Diagnostic performance of circulating tumor DNA as a minimally invasive biomarker for hepatocellular carcinoma: a systematic review and meta-analysis

Search strategy

Two authors (Jiajie Li and Yanqing Lv) independently conducted a comprehensive search for relevant articles in PubMed, Web of Science, Embase, and Cochrane Library. The query terms were as follows: “circulating tumor DNA” OR “circulating DNA” OR “ctDNA” OR “plasma DNA” OR “serum” DNA” OR “blood DNA” and “Liver Neoplasms” OR “Neoplasms, Hepatic” OR “Neoplasms, Liver” OR “Liver Neoplasm” OR “Neoplasm, Liver” OR “Hepatic Neoplasms” OR “Hepatic Neoplasm” OR “Neoplasm, Hepatic” OR “Cancer of Liver” OR “Hepatocellular Cancer” OR “Cancers, Hepatocellular” OR “Hepatocellular Cancers” OR “Hepatic Cancer” OR “Cancer, Hepatic” OR “Cancers, Hepatic” OR “Hepatic Cancers” OR “Liver Cancer” OR “Cancer, Liver” OR “Cancers, Liver” OR “Liver Cancers” OR “Cancer of the Liver” OR “Cancer, Hepatocellular” AND “diagnosis” OR “sensitivity” OR “specificity” OR “accuracy”. The language of all the articles was limited to English and the experiment target was limited to human beings. We also manually screened references from included articles and related reviews in order to expand the search.

Inclusion and exclusion criteria

Studies that met the following criteria were eligible for inclusion: (a) diagnostic accuracy of ctDNA was evaluated in plasma or serum; (b) sufficient data were acquired or could be calculated from the raw data (e.g., SEN, SPE, true positives [TP], false positives [FP], true negatives [FN], and false negatives [FN]); (c) CtDNA markers were used for the first diagnosis of HCC, not for the diagnosis of recurrence and metastasis; (d) controls were cancer-free adults; and (e) articles were published in English and experiment sources were human beings. The main exclusion criteria were as follows: (a) reviews, conference abstracts, meta-analysis, editorials, letters, reply, case report, commentary, short survey, notes, research highlight, and duplicate publications; (b) the sample size of studies was less than 10; (c) we failed to obtain the full text; (d) overlapping publications that included the same population and gene; and (e) there were multiple genes or gene models.

Data extraction & quality assessment

The following data from the eligible data were extracted by two reviewers independently: the first authors’ name, publication year, country or region of origin, study design, sampling time, inspection method, assay indicator, cut-off values, number of participants, and SP, SE, TP, FP, FN, and TN, which were given directly or could be calculated by raw data. Also, based on the revised Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2), the quality of studies was assessed and rated by two authors independently, and the risk of bias and applicability concerns was categorized as low, unclear, or high. If the answer to all the iconic questions in a range was “yes”, then the risk of bias can be assessed as low, if the answer to any of the information questions is “no”, then the risk of bias was judged as “high”, and when there was not enough information, we defined it as unclear risk. Divergences were discussed together to reach a consensus and if an agreement was unable to be met, Huifan Ji made a judgement.

Statistical analysis

We used Stata software (version 16.0; Stata Corporation, TX, USA), meta-disc 1.4, to perform this diagnostic meta-analysis. The pooled sensitivity and specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR), and corresponding 95% confidence interval (95% CI) were calculated using a bivariate generalized linear mixed model to evaluate the test accuracy. Simultaneously, we plotted hierarchical summary receiver operating characteristics (HSROC) and utilized lambda as well as the area under the curve (AUC) which created by summary receiver operating characteristic (SROC) curves to estimate the diagnostic value of ctDNA. The logit estimates of SEN, SPE, and respective variances were used to construct the HSROC curves. Fagan’s nomogram was applied to interpret the clinical utility of ctDNA for HCC (Anthony, 2007). The Spearman correlation coefficient and its corresponding P value were used to identify the presence of the threshold effect. The threshold effect was considered to exist when the P value was lower than 0.05. If heterogeneity resulted from the non-threshold effect, I² and χ2 were used to evaluate heterogeneity between the enrolled articles. I² >50% or P < 0.05 for χ2 suggested significant heterogeneity (Borenstein et al., 2017). Subgroup and meta regression analyses were used to explore the sources of heterogeneity. Finally, we formulated Deek’s funnel plot to examine potential publication bias (Deeks, Macaskill & Irwig, 2005), and trim and fill analysis was applied to assess the effect of bias on the pooled estimate once the publication bias existed.

Study characteristics

Figure 1 shows the process of literature retrieval and inclusion. Through our search strategy, a total of 879 publications were enrolled from different databases outlined in the Materials and methods section. After deleting 216 duplicated articles, 487 articles were excluded after screening their titles and abstracts (Fig. 1). Eventually, the full text of 59 eligible articles (60 studies) was incorporated into this meta-analysis (Huang et al., 2003; Wang et al., 2021; Wong et al., 2000; Wong et al., 2003; Chu et al., 2004; Lin et al., 2005; Yeo et al., 2005; Iizuka et al., 2006; Ren et al., 2006; Wang et al., 2006; Zhang et al., 2007; Chan et al., 2008; Chang et al., 2008; Igetei et al., 2008; Ahmed et al., 2010; El-Shazly et al., 2010; Hu et al., 2010; Huang et al., 2011; Iizuka et al., 2011; Yang et al., 2011; Chen et al., 2012; Huang et al., 2012; Mohamed et al., 2012; Chen et al., 2013; Sun et al., 2013; Zhang et al., 2013; Han et al., 2014; Huang et al., 2014; Ji et al., 2014; Kuo et al., 2014; Li et al., 2014; Yang et al., 2014; Dong et al., 2015; Huang et al., 2015; Ramadan et al., 2015; Teng et al., 2016; Dong et al., 2017; Hu et al., 2017; Mansour et al., 2017; Tian et al., 2017; Wu et al., 2017; Gai et al., 2018; Li et al., 2018; Oussalah et al., 2018; Wei et al., 2018; Linlin et al., 2018; Bai et al., 2019; Kisiel et al., 2019; Marchio et al., 2019; Pasha, Mohamed & Radwan, 2019; El-Bendary et al., 2020; He et al., 2020; Kotoh et al., 2020; Li et al., 2020a; Liu et al., 2020; Qian et al., 2020; Akuta et al., 2021; Lewin et al., 2021; Xie et al., 2021).

Baseline characteristics

Table S1 summarizes the characteristics of all 59 papers (n = 92). All the included studies included quantitative analysis to measure ctDNA concentration and single-gene methylation concentration (n = 19) (Wang et al., 2021; Iizuka et al., 2006; Ren et al., 2006; Chan et al., 2008; El-Shazly et al., 2010; Iizuka et al., 2011; Yang et al., 2011; Chen et al., 2012; Huang et al., 2012; Mohamed et al., 2012; Chen et al., 2013; He et al., 2020; Kotoh et al., 2020; Li et al., 2020a; Xie et al., 2021; Mansour et al., 2017; Gai et al., 2018; Linlin et al., 2018; El-Bendary et al., 2020) and qualitative analysis of tumor-specific ctDNA single gene-mutation and methylation (n = 40) (Wong et al., 2000; Wong et al., 2003; Chu et al., 2004; Lin et al., 2005; Yeo et al., 2005; Wang et al., 2006; Zhang et al., 2007; Chang et al., 2008; Igetei et al., 2008; Ahmed et al., 2010; Hu et al., 2010; Huang et al., 2011; Sun et al., 2013; Zhang et al., 2013; Han et al., 2014; Huang et al., 2014; Ji et al., 2014; Kuo et al., 2014; Li et al., 2014; Yang et al., 2014; Dong et al., 2015; Huang et al., 2015; Ramadan et al., 2015; Teng et al., 2016; Dong et al., 2017; Hu et al., 2017; Tian et al., 2017; Wu et al., 2017; Li et al., 2018; Oussalah et al., 2018; Wei et al., 2018; Bai et al., 2019; Kisiel et al., 2019; Marchio et al., 2019; Pasha, Mohamed & Radwan, 2019; Liu et al., 2020; Qian et al., 2020; Akuta et al., 2021; Lewin et al., 2021). Of these papers, 18 articles evaluated the diagnostic performance of ctDNA combined with AFP assay in HCC (Wang et al., 2021; Chan et al., 2008; Huang et al., 2012; Chen et al., 2013; Han et al., 2014; Huang et al., 2014; Kuo et al., 2014; Li et al., 2014; Yang et al., 2014; Dong et al., 2015; Teng et al., 2016; Hu et al., 2017; Tian et al., 2017; Pasha, Mohamed & Radwan, 2019; Kotoh et al., 2020; Li et al., 2020a; Liu et al., 2020; Qian et al., 2020). In the qualitative analysis subgroup, our study enrolled a total population of 3,072 HCC patients and 3,413 control individuals (2,064 patients with benign liver disorders or liver cirrhosis, 1,001 healthy volunteers, and 308 non-cancer controls in a combined group of patients with benign liver disorders and healthy controls). Patients with chronic hepatitis, benign hepatic lesions, and cirrhosis were selected as the control group in 13 articles (Chu et al., 2004; Wang et al., 2006; Chang et al., 2008; Ahmed et al., 2010; Huang et al., 2011; Huang et al., 2014; Ji et al., 2014; Huang et al., 2015; Oussalah et al., 2018; Kisiel et al., 2019; Qian et al., 2020; Akuta et al., 2021; Lewin et al., 2021) , while eight articles with only healthy control groups were chosen in this subgroup (Yeo et al., 2005; Zhang et al., 2007; Igetei et al., 2008; Hu et al., 2010; Zhang et al., 2013; Kuo et al., 2014; Bai et al., 2019; Marchio et al., 2019) . Other articles combined healthy volunteers and chronic hepatitis or liver cirrhosis patients as the control group. The majority of articles were conducted in Asia (n = 31) (Huang et al., 2003; Wong et al., 2000; Wong et al., 2003; Chu et al., 2004; Lin et al., 2005; Yeo et al., 2005; Wang et al., 2006; Zhang et al., 2007; Chang et al., 2008; Hu et al., 2010; Huang et al., 2011; Sun et al., 2013; Zhang et al., 2013; Han et al., 2014; Ji et al., 2014; Kuo et al., 2014; Li et al., 2014; Yang et al., 2014; Dong et al., 2015; Huang et al., 2015; Teng et al., 2016; Dong et al., 2017; Hu et al., 2017; Tian et al., 2017; Wu et al., 2017; Li et al., 2018; Wei et al., 2018; Bai et al., 2019; Liu et al., 2020; Qian et al., 2020; Akuta et al., 2021), five in Africa (Igetei et al., 2008; Ahmed et al., 2010; Ramadan et al., 2015; Marchio et al., 2019; Pasha, Mohamed & Radwan, 2019), three in America (Huang et al., 2014; Kisiel et al., 2019; Lewin et al., 2021) , and one in Europe (Oussalah et al., 2018). A total of 37 studies looked at ctDNA methylation, while three evaluated single-gene mutation. In terms of study type, seven studies were retrospective, four were prospective, and the rest (n = 31) did not clearly state the study design. Among those factors known at the time of collection, sampling time was either before treatment or surgery (n = 19), and samples were obtained from plasma (n = 15), serum (n = 24), or plasma/serum (n = 1). There were 23 articles with sample sizes ≥100, while the remaining sample sizes<100.

In the quantitative analysis subgroup, there were a total of 1,446 HCC patients and 1,835 non-cancer control participants (966 patients, 650 healthy volunteers, and 219 participants in a mixed benign liver disorders and healthy control group). Patients with liver cirrhosis, HCV infection, and HBV infection were chosen as the control group in nine publications (Iizuka et al., 2006; El-Shazly et al., 2010; Iizuka et al., 2011; Chen et al., 2012; Huang et al., 2012; Mohamed et al., 2012; Mansour et al., 2017; Linlin et al., 2018; Kotoh et al., 2020) , and only one study singly chose healthy volunteers as control group (Chen et al., 2013). Among these studies, all but four articles were conducted in Egypt (El-Shazly et al., 2010; Mohamed et al., 2012; Mansour et al., 2017; El-Bendary et al., 2020) , and the others were all studied in Asia (n = 15). Eight studies evaluated the performance of ctDNA concentration as a diagnostic indicator, 10 studies chose single gene methylation concentration, and one selected hTERT concentration. As for study design, the majority of them were not described clearly (n = 16), but two were prospective studies, and one was a retrospective study. Nine studies had a known time of collection before treatment and surgery. Samples were obtained from plasma (n = 9) and serum (n = 10). Twelve publications had sample sizes ≥100 and the rest had sample sizes smaller than 100. The assay methods used to measure the concentrations of ctDNA were real-time quantitative polymerase chain reaction (RT-qPCR) (n = 5), droplet digital PCR DNA (Dd-PCR) (n = 3), quantitative methylation specific PCR (Q-MSP) (n = 4), quantitative PCR (QPCR) (n = 4), ultraviolet transilluminator (n = 1), and Qubit dsDNA (n = 1). For ctDNA mixed with the AFP subgroup, 18 papers were studied with 1,790 HCC patients and 1,614 non-cancer participants.

Quality assessment

As shown in Fig. 2, the majority of enrolled studies included four criteria: patient selection, index test assessment, reference standard assessment, and a flow and timing assessment. Publications were judged as having high risk in one field, which was consider as having an overall high risk of bias. However, two studies were excluded due to the high risk of bias in patient selection and reference concerns (Lleonart et al., 2005; Marchio et al., 2018) , and four articles (Chan et al., 2008; Ahmed et al., 2010; Huang et al., 2012; Lewin et al., 2021) had a high risk of bias regarding patient selection and others had a risk of bias in index text (He et al., 2020) or reference standards (Huang et al., 2003). Additionally, due to missing information, the risk of bias could not be assessed for another 16 studies (Yeo et al., 2005; Iizuka et al., 2006; Igetei et al., 2008; Chen et al., 2013; Li et al., 2014; Yang et al., 2014; Ramadan et al., 2015; Dong et al., 2017; Mansour et al., 2017; Li et al., 2018; Wei et al., 2018; Linlin et al., 2018; Bai et al., 2019; Marchio et al., 2019; El-Bendary et al., 2020; Xie et al., 2021). Because many case-control studies were enrolled in this meta-analysis, the selection of patients was the main bias risk across the included publications.

Diagnostic performance

Diagnostic accuracy of the quantitative and qualitative analysis of ctDNA for HCC

The qualitative detection of ctDNA discriminated HCC patients from control individuals with a SEN of 0.50 (95% CI [0.43–0.56], I² statistic: 94.01%) and a SPE of 0.90 (95% CI [0.86–0.93], I² statistic: 95.84%) (Fig. 3). In addition, the pooled PLR was 4.89 (95% CI [3.66–6.53], I² statistic:92.21%), NLR was 0.56 (95% CI [0.50–0.63], I² statistic:95.63%), DOR was 8.72 (95%CI [6.18–12.32], I² statistic:100%), and the AUC for the SROC curve was 0.78 (95%CI [0.74–0.81]). The HSROC graph was plotted and the value of beta was 0.23 (95% CI [−0.05–0.56]), z value was 1.61, and p value was 0.106, which indicated that the graph was symmetrical. Lambda was 1.93 (95% CI [1.56–2.33]), which suggested a moderate level of diagnostic value (Fig. 4A). The Spearman correlation coefficient was 0.367 and p-value = 0.003, indicating that heterogeneity among studies was derived from non-threshold effects. In the same way, the pooled SE and SP for the diagnostic performance of the quantitative detection in HCC were 0.69 (95% CI [0.63–0.74], I² statistic:86.32%) and 0.84 (95% CI [0.77–0.89], I² statistic:92.87%), respectively (Fig. 5). The PLR was 4.36 (95% CI [3.02–6.30], I² statistic: 88.24%) and NLR was 0.37 (95% CI [0.31–0.43], I² statistic: 82.48%). The pooled DOR was 11.87 (95%CI [7.78–18.12]; I² statistic:100%) and the AUC for the SROC curve was 0.81 (95% CI [0.77–0.84]). In addition, the value of beta was 0.62 (95% CI [0.17–1.07]), z statistic was 2.68, and p value was 0.007, indicating that the HSROC was asymmetric. Lambda was 2.32 (95%CI [1.96–2.69]), which indicated a moderate level of diagnostic value (Fig. 4B). The Spearman correlation coefficient of 0.494 and the p value of 0.014 indicated that there was no threshold effect.

Diagnostic performance of ctDNA combined with AFP assay for HCC

Using the combination of ctDNA and AFP as detective indicators distinguished HCC patients from non-cancer control participants with a SEN of 0.82 (95% CI [0.77–0.86], I² statistic: 85.81%) and a SPE of 0.84 (95% CI [0.76–0.90], I² statistic:93.32%) (Fig. 6). The combined PLR, NLR, and DOR was 5.13 (95% CI [3.31–7.96], I² statistic:92.31), 0.22 (95% CI [0.17–0.28], I² statistic:85.91), 23.63 (95% CI [12.82–23.56]; I² statistic:100%), respectively. The results obtained by the HSROC model showed that the value of beta was 0.56 (95% CI [−0.03–1.15]), z statistic was 1.87, and p value was 0.062. The estimate for the “Lambda” and its 95% confidence interval was 3.24 (95% CI [2.63–3.85]), suggesting a high level of diagnostic value (Fig. 4C).

Diagnostic value of circulating mSEPT9 and ctDNA concentration for HCC

We also analyzed the diagnostic performance of mSEPT9 and ctDNA concentration. The SEN, SPE, PLR, NLR, DOR, AUC, and lambda values of mSEPT9 were 0.80 (95% CI [0.71–0.87], I² statistic:85.27%), 0.77 (95% CI [0.68–0.85], I² statistic:80.97%), (Fig. 7), 3.57 (95% CI [2.29–5.56], I² statistic:74.47%), 0.25 (95% CI [0.15–0.42], I² statistic:85.67%), 14.06 (5.64–35.05, I² statistic:85.48%), 0.86, and 2.64 (95% CI [1.73–3.55]), (Fig. 4D), respectively. The SEN, SPE, PLR, NLR, DOR, AUC, and lambda values of ctDNA concentration were 0.63 (95% CI [0.56–0.70], I² statistic:1.38%), 0.86 (95% CI [0.74–0.93], I² statistic:83.85%) (Fig. 8), 4.61 (95% CI [2.50–8.48], I² statistic:57.20%), 0.42 (95% CI [0.36–0.50], I² statistic:0.00%), 10.86 (5.60–21.07, I² statistic:47.79%), 0.73, and 2.01 (95% CI [1.48–2.55]) (Fig. 4E), respectively.

Subgroup and meta regression analyses

Subgroup analysis was applied based on different covariates to explore the potential sources of heterogeneity: region (Asia vs. Africa), sample source (plasma vs. serum), control type (benign disease vs. healthy controls), sample size (≥100 vs. <100), publication year (2000–2010 vs. 2011–2021), assay methods (RT-qPCR vs. other methods in the quantitative study; MSP vs. other methods in the qualitative studies)(Table 1), and single gene methylation vs. single gene mutation in qualitative studies. It should be noted that the studies that did not distinguish patients from healthy controls were not enrolled in the control type subgroup analysis, and the number of studies in Europe (n = 2) and America (n = 3) was not sufficient enough to perform subgroup analysis. The qualitative analysis of the ctDNA subgroup based on different regions revealed that the sampling area from Asia showed better diagnostic performance (DOR:8.83, AUC:0.77) than the area in Africa (DOR:3.22, AUC:0.66). Another subgroup analysis associated with years suggested that studies from 2000–2010 (DOR:26.83, AUC:0.75) had SEN and SPE values of 0.39 and 0.98, respectively, and had lower sensitivity but greater specificity than studies from 2011–2021 (DOR:7.40, AUC:0.79), which had SEN and SPE values of 0.54 and 0.86, respectively. As for sample source, when we compared the samples collected from plasma, the sample source from serum did not show a great difference with a DOR of 5.77 vs 10.61 and an AUC of 0.76 vs. 0.77, which were quite different from the results in the quantitative subgroup. Also, studies with a sample size of ≥100 cases (DOR:7.80, AUC:0.78) were not drastically different when compared with studies with a sample size of <100 cases (DOR: 10.99, AUC: 0.78). Additionally, the single-gene mutation subgroup (DOR:5.11, AUC:0.53) had rather low SEN and higher SPE when compared with the single-gene methylation subgroup (DOR:9.05, AUC:0.79). Similarly, the quantitative analysis of the ctDNA subgroup showed a SEN of 0.65 and a SPE of 0.92, and sampling from plasma (DOR:20.54, AUC:0.87) achieved a greater diagnostic value compared to the subgroup that sampled from serum (DOR:8.00, AUC:0.79), which had a SEN of 0.71 and SPE of 0.76. Likewise, subgroup analyses related to control type showed that studies had satisfactory diagnostic value in discriminating HCC patients from healthy volunteers (DOR:59.26, AUC:0.91) compared with those using benign liver disorder patients. In terms of assay methods, research utilizing RT-PCR detective methods (DOR:22.69, AUC:0.88) revealed superior diagnostic accuracy in discriminating HCC patients from the control group with a SEN of 0.75 and SPE of 0.88 compared with research using other detective methods (DOR:10.20, AUC:0.79) with a SEN of 0.68 and SPE of 0.83, respectively.

We also performed a multivariable meta-regression to further explore the source of heterogeneity (Table 2). The results indicated that the study region and control type may be the source of heterogeneity in the qualitative analysis subgroup. Meanwhile, none of the study characteristics shown above generated significant heterogeneity in the quantitative analysis group.

Clinical effect

Based on Fagan’s nomogram, when we set the pretest probability to 25%, ctDNA increased the probability of a positive value to 62%, while there was a 16% probability that it ignored HCC patients with a negative test (Fig. 9A). Similarly, based on a 50% pretest probability, the probability of a correct detection increased to 83% after a 36% probability of a negative test result (Fig. 9B). When setting the pretest probability to 75%, the probability of a positive detection increased to 94%, and the probability of HCC patients being ignored increased to 63% (Fig. 9C). In the quantitative subgroup, the posttest probability increased to 59%, 81%, and 93% when we set the pretest probability to 25%, 50%, and 75% respectively, with lower post negative test results in discriminated HCC patients when compared with the qualitative analysis subgroup. Therefore, ctDNA may help AFP and ultrasounds initially screen for HCC.

Publication bias

We utilized Deek’s funnel plot asymmetry test to research publication bias in the included studies (Fig. S1A & S1B). Our results showed that there was no significant publication bias in the quantitative analysis group with a coefficient of −1.90 (95% CI [−24.12–20.31]; p = 0.86), while a coefficient of 11.80 (95% CI [1.70–21.90], p = 0.023) indicated a study bias among studies that used a qualitative methodological approach. Trim and fill analysis (Chan et al., 2008) was used to correct funnel plot asymmetry from publication bias. The log OR was used as the effect estimate to execute the test. The pooled log OR was 0.685 (95% CI [0.648–0.723], p = 0.00) for the mixed model before we filled in the missing studies and the outcome altered to 0.668 (95% CI [0.632–0.705], p = 0.00) (Fig. S2). Also, we did not find publication bias in the group of ctDNA combined with AFP assay.