results provide unexpected confirmation. In comparison, GADA is highly valuable for the early diagnosis of HCC, and both are significantly better than GALAD, AFP, and DCP (P <0.001) [Figure 4C and Table 4].

Figure 4.

Performance of the GADA and GALAD models for the detection of early-stage HBV-HCC. (A) Modeling group; (B) internal validation group; and (C) external validation group. GALAD: Gender, Age, alpha-fetoprotein heterogeneity, alpha-fetoprotein, and des-gamma-carboxy prothrombin; HBV: Hepatitis B virus; HCC: Hepatocellular carcinoma.

Table 4.

Values of AFP, AFP-L3, and DCP alone and the GADA and GALAD model in the diagnosis of early HCC.

Variables	AUC (95% CI)	Cutoff	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Youden index	P
Modeling group
GADA	0.858 (0.837–0.877)	−2.34	68.03	87.90	78.94	80.48	0.56	<0.001*
GALAD	0.798 (0.774–0.820)	1.49	58.50	81.28	30.00	93.50	0.47	<0.001*
AFP	0.612 (0.584–0.639)	20.00	42.18	73.74	18.00	90.30	0.23	<0.001*
DCP	0.752 (0.727–0.776)	40.00	60.54	89.20	43.42	94.29	0.50	<0.001*
AFP-L3%	0.627 (0.599–0.654)	0.02	53.79	68.69	50.22	71.67	0.22	<0.001*
Internal validation
GADA	0.868 (0.836–0.895)	−2.34	70.91	87.76	77.23	83.70	0.64	<0.001*
GALAD	0.798 (0.761–0.831)	1.49	50.91	83.47	25.70	93.80	0.49	<0.001*
AFP	0.590 (0.548–0.632)	20.00	32.73	78.57	14.60	91.20	0.26	<0.001*
DCP	0.783 (0.746–0.817)	40.00	60.00	88. 78	37.50	95.20	0.51	<0.001*
AFP-L3%	0.626 (0.583–0.667)	0.02	50.00	73.05	52.14	71.33	0.23	<0.001*
External validation
GADA	0.903 (0.898–0.907)	−2.34	88.16	74.50	1.64	99.92	0.62	<0.001*
DCP	0.816 (0.811–0.822)	40.00	64.79	85.78	2.15	99.80	0.49	<0.001*
AFP	0.736 (0.730–0.742)	20.00	38.50	97.89	8.09	99.70	0.35	<0.001*

^*P <0.05. AFP: Alpha-fetoprotein; AFP-L3: Alpha-fetoprotein heterogeneity; AUC: Area under the receiver operating characteristic curve; DCP: Des-gamma-carboxy prothrombin; PPV: Positive predictive value; NPV: Negative predictive value.

Discussion

We validated the GALAD model in distinguishing HCC from HBV-related CLD in a Chinese cohort. Using the original GALAD formula to differentiate HCC from CLD, we achieved an AUC value of 0.907, which is significantly lower than that reported in the initial study conducted with the UK cohort (AUC 0.97)^[10] and other studies reporting AUC values ranging from 0.93 to 0.98.^[11,19–21] This difference is attributed mainly to the causative difference between China and Europe/the United States. Unlike the original study cohort, in which chronic HBV infection-related HCC accounted for only a very small proportion of cases,^[10] chronic HBV infection accounted for more than 80% of HCC cases in China and more than 50% of HCC cases worldwide.^[22,23] In Europe and the United States, HCV infection, alcoholic liver disease, and nonalcoholic fatty liver disease play key roles in the occurrence of HCC.^[24] In addition, the age of onset of HCC varies in different parts of the world. An analysis of 18,031 patients with HCC in 14 countries revealed that the average ages of onset of HCC in Japan, Europe, and North America were 69, 65, and 62 years, respectively, whereas the average onset ages of HCC in China and Republic of Korea were 52 and 59 years, respectively.^[24] In a study that included 14,891 patients with HCC from 2016 to 2018 in China,^[25] the average age of onset of HCC was 55 years, which further confirmed that the average age of onset of HCC in China is lower than that in Western countries. This result is similar to the average age of onset in our study cohort, further highlighting the importance of developing region-specific diagnostic models. Such models can better account for demographic and epidemiological variations, improving diagnostic accuracy and clinical applicability in specific populations.

In this study, we found that, compared with AFP and AFP-L3%, DCP had greater sensitivity for diagnosing HCC, with the goal of ensuring greater diagnostic accuracy. This method plays an important role in the screening and auxiliary diagnoses of early HCC and can avoid missed diagnoses in some patients with early HCC. Since DCP was reported to be elevated in the plasma of patients with HCC, DCP has been considered a specific marker for the diagnosis of HCC.^[26–29] Serum DCP levels^[27] are elevated in 44–81% of patients with HCC, and a high level of DCP is considered associated with large tumors and HCC recurrence. In 2002, the Japanese guidelines recommended AFP-L3 and DCP as two additional serum tests for early-stage HCC detection and included them in the National Health Insurance.^[30] However, studies have indicated that, compared with patients not receiving antiviral treatment, patients receiving antiviral treatment had significantly lower median levels of AFP (P <0.001), and there were significantly fewer patients with abnormal AFP levels in the antiviral group.^[31,32] Moreover, AFP levels are also increased in patients with chronic or active liver disease, pregnancy, gonadal embryogenic tumors, and some digestive tract tumors. All these factors limit the clinical application of AFP to a certain extent. Therefore, the contribution of the AFP would be smaller in our GADA model than in the GALAD model. In our study, among the various biomarkers, DCP demonstrated the best performance in differentiating HCC from CLD. However, previous studies have reported inconsistent results regarding the superiority of specific biomarkers. Some studies favor DCP,^[33–35] whereas others support AFP as a better marker.^[36–38] Nevertheless, there is a consistent finding across the literature: the combination of both biomarkers provides superior diagnostic performance compared with the use of a single biomarker alone.^[38–42]

Our decision to simplify the model by excluding AFP-L3 was driven by several factors: its high measurement complexity, cost, sample volume requirement, and, most importantly, its suboptimal sensitivity as a single biomarker for HCC, particularly in HBV-endemic settings. Recent research has also supported the limited utility of AFP-L3% as a reliable biomarker for HCC, emphasizing the need for more robust diagnostic tools.^[43] A meta-analysis including 12 studies assessed the sensitivity of AFP-L3% in diagnosing HCC. The results revealed that the sensitivity of AFP-L3% in diagnosing HCC was 48.30%.^[44] In a Korean cohort study, the sensitivity of AFP-L3% in diagnosing HCC was only 18.80%.^[36] Studies from others and our previous study revealed a strong correlation between AFP-L3 and AFP in a cohort of HBV-HCC patients.

We conducted external validation of the GADA model in a cohort of 20,908 patients from 10 hospitals across China, assessing its diagnostic accuracy and robustness across diverse populations and clinical settings. The overall AUC for the GADA model reached 0.919 (95% CI: 0.916–0.923), with high sensitivity (86.00%) and specificity (81.88%), effectively improving diagnostic efficiency and reducing the occurrence of misdiagnoses and missed diagnoses. The external validation results for early HCC were particularly promising, with an impressive AUC of up to 0.903 (95% CI: 0.898–0.907), a sensitivity of 88.16%, and a specificity of 74.50%. In addition, in our single-center validation results, AFP outperformed DCP in the diagnosis of HCC in a limited number of centers, which may be attributed to the limited number of HCC patients. However, the GADA model performed well, indicating that the model exhibits good robustness and generalization across different patient populations and medical settings [Supplementary Table 14–23, http://links.lww.com/CM9/C981]. These findings suggest that the model can provide a valuable basis for the early diagnosis of HCC in clinical practice. In our study, the reason for not standardizing cutoff values across centers was due to variations in testing equipment and laboratory conditions, which may have led to data discrepancies. Instead, each center’s cutoff value was determined based on the optimal combination of sensitivity and specificity to better reflect diagnostic performance across different centers. Notably, the cutoff values in the Qinghai region were relatively low. We also found that the optimal cutoff value for DCP in diagnosing HCC in this region was substantially greater than that in other regions (112 mAU/mL) [Supplementary Table 20, http://links.lww.com/CM9/C981]. This discrepancy may be related to the region’s geographical location and the unique lifestyle habits of its residents, which warrant further in-depth research.

In summary, within a Chinese cohort of patients infected with HBV, the GADA model, constructed using only four variables, was developed by simplification of the GALAD parameters. The GADA model outperformed the original GALAD model. Furthermore, the GADA model’s robustness and generalization across various clinical settings were confirmed through external validation in a large multicenter cohort.

Acknowledgments

We would like to express our sincere gratitude to all the patients who participated in this study and provided valuable medical data. We also thank the healthcare professionals and staff at Mengchao Hepatobiliary Hospital of Fujian Medical University, Shenzhen Third People’s Hospital, Peking University Shenzhen Hospital, The First Affiliated Hospital of Zhejiang University School of Medicine, Sanming First Hospital, Peking University People’s Hospital, Qinghai Provincial People’s Hospital, The Second Affiliated Hospital of Chongqing Medical University, The Second Hospital of Shanxi Medical University, The First Affiliated Hospital of Fujian Medical University, Wuhan Union Hospital, and Shenzhen New Industries Biomedical Engineering Co., Ltd. Their critical roles in patient enrollment and data collection were essential. This research would not have been possible without their dedication and support.

Funding

This work was supported by the National Natural Science Foundation of China Grant (No. 82272433) and the fund of Fujian Provincial Key Laboratory of Hepatic Drug Research (No. 2022-YF-0050).

Conflicts of interest

None.

Supplementary Material

cm9-139-3110-s001.docx ^{(1.4MB, docx)}