Diagnostic Performance Comparison of AFP, PIVKA-II, GALAD Model, and ASAP Model Across Two Chemiluminescence Immunoassay Platforms for Hepatocellular Carcinoma

Introduction

In 2020, approximately 905,677 new cases of liver cancer were diagnosed globally, resulting in over 830,000 deaths.¹ In China, liver cancer is the second leading cause of cancer-related mortality.² Hepatocellular carcinoma (HCC), which accounts for more than 80% of primary liver cancer cases, poses a conundrum of rapid progression, early metastasis, and a lack of early-stage symptoms.³ The overall five-year survival rate for HCC is only 18%;⁴ however, this increases to over 70% for patients diagnosed and treated at an early stage.⁵ Therefore, early diagnosis and effective monitoring of HCC progression are crucial for improving patient outcomes and reducing mortality.

Serum markers such as alpha-fetoprotein (AFP) and protein induced by vitamin K absence or antagonist-II (PIVKA-II) are widely employed for diagnosing and monitoring HCC.⁶ These biomarkers provide non-invasive options for detecting HCC, but their individual diagnostic performance remains limited: meta-analyses have reported moderate sensitivity (61% for AFP, 66% for PIVKA-II) and specificity (87% for AFP, 88% for PIVKA-II).^7–9

To mitigate the limitations of inadequate performance of individual serum biomarkers, composite diagnostic models integrating clinical parameters have been developed. A prominent example is the GALAD model, first introduced in 2014.¹⁰ This model, constructed using the logistic regression method, incorporates five parameters (gender, age, AFP-L3, AFP, and PIVKA-II) and was initially established using the mTASWako i30 autoanalyzer. Since its development, the GALAD model has been extensively validated across diverse populations and etiologies, consistently demonstrating high sensitivity and specificity.^10–14 Another representative model is the ASAP model, proposed in 2019.¹⁵ Also built based on the logistic regression method, ASAP differs from GALAD in that it excludes AFP-L3 and relies on four parameters (gender, age, AFP, and PIVKA-II); it was originally developed and tested on the ARCHITECT immunoassay platform. Subsequent external validations have further confirmed the ASAP model’s robust diagnostic utility.^15–18

Notably, recent studies have primarily focused on comparing diagnostic performance for HCC: between these models, between the models and individual serum biomarkers, and across different models in etiologically distinct populations.^19–22 However, no systematic studies have investigated whether the performance of these models varies when applied to biomarker data generated by different analytical platforms. Importantly, these distinct detection platforms differ in key technical parameters, including reagent formulations, detection thresholds, signal amplification mechanisms, and instrument calibration protocols. Such platform-related differences may introduce quantitative biases in biomarker measurements, which can alter model outputs and further affect diagnostic decisions. The primary objective of this study is to compare the clinical performance of AFP, PIVKA-II, the GALAD model, and the ASAP model for HCC diagnosis across two distinct detection platforms.

Methods

Participants

A total of 518 participants were enrolled from Peking Union Medical College Hospital (Beijing, China) between 2011 and 2021. The cohort comprised 102 HCC patients, 117 with benign liver disease (BLD), 38 with cholangiocarcinoma (CCA), 96 with colorectal cancer (CRC), 65 with metastatic hepatic carcinoma (MHC), and 100 healthy controls. HCC was defined according to the American Association for the Study of Liver Diseases (AASLD) Practice Guidelines for the management of HCC (updated version, 2010).²³ BLD cases included patients with hepatic cysts or hepatic hemangiomas confirmed by ultrasound or computed tomography. CCA, CRC, and MHC diagnoses were pathologically confirmed. Demographic and clinical data from electronic medical records were retrieved, including patient gender, age, and alanine aminotransferase (ALT) levels.

Measurement of Serum AFP, PIVKA-II and AFP-L3

Three to five milliliters of venous blood were collected from each participant. Samples were centrifuged at 1,500 ×g for 10 min, and the obtained serum was aliquoted and stored at −80°C until analysis. Frozen samples were thawed at room temperature prior to testing. Serum AFP and PIVKA-II levels were measured using the Abbott ARCHITECT i2000 automated analyzer (Illinois, USA) and the Hotgen Biotech C2000 automated analyzer (Beijing, China). Serum AFP-L3 was measured using the Hotgen Biotech C2000 automated analyzer. The detection methods employed on both the Abbott and Hotgen platforms are chemiluminescence microparticle immunoassays. All tests were performed as per the manufacturers’ specifications. Results from the Hotgen Biotech C2000 were designated as AFP_Hotgen, PIVKA-II_Hotgen, and AFP-L3_Hotgen; Results from the Abbott ARCHITECT i2000 were indicated as AFP_Abbott and PIVKA-II_Abbott. PIVKA-II_Hotgen was expressed in ng/mL, whereas PIVKA-II_Abbott was reported in mAU/mL (1 mAU/mL = 1 ng/mL per manufacturer’s specifications).

Definition of GALAD and ASAP Models

GALAD and ASAP models were calculated according to previous studies to evaluate their diagnostic value for HCC.^10,15 GALAD score= −10.08 + 0.09 × [Age] + 1.67 × [Gender (1 for males, 0 for females)] + 0.04 × [AFP-L3%] + 2.34 × log₁₀ [AFP] + 1.33 × log₁₀ [PIVKA-II]. ASAP score= −7.57711770 + 0.04666357 × Age - 0.57611693 × [Gender (0 for males, 1 for females)] + 0.42243533 × ln [AFP] + 1.10518910 × ln[PIVKA-II].

Statistical Analysis

For the evaluation of HCC diagnostic performance and the calculation of GALAD and ASAP scores, all specimens were included. Values exceeding the upper limit of detection for the Abbott platform (>2000 ng/mL for AFP and >30000 mAU/mL for PIVKA-II) were censored at these respective thresholds. Subsequently, for the inter-platform comparison, specimens with analyte levels exceeding the Abbott system’s upper limits (13 for AFP and 3 for PIVKA-II) were excluded, yielding 505 and 515 valid paired measurements for AFP and PIVKA-II, respectively.

Statistical analyses were performed using SPSS 18.0, GraphPad Prism 10.0, Origin 2021, and MedCalc for data analyses and graphical representations. The Kolmogorov–Smirnov test was employed to assess the distribution of variables. The Mann–Whitney U-test and Kruskal–Wallis H-test were utilized to compare differences between groups. Receiver operating characteristic (ROC) curves and area under the curve (AUC) with 95% confidence intervals (CIs) were conducted to elucidate the diagnostic values of AFP, PIVKA-II, AFP-L3 and composite models. DeLong’s test was used for AUC comparisons. Parameters related to diagnostic efficiency, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated.

Spearman correlation analysis was utilized to evaluate the correlation between the two platforms. Passing-Bablok regression was applied to assess systematic and proportional bias. Bland-Altman plots were used to determine the consistency between two variables. P values < 0.05 were regarded as statistically significant.

Results

Clinical Characteristics of Participants

In this study, the clinical characteristics of 518 enrolled participants, including age, gender, serum ALT levels, and concentrations of serum AFP, PIVKA-II, and AFP-L3, are presented in Table 1. In the HCC group, the mean age of patients was 56.91 years, with 75.5% of male patients, and 55.9% of patients with elevated ALT levels (>40 IU/L). AFP, PIVKA-IIand AFP-L3 levels in the HCC group, measured on both platforms, were significantly elevated compared to all other groups (P < 0.001, Figure 1). Inter-group comparisons were performed using Mann–Whitney U-test and Kruskal–Wallis H-test for non-parametric data. Figure 1 visually compares the distributions of AFP, PIVKA-II, and AFP-L3 across all groups.

Table 1 Clinical Features of Participants Enrolled in This Study

Figure 1 Expression levels of AFP_Hotgen (a), AFP_Abbott (b), PIVKA-II_Hotgen (c), PIVKA-II_Abbott (d) and AFP-L3_Hotgen (e) in HCC group and five non-HCC groups. Abbreviations: HCC, hepatocellular carcinoma; CCA, cholangiocarcinoma; CRC, colorectal cancer; MHC, metastatic hepatic carcinoma; BLD, benign liver disease; AFP, alpha fetoprotein; PIVKA-II, protein induced by vitamin K absence or antagonist-II; AFP-L3, alpha-fetoprotein lens culinaris agglutinin 3. Notes: ***P<0.001.

Diagnostic Performance of AFP, PIVKA-II, AFP-L3, GALAD Model, and ASAP Model Determined by Two Platforms

ROC analysis was used to evaluate the diagnostic performance of AFP, PIVKA-II, AFP-L3, GALAD model, and ASAP model for HCC, with results shown in Table 2. Notably, the table provides sensitivity, specificity, PPV, and NPV corresponding to both ROC-derived optimal cut-off values and commonly used clinical cut-offs,²⁴ enabling clinicians to better evaluate and apply these biomarkers in practice. DeLong’s test was used for AUC comparisons. The AUC of AFP_Hotgen was 0.821 (95% CI: 0.786–0.853), and that of AFP_Abbott was 0.846 (95% CI: 0.811–0.876). The diagnostic efficiency of Hotgen Biotech was similar to Abbott Architect (P = 0.0873). For PIVKA-II, AUC comparison revealed that the diagnostic efficiency of the Abbott platform was mildly higher than that of the Hotgen platform (P = 0.0017). The AUC of AFP-L3_Hotgen was 0.714 (95% CI: 0.673–0.752).

Table 2 Diagnostic Performance of Biomarkers and Models for HCC: ROC-Derived Optimal Cut-Offs and Commonly Used Clinical Cut-Offs

The GALAD_Hotgen model and the ASAP_Hotgen model exhibited significantly higher AUC values for HCC diagnosis compared to the individual serum markers AFP_Hotgen, PIVKA-II_Hotgen, and AFP-L3_Hotgen (P < 0.05). Meanwhile, the GALAD_Abbott model and ASAP_Abbott model also demonstrated a notable increase in AUC values for HCC diagnosis compared to the individual serum markers AFP_Abbott and PIVKA-II_Abbott (P < 0.05). Furthermore, when comparing different models within the same platform, the AUCs of the GALAD_Hotgen and ASAP_Hotgen models were identical, with both achieving an AUC of 0.872. The ASAP_Abbott model (AUC = 0.913) exhibited a slightly higher AUC than the GALAD_Abbott model (AUC = 0.901), though this difference did not reach statistical significance (p = 0.0569). For comparisons of the same model across different platforms, the ASAP_Abbott model (AUC = 0.913) demonstrated a significantly higher AUC than the ASAP_Hotgen model (AUC = 0.872) (P = 0.0003). Similarly, the GALAD_Abbott model (AUC = 0.901) yielded a significantly higher AUC compared to the GALAD_Hotgen model (AUC = 0.872) (P = 0.0001).

It is worth explaining that due to the unavailability of AFP-L3 testing on the Abbott platform, when evaluating the GALAD_Abbott model, AFP-L3 was measured using the Hotgen platform, while other biomarkers were tested on the Abbott platform.

Comparison of Serum AFP and PIVKA-II Levels Between Hotgen Biotech and Abbott Architect

Passing-Bablok regression analysis revealed that the regression equation for AFP is AFP_Hotgen = 1.116 * AFP_Abbott + 0.221 (P = 0.16). The slope’s 95% confidence interval (CI) was 1.066 to 1.162, and the intercept’s 95% CI ranged from 0.111 to 0.373. Spearman correlation analysis revealed a correlation coefficient was 0.573 (P <0.001). The regression equation for PIVKA-II is PIVKA-II_Hotgen = 0.093 * PIVKA-II_Abbott + 2.344 (P <0.01). The slope’s 95% CI was 0.078 to 0.108, and the intercept’s 95% CI ranged from 2.052 to 2.627. The correlation coefficient from Spearman correlation analysis was 0.460 (P <0.001). The scatter plots comparing AFP and PIVKA-II between the two platforms were displayed in Figure 2a and 2b, respectively.

Figure 2 Quantitative comparison of serum AFP and PIVKA-II levels between Hotgen Biotech and Abbott Architect. (a) Scatter plot between AFP_Hotgen and AFP_Abbott. (b) Scatter plot between PIVKA-II_Hotgen and PIVKA-II_Abbott. (c) Bland-Altman plot between AFP_Hotgen and AFP_Abbott. (d) Bland-Altman plot between PIVKA-II_Hotgen and PIVKA-II_Abbott. Abbreviations: AFP, alpha fetoprotein; PIVKA-II, protein induced by vitamin K absence or antagonist-II; SD, standard deviation.

Bland-Altman plots were used to compare AFP and PIVKA-II differences between the two platforms. According to the external quality control criteria of the China National Center for Clinical Laboratories, the allowable bias for both markers is ±30%, with 80% of data required to fall within this range. For AFP, the mean bias between the Hotgen and Abbott platforms was 44.32%, and only 57.43% (290/505) of the data points fell within the ±30% allowable bias range (Figure 2c), indicating poor consistency. For PIVKA-II, the mean bias between Hotgen and Abbott was −92.02%, with merely 9.32% (48/515) of data within the allowable range (Figure 2d). Notably, PIVKA-II concentrations measured by Hotgen were lower than those by Abbott in most samples.

Discussion

In this study, we evaluated the diagnostic performance of individual serum biomarkers (AFP, PIVKA-II, AFP-L3), and the composite models (GALAD and ASAP) for HCC across two distinct immunoassay platforms.

AFP levels were significantly higher in the HCC group than in all non-HCC groups across both platforms. The diagnostic efficacy of AFP was clinically favorable on both systems, with AUCs of 0.821 (Hotgen) and 0.846 (Abbott), showing no significant inter-platform difference (p > 0.05). However, its sensitivity remained insufficient even at the newly identified optimal cut-off (3~5 ng/mL), reaching only ~70%. When using the commonly used clinical cut-off of 20 ng/mL, sensitivity further decreased to approximately 40%. Consistent with previous literature,^6,7 these results confirm that AFP alone provides inadequate sensitivity for HCC screening, leading to considerable missed diagnoses. Similarly, HCC patients exhibited elevated PIVKA-II levels, and Abbott’s performance was marginally better than Hotgen’s [(AUC: 0.863 vs 0.787, P < 0.05)]. Despite this, both platforms exhibited limited sensitivity and specificity for diagnosing HCC—whether at the optimal cut-offs or commonly used clinical cut-offs—as shown in Table 2. AFP-L3 consistently exhibited low sensitivity. This makes AFP-L3 more useful for excluding HCC than screening it, which is consistent with Debes JD’s observations.²⁵ These findings underscore the constraints of relying on single biomarkers for diagnosing HCC,²⁶ highlighting that the application of composite models is the way forward for the serological diagnosis of HCC.

Nowadays, several composite models—including the GALAD and ASAP models—have emerged as promising tools for HCC diagnosis.^10,15 Consistent with previous literature, our findings demonstrate that both the GALAD and ASAP models exhibited significantly higher AUC values for HCC diagnosis compared to the individual serum markers AFP, PIVKA-II, and AFP-L3.^10–18 Numerous studies have compared the diagnostic performance of the GALAD and ASAP models for HCC, yet consistent conclusions remain elusive. In a multicenter study by Sun et al, the ASAP model (AUC = 0.886) demonstrated superior HCC detection efficacy compared to the GALAD model (AUC = 0.853), particularly in patients with chronic liver diseases and early-stage HCC.¹⁹ For biomarker testing, AFP and PIVKA-II were measured via Abbott ARCHITECT kits, while AFP-L3 was assayed using either Fujirebio’s Immunological Test System or Hotgen Biotech’s micro centrifugal column. In contrast, Le et al reported comparable diagnostic performance between the ASAP (AUC = 0.96) and GALAD (AUC = 0.95) models for overall and early-stage HCC detection; the biomarker detection platform was unspecified.²⁰ Studies by Demirtas, Thanapirom, and their colleagues have focused on model performance across etiological subgroups, with both teams measuring biomarkers using the Wako uTASWako i30 analyzer.^21,22 Both research teams noted that the GALAD model achieved higher sensitivity in non-viral HCC, but discrepancies emerged regarding viral liver disease patients: Demirtas et al found comparable performance between the two models, while Thanapirom et al observed superior efficacy with the ASAP model.

In our study, and consistent with some previous reports, no significant difference in diagnostic performance was found between the GALAD and ASAP models when analyzed on the same platform. However, significant performance differences emerged when the same model was applied across different platforms. For the ASAP model, the AUC on the Abbott platform (0.913) was significantly higher than that on the Hotgen platform (0.871, P = 0.0003). A similar pattern was observed for the GALAD model, with its AUC on the Abbott platform (0.901) being significantly greater than that on the Hotgen platform (0.871, P = 0.0001). Previous studies have primarily focused on the impact of model-specific differences or study population variations, while the critical variable of detection platform has been overlooked. This oversight may partially explain the discrepancies in previous reports regarding which model has superior diagnostic efficacy. This study clearly demonstrates that the detection platform exerts a significant influence on the diagnostic performance of both the GALAD and ASAP models. This finding highlights the need to consider detection platform differences in clinical practice and cross-study comparisons. It also provides valuable insights for the standardized application and accurate interpretation of HCC diagnostic models.

In addition, we systematically compared AFP and PIVKA-II between the Hotgen Biotech and Abbott Architect platforms using various statistical methods. Passing-Bablok regression identified a significant linear relationship for AFP (slope = 1.116, intercept = 0.221, P = 0.16), indicating proportional and systematic biases, whereas PIVKA-II showed no linearity between platforms (P < 0.001). Spearman analysis confirmed moderate correlations (correlation coefficient = 0.573 for AFP, 0.460 for PIVKA-II), but Bland-Altman plots revealed clinically significant biases (mean bias: AFP + 44.32%, PIVKA-II −92.02%). Notably, despite these analytical differences, ROC analyses demonstrated comparable diagnostic efficacy for HCC across platforms—with AFP’s AUC showing approximate consistency, while PIVKA-II had a minor inter-platform difference. This suggests that while absolute values are platform-dependent, the discriminative power remains stable However, it is recommended that the same detection system be used for the same patient throughout treatment and follow-up to avoid discrepancies due to platform differences.²⁷

The key strength of this study lies in its novel dual-platform comparison of both individual biomarkers (AFP, PIVKA-II) and composite models (GALAD, ASAP) for HCC diagnosis. Concurrently, it has been revealed that the detection platform employed for biomarker quantification exerts a significant impact on the diagnostic performance of these models. However, several limitations should be acknowledged: Firstly, tumor staging data for HCC patients were not obtained, making it impossible to assess the value of serum biomarkers and composite models for early HCC screening and diagnosis. Secondly, as a single-center study, it may be subject to selection bias. Thirdly, the study cohort exhibits class imbalance across different subgroups, and no specific correction strategies were implemented during data analysis. Future studies will expand the sample size of minority subgroups and adopt appropriate statistical methods to address this imbalance.