Diagnostic accuracy of tumor markers for hepatocellular carcinoma: a systematic review

Abstract

Background and aims The role of alphafetoprotein (AFP) in the diagnosis and surveillance of hepatocellular carcinoma (HCC) is getting smaller owing to the advances in imaging modalities. The aims of this study were to assess the diagnostic accuracy of tumor markers in small HCC and to find the optimal cutoff value of each tumor marker for efficient surveillance. Methods Studies in all languages were identified by searching MEDLINE from 1982 to 2002. Studies were included when they showed sensitivity and specificity for HCCs 5 cm or smaller and recruited only patients with chronic hepatitis or liver cirrhosis as control. We assessed diagnostic odds ratios (DORs) for the evaluation of diagnostic accuracy of tumor markers and positive likelihood ratios (LRs+) to find the optimal cutoff value. DORs and LRs+ were combined according to the random effect model. The summary receiver operating characteristics (ROC) curve was also assessed. Results Seventeen articles on three tumor markers—AFP, des-gamma-carboxyprothrombin (DCP), and Lens culinaris agglutinin-reactive fraction of AFP (AFP-L3)—were enrolled after full-text evaluation. AFP was inferior to DCP and AFP-L3 in both DOR (4.50 vs. 8.16 and 10.50) and area under the ROC curve (0.647 vs. 0.688 and 0.695). Optimal cutoff values that provide the best LR+ were 200 ng/ml for AFP, 40 mAU/ml for DCP, and 15% for AFP-L3. Conclusions Diagnostic accuracy of AFP in small HCC was substantially limited. Surveillance including other tumor markers with optimal cutoff value should be conducted to confirm the efficacy of the policy.

Introduction

Hepatocellular carcinoma (HCC) is a common worldwide malignancy, with the United States showing an increasing incidence rate [1, 2]. Approximately, 75–80% of primary liver cancers are attributable to persistent viral infections with either hepatitis B virus (HBV) or hepatitis C virus (HCV) [3–6], and the annual incidence of HCC with cirrhosis is 1–7% [7–11].

In the ordinary diagnostic process of HCC, a space-occupying lesion in the liver is first detected by imaging modalities such as ultrasonography and then confirmed by dynamic CT or MRI with contrast media. Typical HCC shows hypervascularity in the arterial phase and washout of contrast media in the portal-venous phase [12, 13]. The final diagnosis was made pathologically when a patient receives percutaneous biopsy, hepatic resection, or liver transplantation.

Alphafetoprotein (AFP) has served as a diagnostic test for HCC since the 1970s, when most patients with HCC were diagnosed at an advanced stage with clinical symptoms [14]. Concentrations higher than 500 ng/ml can be confirmatory in that situation. Nowadays quite a few small HCCs (e.g., 3 cm or smaller) can be detected owing to advances in imaging modalities, and it is known that significant numbers of small HCCs do not secrete a diagnostic level of AFP [15]. Furthermore, AFP levels are elevated both in patients with HCC and in those with chronic liver diseases, and there is a wide overlap between the two groups [16, 17]. Thus, the role of AFP as a diagnostic test is getting smaller.

Another use of AFP is in detecting asymptomatic HCC via periodic screening for high-risk patients. High positive predictive value with an appropriate cutoff value is mandatory for an effective surveillance program.

To date, many tumor markers have been proposed as a complement or substitute for AFP in the diagnosis of HCC, such as des-gamma-carboxyprothrombin (DCP) [18, 19], Lens culinaris agglutinin-reactive fraction of AFP (AFP-L3) [20, 21], and various cytokines [22–24]. The aim of this study is to compare the diagnostic accuracy of AFP in small HCC with other biomarkers and to find the optimal cutoff value of each tumor marker for efficient surveillance.

Methods

We conducted this study as part of a project to establish evidence-based guidelines for the diagnosis and treatment of HCC. This project is supported by the Ministry of Health, Labour and Welfare of Japan.

Study search protocol

First, we settled upon two research questions: (1) How accurate is each tumor marker at diagnosing small HCC? (2) What is the optimal cutoff value of each tumor marker for an effective screening strategy?

To identify relevant articles, we searched MEDLINE (from January 1982 to December 2002), using liver neoplasms and tumor markers as Medical Subject Headings terms (Appendix 1). Subsequently, we reran searches using more specific terms (Appendix 2). A search algorithm was constructed by experts at the International Medical Information Center (Shinanomachi, Shinjuku, Japan). We supplemented these sources by searching the Cochrane Collaboration Library and hand-searching bibliographies of systematic reviews and relevant original articles. Bibliographies were downloaded into a specially designed database made by FileMaker Ver. 5 (FileMaker, Inc., Santa Clara, CA, USA) for the project.

Inclusion criteria

Studies were included when they met all the following criteria: (1) the sensitivity and specificity were described or could be calculated from tables or figures, (2) the maximum size of nodules was 5 cm or smaller, or sensitivity of nodules smaller than 5 cm could be calculated, and (3) only patients with chronic hepatitis or cirrhosis were recruited as control. We then selected tumor markers for which at least three articles were available. We excluded any meeting abstracts not accompanied by full articles, and other incomplete reports.

Two authors independently reviewed the article titles and abstracts identified by the search, evaluated each study for inclusion, and retrieved potentially eligible articles for full-text evaluation. Any discrepancies were settled by a third author.

Data extraction

Two authors independently extracted data from each eligible article. A 2 × 2 table was reconstructed for every tumor marker in the article. When tables could be reconstructed for 2 or more upper limits of tumor size (e.g., 3 and 5 cm), we adopted the larger size up to 5 cm. When articles provided 2 or more tables for 2 or more cutoff values (e.g., 20 and 100 ng/ml for AFP), all available tables were reconstructed separately.

Potential confounders and quality assessment

We considered eight variables to be potential confounders for explaining heterogeneity and interstudy variability: (1) proportion of chronic hepatitis and cirrhosis patients; (2) proportion of patients with hepatitis B and C; (3) year of publication; (4) study design: cohort or case–control study; (5) proportion of histologically proven HCC patients; (6) blinding: whether final diagnosis of HCC was performed independently from the test result; (7) consecutive recruitment of patients; and (8) existence of verification bias: whether only patients with positive test results received the reference standard. The latter four variables were introduced to assess the quality of articles according to the guidelines of Irwig et al. [25]. All these variables were defined a priori; two authors independently evaluated the articles and a third author independently settled discrepancies.

All data were input to a standardized form for assessing the characteristics of enrolled articles.

Statistical analysis

Diagnostic odds ratio

To answer the first research question (i.e., how is the accuracy of each tumor marker at diagnosing small HCC), we assessed the diagnostic odds ratio (DOR) of each tumor marker that represents the comprehensive ability of a diagnostic test according to the following formula. Since there is an inverse association between sensitivity and specificity that differs according to test thresholds, it is inappropriate to estimate their means separately [26].

Assessment for potential confounders

To examine the factors associated with variation in the DOR, a regression model was formulated and tested. We applied linear regression analysis with log DOR as a dependent variable and the previously described eight factors as independent variables. The P value threshold for statistical significance was set at 0.05.

Summary receiver operating characteristic curve

In addition, we applied another approach to combine the results of primary studies, that is, to draw a summary receiver operating characteristic (SROC) curve according to Moses et al. [27]. This model hypothesizes that there is a linear relationship between

and

We applied weighted linear regression analysis for each tumor marker and drew SROC curves. We also calculated the area under the curve (AUC) and its standard error (SE) according to the method of Walter [28].

Assessment for publication bias

To assess the presence of publication bias, we created funnel plots for each diagnostic test. We plotted the inverse of the standard error of the natural logarithm of the DOR against the natural logarithm of the DOR. Its asymmetry was tested by significance test using the linear regression method suggested by Egger et al. [29]. In the regression, the standardized effect, defined as the effect divided by its standard error, is regressed against the precision of the effect, defined as the inverse of the standard error:

The intercept α provides a quantitative measure of the asymmetry and is of major interest. The more the intercept deviates from zero, the more pronounced the asymmetry. Negative values of α will indicate that less precise studies have a more pronounced effect than more precise studies, suggesting publication bias. On the other hand, positive intercept does not suggest selection bias rather than heterogeneity of included studies. A P value less than 0.1 of zero intercept is considered statistically significant.

Positive likelihood ratio

To answer the second question (i.e., what is the optimal cutoff value), we assessed positive likelihood ratio (LR+) on the assumption that a positive result may lead to a confirming test (e.g., dynamic CT). LR+, which is always discussed with negative likelihood ratios, was calculated on the following formula and express how much more frequent the positive results is among subjects with disease than among subjects without disease, independent from pretest probability or disease prevalence.

LRs+ above 10 are considered to provide a strong evidence to rule in diagnoses respectively in most circumstances [30]. Positive predictive values (= posttest probabilities) can be calculated on the following formula.

When the sensitivity or specificity equaled zero, we added 0.5 to each cell of the original 2 × 2 table to avoid dividing by zero. We calculated the pooled DORs and LRs and their 95% confidence intervals (CIs) for random effect models on the basis of DerSimonian and Laird [31]. We assessed the heterogeneity of the odds ratio by a homogeneity chi-square test.

We used MetaDiSc statistical software [32] and Comprehensive Meta Analysis Ver. 1.0 (Biostat, Englewood, NJ, USA) for the analysis of LRs and DOR, and S-PLUS 2000 (MathSoft Inc., Seattle, WA, USA) for regression analysis and drawing SROC curves.

Role of the funding sources

The funding sources had no role in the choice of topic; the collection, analysis, or interpretation of the data; or the decision to submit the manuscript for publication.

Results

The primary MEDLINE search identified 2,685 articles potentially relevant to the topic. After reviewing titles and abstracts of these articles, the following three tumor markers were found in at least three articles that provided sensitivity and specificity: AFP, DCP, and AFP-L3. We ran another MEDLINE search using more specific terms (Appendix 2) and identified 985 additional articles to be reviewed. We also identified two systematic reviews on detecting HCC in patients with chronic hepatitis C [33, 34] and checked their bibliographies.

After reviewing titles and abstracts, 54 articles were retrieved for full-text evaluation. Thirty-seven of them were excluded: 32 did not provide sensitivity for tumors less than 5 cm in diameter [15, 19, 21, 22, 35–62], 3 recruited patients without chronic hepatitis [46, 52, 63], and 5 articles did not provide specificity [56, 64–67]. Finally, 17 articles were included in our study (Fig. 1, Table 1) [68–84]. Eleven studies included data on more than one tumor marker [70, 71, 73–77, 79–81, 83], and 4 of them provided sensitivity and specificity of a combination of two tumor markers [70, 74, 77, 79]. As 8 studies provided sensitivity and specificity with more than one cutoff value [68, 70, 71, 74, 76, 80, 83, 85], we could evaluate 50 DORs in total.

Fig. 1.

Study inclusion flow diagram

Table 1.

Characteristics of included studies

Author (reference)	Diagnostic test	Study design	Country	Patients with HCC				Control
				n	Etiology	Characteristics of HCC	Modalities of diagnosis	n	Etiology	Characteristics of patients
Cottone et al. [68]	AFP (RIA)	Co	Italy	30	7% with HBV, 73% with HCV	100% ≤5 cm	100% by pathology	147	12% with HBV, 71% with HCV	100% with LC
Fujiyama et al. [69]	AFP (RIA), DCP (EIA)	CC	Japan	120	NR with HBV and HCV	34% ≤5 cm, 10% ≤2 cm	NR	334	NR with HBV and HCV	41% with LC, 59% with CH
Ikoma et al. [70]	AFP (EIA), DCP (EIA)	CC	Japan	63	14% with HBV, 70% with HCV	56% ≤3 cm, 20% ≤2 cm	33% by pathology, 67% by imaging	188	22% with HBV, 84% with HCV	24% with LC, 76% with CH
Kasahara et al. [71]	AFP (EIA), DCP (EIA)	CC	Japan	112	NR with HBV and HCV	15% ≤5 cm	100% by imaging	403	NR with HBV and HCV	28% with LC, 72% with CH
Maringhin et al. [72]	AFP (RIA)	Co	Italy	146	80% with HBV	38% ≤5 cm	50% by pathology	217	65% with HBV	100% with LC
Mita et al. [73]	AFP (EIA), DCP (sEIA), AFP-L3 (EP)	CC	Japan	91	17% with HBV, 70% with HCV	47% ≤2 cm	100% by pathology or imaginge	57	32% with HBV, 40% with HCV	100% with LC
Nomura et al. [74]	AFP (EIA), DCP (EIA)	CC	Japan	27	7% with HBV, 89% with HCV	100% ≤3 cm	100% by pathology	101	NR with HBV and HCV	68% with LC, 32% with CH
Nomura et al. [75]	AFP (EIA), DCP (EIA, ECLIA, IRMA), AFP-L3 (EP)	CC	Japan	36	NR with HBV and HCV	100% ≤3 cm, 53% ≤2 cm	100% by pathology	49	NR with HBV and HCV	100% with LC
Oka et al. [76]	AFP (variousa), AFP-L3 (EP)	CC	Japan	388	9% with HBV, 78% with HCV	32% ≤2 cm	26% by pathology, 74% by imaging	212	100% with HBV or HCV	100% with CH or LC
Saitoh et al. [78]	AFP (RIA), DCP (ABC)	CC	Japan	115	17% with HBV, 80% with HCV	100% ≤2 cm	72% by pathology, 28% by imaging		NR with HBV and HCV	45% with LC, 55% with CH
Sassa et al. [77]	AFP (RIA), DCP (EIA), AFP-L3 (EP)	CC	Japan	61	15% with HBV, 84% with HCV	100% ≤2 cm	92% by pathology	134	17% with HBV, 79% with HCV	56% with LC, 44% with CH
Shimauchi et al. [79]b	AFP (RIA), DCP (sEIA), AFP-L3 (EP)	Co	Japan	21	14% with HBV, 76% with HCV	95% ≤3 cm, 67% ≤2 cm	100% by pathology	57	14% with HBV, 79% with HCV	100% with LC
Shiraki et al. [80]	AFP (RIA), DCP (EIA), AFP-L3 (EP)	Co	Japan	51	20% with HBV, 80% with HCV	80% ≤3 cm, 57% ≤2 cm	88% by pathology, 12% by imaging	21	24% with HBV, 76% with HCV	100% with LC
Suehiro et al. [81]	AFP (NR), DCP (EIA)	CC	Japan	185	NR with HBV and HCV	39% ≤6 cm, 21% ≤3 cm	49% by pathologyf	85	NR with HBV and HCV	100% with LC
Taketa et al. [82]c	AFP (RIA/EIA), AFP-L3 (EP)	CC	Japan	167	21% with HBV, 50% with HCV	87% ≤5 cm, 68% ≤3 cm	51% by pathology, 49% by imaging	181	33% with HBV, 44% with HCV	69% with LC, 31% with CH
Tanabe et al. [83]d	AFP (RIA), DCP (EIA)	CC	Japan	571	NR with HBV and HCV	32% ≤5 cm, 8% ≤2 cm	100% by pathology or imaging	28	NR with HBV and HCV	100% with LC
Tsai et al. [84]	AFP (RIA), DCP (EIA, SC)	CC	Taiwan	39d	62% with HBV, NR with HCV	100% ≤5 cm, 54% ≤2 cm	100% by pathology or imaging	54	NR with HBV and HCV	52% with LC, 48% with CH

^aAccording to each hospital’s available kit

^bSpecificity for AFP was not reported

^cSensitivity was only reported on AFP-L3

^dDCP was reported in only 31 patients with HCC

^eExact percentage was not reported

^f100% pathologically proven for HCCs smaller than 3 cm in diameter

Abbreviations: ABC, avidin–biotin complex method; AFP, alphafetoprotein; AFP-L3, Lens culinaris agglutinin-reactive fraction of AFP; CC, case–control study; CH, chronic hepatitis; Co, cohort study; DCP, des-gamma-carboxyprothrombin; ECLIA, electrochemiluminescence assay; EIA, enzyme immunoassay; sEIA, sensitive enzyme immunoassay; EP, electrophoresis; HBV, hepatitis B virus; HCV, hepatitis C virus; LC, liver cirrhosis; NR, not reported; RIA, radioimmunoassay

Sensitivity, specificity, and DOR

Diagnostic odds ratios calculated for each tumor marker were stratified by cutoff values and combined (Fig. 2). Combined odds ratios (95% CI) for AFP, DCP, and AFP-L3 according to the random effects model were 4.50 (2.82–7.18), 8.16 (4.95–13.48), and 10.50 (5.39–20.49), respectively. There was significant heterogeneity among studies in AFP and DCP. The DORs of the combinations of two markers are shown in Table 2.

Fig. 2.

Forest plots of DOR of AFP (a), DCP (b), and AFP-L3 (c). Diagnostic odds ratios were combined using cutoff value as stratification factor. AFP, alphafetoprotein; AFP-L3, Lens culinaris agglutinin-reactive fraction of AFP; DCP, des-gamma-carboxyprothrombin

Table 2.

Diagnostic odds ratio of the combination of two diagnostic tests

Author (reference)	Test	Cutoff value	Sensitivity	Specificity	DOR (95% CI)
Ikoma et al. [70]	AFP + DCP	20 ng/ml, 16 mAU/ml	0.83	0.84	25.46 (9.73–66.6)
Sassa et al. [77]	AFP + DCP	200 ng/ml, 40 mAU/ml	0.48	0.99	59.81 (13.56–263.8)
Nomura et al. [74]	AFP-L3 + DCP	10%, 40 mAU/ml	0.42	0.90	6.29 (2.02–19.6)
Shimauchi et al. [79]	AFP-L3 + DCP	10%, 40 mAU/ml	0.67	0.90	17.0 (4.92–58.8)
Sassa et al. [77]	AFP + AFP-L3	200 ng/ml, 10%	0.25	0.99	43.4 (5.57–337.5)

Abbreviations: AFP, alphafetoprotein; AFP-L3, Lens culinaris agglutinin-reactive fraction of AFP; DCP, des-gamma-carboxyprothrombin

Assessment for potential confounders

Table 3 shows the results of regression analysis for potential confounders. We excluded blinding in the reference standard from the analysis because none of the included studies provided sufficient information regarding blinding. A significant decrease in the odds ratio of DCP (slope = −0.551, P = 0.02) was also observed when all patients had pathologically proven HCC. The diagnostic accuracy of AFP significantly deteriorates when the majority of patients were HCV-positive (slope = −1.038, P = 0.03).

Table 3.

Results of regression analysis for potential confounders

Variable	AFP		DCP		AFP-L3
	Slope	P	Slope	P	Slope	P
Liver functiona	0.382	0.89	−0.183	0.70	NA	NA
Etiologyb	−1.038	0.03	NA	NA	0.044	0.94
Publication yearc	−0.573	0.07	0.147	0.57	−0.404	0.38
Study designd	0.194	0.56	0.009	0.98	−0.190	0.69
Diagnosise	−0.618	0.07	−0.551	0.02	−0.464	0.25
Consecutive recruitmentf	0.160	0.63	−0.093	0.75	−0.190	0.69
Verification biasg	0.845	0.19	NA	NA	NA	NA

^a0, less than 50% of patients had liver cirrhosis; 1, more than 50% of patients had liver cirrhosis

^b0, less than 50% of patients were HCV-positive; 1, more than 50% of patients were HCV-positive

^c0, year of publication 1994 or earlier; 1, year of publication 1994 or later

^d0, case–control study; 1, cohort study

^e0, not all patients were diagnosed by pathology; 1, all patients were diagnosed by pathology

^f0, nonconsecutive recruitment of patients; 1, consecutive recruitment of patients

^g0, present; 1, absent

Values in boldface have statistical significance

Abbreviations: AFP, alphafetoprotein; AFP-L3, Lens culinaris agglutinin-reactive fraction of AFP; DCP, des-gamma-carboxyprothrombin; NA, not applicable because all studies were categorized into a single group

SROC analysis

Figure 3 shows the scatterplot of (1—specificity) against sensitivity in each study for the diagnostic test and SROC curves. AFP-L3 and DCP showed higher diagnostic accuracy than AFP (Table 4). We also plotted the combined sensitivity and specificity of AFP + DCP, AFP-L3 + DCP, and AFP + AFP-L3 provided in four studies [70, 74, 77, 79].

Fig. 3.

SROC curves for 3 diagnostic tests for HCC. Solid line and circles, AFP; dashed line and triangles, DCP; dotted line and diamonds, AFP-L3; asterisks, AFP + DCP; pluses, DCP + AFP-L3; cross, AFP + AFP-L3. AFP, alphafetoprotein; AFP-L3, Lens culinaris agglutinin-reactive fraction of AFP; DCP, des-gamma-carboxyprothrombin; HCC, hepatocellular carcinoma; SROC, summary receiver operating characteristics

Table 4.

Results of SROC analyses for three diagnostic tests for HCC

Test	Intercept	Slope	AUC	SE (AUC)
AFP	0.940	−0.321	0.647	0.027
DCP	1.207	−0.230	0.688	0.083
AFP-L3	1.266	−0.262	0.695	0.166

Abbreviations: AFP, alphafetoprotein; AFP-L3, Lens culinaris agglutinin-reactive fraction of AFP; AUC, area under the curve; DCP, des-gamma-carboxyprothrombin; HCC, hepatocellular carcinoma; SE, standard error; SROC, summary receiver operating characteristics curve

Assessment for publication bias

Funnel plots of three tumor markers showed asymmetry (Appendix Fig. A1). Intercepts and P values of AFP, DCP, and AFP-L3 were 3.59 (P = 0.0003), 5.30 (P = 0.0001), and 6.78 (P = 0.0047), respectively. As previously described, a positive intercept does not suggest publication bias but rather heterogeneity of included studies.

Fig. A1.

Funnel plots of DOR of AFP (a), DCP (b), and AFP-L3 (c). Inverse of standard error of natural logarithm of the odds ratio against natural logarithm of the DOR is plotted. The area of the circle is proportional to the inverse of standard error of natural logarithm of the odds ratio. AFP, alphafetoprotein; AFP-L3, Lens culinaris agglutinin-reactive fraction of AFP; DCP, des-gamma-carboxyprothrombin; DOR, diagnostic odds ratio

Positive likelihood ratio

LRs+ of each tumor marker stratified by cutoff values were calculated and combined (Fig. 4). To assess the optimal cutoff value, we compared LRs+ with the following cutoff values since there were three or fewer studies with other cutoff values: 20 and 200 ng/ml of AFP, 40 and 100 mAU/ml of DCP, and 10% and 15% of AFP-L3. LRs+ with a cutoff value of 200 ng/ml of AFP (5.85, 95% CI: 1.49–22.93), 40 mAU/ml of DCP (12.60, 6.65–23.87), and 15% of AFP-L3 (13.10, 3.89–44.97) were better than those with, respectively, 20 ng/ml of AFP (2.45, 1.74–3.45), 100 mAU/ml of DCP (4.91, 2.43–9.91), and 10% of AFP-L3 (4.89, 2.77–8.61).

Fig. 4.

Forest plots of positive likelihood ratio of AFP (a), DCP (b), and AFP-L3 (c). Likelihood ratios were combined on a cutoff value basis. AFP, alphafetoprotein; AFP-L3, Lens culinaris agglutinin-reactive fraction of AFP; DCP, des-gamma-carboxyprothrombin; LR, likelihood ratio

Discussion

Our literature search identified two potential candidates for substituting or complementing AFP in the diagnosis of HCC, namely, DCP and AFP-L3. DCP, also known as prothrombin induced by vitamin K absence-II (PIVKA-II), is an abnormal prothrombin protein that is present at higher levels in the serum of HCC patients. Since the report by Liebman et al. [18], DCP has been recognized as not only a highly specific marker for HCC but also a predictor of prognosis of HCC patients [86, 87]. AFP-L3 is a fucosylated variant of AFP that reacts with Lens culinaris agglutinin A and can differentiate an increase of AFP due to HCC from that due to benign liver disease [20, 21, 82].

In this study, the diagnostic accuracy of AFP was found to be inferior to that of DCP and AFP-L3 in both meta-analysis with random effects model and SROC analysis. It is apparent that the inferiority of AFP in DORs is due to its lack of specificity as compared with the other two markers. The false-positive rate of AFP with a cutoff value of 20 ng/ml was 0.14–0.32. It should be stated that the role of AFP as a diagnostic test is almost over in the era of advanced imaging technologies. Instead, the other two markers, which showed superior specificity around 0.95, will play a role as a confirmatory test.

During the literature-searching process, we identified two systematic reviews of the diagnostic accuracy of AFP [33, 34]. Both of them focused on HCV-positive patients based on the concept that the natural history of hepatitis C is different from that of hepatitis B [6, 40]. Elevated AFP has been observed more frequently in patients with chronic hepatitis B than C [46, 57], and the present results of meta-regression on the etiology of liver disease may support this finding. Diagnostic accuracy deteriorated when the majority of patients were HCV positive. In contrast, we did not encounter this phenomenon with AFP-L3.

We also assessed potential confounders other than liver disease etiology. The presence or absence of liver cirrhosis, which strongly influences the risk of HCC in HCV-positive patients, did not affect the DOR. The year of publication, which is associated with improvement in the diagnostic accuracy of the reference standard (e.g., dynamic CT), also did not show significant influence on the DOR.

Research on randomized trials and observational studies has suggested that the major reason for not publishing is that investigators do not submit studies with negative results for publication [88]. Publication bias is more serious in observational studies than in interventional studies. Most of the studies on DCP and AFP-L3 included in the current analysis assessed the diagnostic accuracy in comparison with AFP. Thus, there might also be publication bias when true results suggesting inferiority of DCP or AFP-L3 to AFP were possibly left unpublished, but funnel plot analysis could obviously not reveal the existence of such publication bias.

Study design features significantly affect the entire results of a study on diagnostic tests [89–91]. In a case–control study, to prevent the possibility of selection bias, patients should be recruited consecutively from a relevant clinical population covering the spectrum of disease that is likely to be encountered in the current or future use of the test. In surveillance as a cohort study consisting of periodical tests, verification bias exists when the decision to perform the reference test is based on the result of the test under examination. The results of regression analysis showed no significant difference between well-designed studies and others. This might be because the number of quality studies was too small.

As it is inappropriate to conduct surveillance for HCC without ultrasound in a high-risk population [15, 92], the major issue is whether to combine ultrasound and tumor markers in the surveillance. The fact that DCP and AFP-L3 showed LRs+ of larger than 10, which are interpreted as conclusive, make it quite reasonable to consider inclusion of those markers in surveillance. In fact, DCP with a cutoff value of 40 mAU and AFP-L3 with a cutoff value of 15% were adopted in the Japanese guideline for the diagnosis and treatment of HCC [93]. The higher diagnostic accuracy shown in the combination of two tumor markers, rather than by AFP alone, suggests that inclusion of two or three markers may improve the efficiency of surveillance programs.

The aim of this study is not to directly assess the effectiveness of surveillance programs for HCC but to find potential candidates to be included in the surveillance. The effectiveness of surveillance depends on various factors—disease prevalence, identification of high-risk populations, interval of diagnostic tests, type of confirmation tests, and treatment modality when disease is confirmed—and should also finally be assessed in terms of cost-effectiveness. Inclusion of new diagnostic tests in surveillance programs may make it possible to detect additional small HCCs. However, it should be verified whether improvement of effectiveness is always justified by increased cost. In conclusion, surveillance of HCC in high-risk populations that includes DCP and/or AFP-L3 should be conducted and assessed on the basis of acceptable cost-effectiveness.

Abbreviations

AFP

Alphafetoprotein

AFP-L3

Lens culinaris agglutinin-reactive fraction of AFP

AUC

Area under the curve

CI

Confidence interval

DCP

Des-gamma-carboxyprothrombin

DOR

Diagnostic odds ratio

HBV

Hepatitis B virus

HCC

Hepatocellular carcinoma

HCV

Hepatitis C virus

LR+

Positive likelihood ratio

ROC

Receiver operating characteristics

SE

Standard error

Appendix

Appendix 1.

Primary research

	Key words	No. of articles
1	Liver Neoplasms [MAJR] AND 1982:2002 [DP]	36,650
2	1 AND (Human [MeSH] OR hominidae [MeSH])	28,685
3	2 AND Tumor Markers, Biological [MeSH]	2,685

Appendix 2.

Secondary research

	Key words	No. of articles
1	Liver Neoplasms [MAJR] AND 1982:2002 [DP]	36,650
2	1 AND (Human [MeSH] OR hominidae [MeSH])	28,685
4	2 AND (alpha-Fetoproteins [MeSH] OR alpha-fetoprotein [TW] OR alphafetoprotein [TW] OR AFP [TW])	2,587
5	2 AND (prothrombin [MeSH] OR des-gamma-carboxyprothrombin [TW] OR des-gamma-carboxy-prothrombin [TW] OR desgammacarboxyprothrombin [TW] OR DCP [TW] OR PIVKA [TW])	243
6	4 OR 5	2,650
7	3 OR 6	3,670