The first 2 authors contributed equally to this article.
Article first published online: 10 JAN 2011
Copyright © 2011 American Cancer Society
Volume 117, Issue 13, pages 2951–2960, 1 July 2011
How to Cite
Tsai, H.-W., Lin, Y.-J., Lin, P.-W., Wu, H.-C., Hsu, K.-H., Yen, C.-J., Chan, S.-H., Huang, W. and Su, I.-J. (2011), A clustered ground-glass hepatocyte pattern represents a new prognostic marker for the recurrence of hepatocellular carcinoma after surgery. Cancer, 117: 2951–2960. doi: 10.1002/cncr.25837
We thank the technical assistance of Microbial Genomics Core of National Taiwan University Research Center for Medical Excellence Division of Genomic Medicine for the virus quantification (and genotyping).
- Issue published online: 17 JUN 2011
- Article first published online: 10 JAN 2011
- Manuscript Accepted: 10 NOV 2010
- Manuscript Revised: 9 NOV 2010
- Manuscript Received: 14 OCT 2010
- pre-S mutant;
- ground-glass hepatocyte;
- hepatitis B virus;
- hepatocellular carcinoma;
The recurrence of hepatocellular carcinoma (HCC) after hepatectomy is a serious event. It has been demonstrated that different ground-glass hepatocyte (GGH) patterns harbor specific hepatitis B virus (HBV) pre-S deletion mutants and represent preneoplastic lesions in chronic HBV infection. In the current study, the authors investigated whether a specific GGH pattern in nontumorous liver tissues was associated with the recurrence of HBV-related HCC after surgery.
Clinicopathologic data from 82 patients with HBV-related HCC were reviewed. GGH patterns were assessed on hematoxylin and eosin-stained sections. Tissue hepatitis B surface antigen (HBsAg) expression was evaluated by immunohistochemical staining. Serum profiles of pre-S status, viral load, and HBV genotype were determined and correlated with clinical recurrence and survival after surgery.
The results indicated that the clustered pattern of GGHs or HBsAg expression was associated significantly with decreased local recurrence-free survival (LRFS) during a mean follow-up of 46.4 months (P<.001). This biomarker was comparable to or better than the prognostic value of other parameters, such as multifocal tumors (P = .022), satellite nodules (P = .005), small cell dysplasia (P = .045), or elevated viral load (P = .027), to predict recurrent HCC. Multivariate analysis also revealed that type II GGHs, which expressed marginal HBsAg and consistently clustered in nodules, were independent variables associated with LRFS (P<.001) and overall survival (P = .003).
The current results indicated that the assessment of GGH patterns or HBsAg expression in nontumorous liver tissues provides an easily recognized, new risk marker for the recurrence of HBV-related HCC after hepatic resection. Cancer 2011. © 2011 American Cancer Society.
Hepatocellular carcinoma (HCC) is 1 of the most common cancers in the world. The majority of HCCs are attributable to either hepatitis B virus (HBV) or hepatitis C virus infection.1-3 The prognosis for patients with HCC remains grave. Although several nonsurgical modalities have been developed for HCC therapy, surgical resection remains the most important approach for curative treatment.4 However, the recurrence rate after surgery is high, ranging from 38% to 61.5% in 5 years.5-7 Therefore, to identify the risk factors associated with HCC recurrence, proper patient management is important. In past decades, tumor factors, such as multifocal tumors, satellite spreading, vascular invasion, and clinical staging, were among the potential risk factors associated with HCC recurrence after surgery.8 Along with these tumor factors, several viral factors, such as high viral load, genotype C, and pre-S deletion mutants in serum, also have been identified as predictors of HCC development.9-12 Recently, it was demonstrated that gene expression profiles in nontumorous liver tissues play a more significant role in predicting late recurrences and survival than tumor factors.13 Therefore, the recognition of pathology in nontumorous liver tissues from patients with HCC may provide an informative parameter with which to identify a group of high-risk patients who may develop recurrent HCC after undergoing surgery.
Previously, we reported that “ground-glass” hepatocytes (GGHs)14, 15 harbor pre-S mutant, large-surface proteins (LHBs)16 and represent precursor lesions of HBV-related HCC.11 There are 2 major types of GGHs: Type I GGHs harbor pre-S1 deletion mutants and usually appear in a globular or inclusion-like pattern in hepatitis B surface antigen (HBsAg) immunostaining; whereas type II GGHs harbor pre-S2 mutants, exhibit HBsAg immunostaining at the cell periphery, and consistently cluster in groups.17 The different patterns of GGHs and HBsAg immunostaining may represent different replication stages of chronic HBV infection.16 Previously, we also demonstrated that the pre-S mutant proteins in GGH are accumulated in the endoplasmic reticulum and induce endoplasmic reticulum stress signals, which further induce genomic instability through oxidative stress.11 Pre-S mutants also may up-regulate cyclooxygenase-2 through neuronal factor-κB and p38 microtubule-associated protein kinase, and they may up-regulate cyclin A to induce cell cycle progression.18, 19 It also was noted that GGHs overexpressed vascular endothelial growth factor-A (VEGF-A) the with activation of protein kinase B/mammalian target of rapamycin (Akt/mTOR) signals, through which GGHs may exhibit a growth advantage.20 A study in transgenic mice revealed that pre-S mutant LHBs can induce dysplastic changes.11 Therefore, GGHs may represent preneoplastic lesions, and the pre-S deletion mutant proteins are potential viral oncoproteins.11
In the current study, we hypothesized that the presence of GGHs or the patterns of HBsAg immunostaining in nontumorous liver tissues may represent a new prognostic marker for disease recurrence or survival in patients with HCC after surgery in view of the preneoplastic nature of GGHs, which may accelerate to regenerate and grow into full-blown HCC after surgery. The prognostic value of GGH patterns was evaluated and compared with other clinicopathologic parameters, serum profiles of pre-S status, viral load, and genotype in 82 patients with HBV-related HCC who underwent surgical resection.
MATERIALS AND METHODS
Patients, Samples, and Clinicopathologic Data
From November 2001 to September 2007, 82 consecutive patients who had HBV-related HCC underwent hepatectomy at National Cheng Kung University Hospital (NCKUH). Tissue and serum samples from all 82 patients were collected by the tissue bank at NCKUH. The clinicopathologic records were analyzed. Serum samples that were obtained before hepatectomy were analyzed for pre-S mutant status, HBV viral load, and genotype. Archival tissue blocks from 82 HCC samples were retrieved, and at least 3 tissue sections that contained both tumor (1 section) and nontumorous liver tissues (2 sections) from each patient were analyzed for histopathology, the presence or pattern of GGHs, and HBsAg immunohistochemical expression. The histology of liver parenchyma was analyzed using the system described by Knodell et al.21 The differentiation of HCC was analyzed using the Edmondson and Steiner grading system.22 This study was approved by the Human Experiment and Ethics Committee of NCKUH (no. HR-95-103), and informed consent was obtained from all patients.
Immunohistochemical Analysis and Scoring System
The immunohistochemical staining of HBsAg was performed as described previously.17 The optimal dilution of anti-HBsAg was determined by using liver tissue from an HBV carrier as a positive control. The StrAviGen Super Sensitive MultiLink kit (BioGenex Laboratories, San Ramon, Calif) was used to detect the resulting immune complex. Peroxidase activity was observed by using an aminoethyl carbazole substrate kit (Zymed Laboratories, San Francisco, Calif). Finally, the sections were counterstained with hematoxylin. For negative controls, nonimmune mouse immunoglobulin was substituted for primary antibody in the incubation. The HBsAg staining patterns were classified as follows: Pattern A, weak HBsAg immunostaining in hepatocytes with homogeneous, faint, cytoplasmic staining; Pattern B, type I GGH pattern with dense, globular, or “inclusion-like” staining; and Pattern C, type II GGH pattern identified as hepatocytes with surface antigen staining at the cell margin or periphery.23, 24 The presence of immunostaining or the GGH pattern was assessed in the 2 sections of nontumorous liver parenchyma by 2 pathologists (H.-W.T. and I.-J.S.).23, 24 A semiquantitative expression scoring system was used with scores from 0 to 4 corresponding to the percentage of positive immunostaining in 0%, <5%, 5% to 9%, 10% to 29%, and ≥30% of hepatocytes, respectively.25
Pre-S Mutation Detection in Serum Samples
To correlate the status of pre-S mutants in serum with HCC recurrence, serum samples (200 μL) from all patients were used for DNA extraction with the QIAamp DNA Mini Kit (Qiagen, Hilden, Germany). The following primers were used for polymerase chain reaction (PCR) analysis and were conserved among all reported HBV sequences: The sense primer was 5′-GCGGGT CACCATATTCTTGG-3′ (nucleotides 2812-2837), and the antisense primer was 5′-GAGTCTAGACTCT GCGGTAT-3′ (nucleotides 255-236), in which the defined region covered the entire pre-S1 and pre-S2 genes. The PCR program was set for 40 cycles. The antisense primer was replaced by 5′-TAACACGAGCAGGGG TCCTA-3′ (nucleotides 199-180) in the nested PCR mixture. The amplified DNA was purified with the MinElute PCR Purification Kit (Qiagen). For samples in which multiple pre-S PCR products were revealed, Escherichia coli thymine-adenine (TA) cloning was used. Three microliters of PCR product were ligated with 50 ng of TA cloning vector (pCR 2.1; Invitrogen, Carlsbad, Calif) at 14°C overnight. The next day, the ligation product was transformed into an E. coli DH5α strain. The pre-S insert DNA was examined using colony PCR. Then, the deletion region of the pre-S DNA was detected using a pre-S gene chip, as described in our previous study.26
HBV Viral Load in Sera and Genotyping
The associations between GGH patterns and HBV serum profiles were calculated by using the chi-square test. The expression scores of HBsAg between patients who had HCC with and without a recurrence were compared using the Fisher exact test. The correlations between viral load, pre-S mutation status, and HBV genotype were calculated by using a 1-way analysis of variance, an independent 2-sample t test, and the Fisher exact test, as appropriate. The local recurrence-free survival (LRFS) and overall survival (OS) rates were calculated by using the Kaplan-Meier method, and the log-rank test was used to assess the significance of differences between groups. Cox proportional hazards regression models with a forward stepwise selection method were used to assess the independence of different factors, and only those prognostic variables that had P values <.15 in univariate analysis were included in the model. P values <.05 were considered statistically significant.
Patient Profiles and Clinical Outcomes
The clinicopathologic data and serum HBV viral profiles of the 82 patients are summarized in Table 1. There were 62 men and 20 women, and the mean age of all patients was 54.4 years. The mean duration of follow-up was 46.4 months (range, 1.8-92.9 months). Forty-one patients (50%) developed a local recurrence of HCC in the liver at a mean follow-up of 16.7 months after surgery (range, 2-74.2 months), and 28 patients (34.1%) died of disease and had a mean survival of 25.1 months after surgery (range, 1.8-73.9 months). After surgery, 17 patients received anti-HBV drugs, which included lamivudine in 2 patients, adefavir in 1 patient, entecavir in 11 patients, and lamivudine switched to entecavir in 3 patients. The mean duration of anti-HBV drug treatment was 24 months (range, 1-60.8 months). None of the patients received anti-HBV drugs before surgery.
|Variable||No. of Patients|
|Age: Mean (range), y||54.4, 21∼78|
|HCC differentiation: Well/moderate/poor||5/58/19|
|Multifocal tumor: No/yes||79/3|
|Satellite nodule: No/yes||67/15|
|Size: Mean (range), cm||4.8 (0.7-16)|
|Vascular invasion: No/microscopic/major branches||41/36/5|
|Child-Pugh class: A/B||74/8|
|Okuda stage: I/II||57/25|
|CLIP score: 0/1/2/3/4||33/30/11/7/1|
|BCLC stage: A1-3/A4/B||48/4/30|
|Knodell inflammation score: ≤5/≥6||54/28|
|Knodell fibrosis score: 0-1/3/4||7/40/35|
|Large cell changes: No/yes||5/77|
|Small cell changes: No/yes||40/42|
|Weak staining score, median||3|
|Sporadic type I GGH score, median||1|
|Clustered type I GGH score, median||1.5|
|Type II GGH score, median||2|
|Serum viral load: Median (range), copies/mL||34,550 (0-4.59×108)|
|Serum pre-S mutation: Wild type/pre-S1/ pre-S1 and S2||28/19/31|
|Anti-HBV drug after surgery: No/yes||65/17|
Immunohistochemical Evaluation of HBsAg Expression
HBsAg expression was detected in the nontumorous liver tissues from 79 patients (96.3%). Weak-staining hepatocytes (Fig. 1Aa), which presumably harbored wild-type HBsAg,17 were identified in 76 patients (92.7%). Type I GGHs, which harbored a pre-S1 deletion mutant,17 were identified in 68 patients (82.9%). A sporadic distribution pattern (Fig. 1Ab) was observed in 56 patients (68.3%), and a clustered GGH pattern (Fig. 1Ac) was observed in 56 patients (68.3%). A marginal HBsAg immunostaining pattern or a type II GGH pattern, in which GGHs were distributed either consistently in clusters or extensively in entire cirrhotic nodules (Fig. 1Ad),17 was observed in 50 patients (61%). Most patients (n = 67; 81.7%) had a combination of multiple patterns of HBsAg immunostaining: A combination of weak HBsAg staining in hepatocytes and a type I GGH pattern was observed in 18 patients (26.9%); a combination of weak-staining hepatocytes and the type II GGH pattern was observed in 1 patient (1.5%); and a combination of weak-staining hepatocytes, type I GGH patterns, and the type II GGH pattern was observed in 48 patients (71.6%). The median HBsAg expression scores were 3, 1, 1.5, and 2 for weak-staining hepatocytes, a sporadic type I GGH pattern, a clustered type I GGH pattern, and the type II GGH pattern, respectively. In HCC samples, immunohistochemical staining for HBsAg revealed focal, weak, cytoplasmic staining in 7 tumors (8.5%); moderately intense staining in <10% of the tumor area in 10 tumors (12.2%); and moderately intense staining in >10% of the tumor area in 13 tumors (15.9%). The moderately intense staining patterns were either cytoplasmic or membranous. Most tumors had cytoplasmic staining, including homogeneous cytoplasmic staining in 10 tumors, globule staining that resembled type I GGHs in 10 tumors, marginal staining that represented type II GGHs in 4 tumors, and dot-like staining in the perinuclear area in 3 tumors. A focal membranous staining pattern was observed in only 2 tumors.
Serum Profile of Pre-S Status, Viral Load, and HBV Genotype
Pre-S mutants were detectable in serum samples from 50 patients (61%). The location of the deletion site was identified in the pre-S1 region in 19 patients (23.2%) and in both the pre-S1 and pre-S2 regions in 31 patients (37.8%). The median of viral load was 3.46 × 104 copies/mL (range, from undetectable to approximately 4.59 × 108 copies/mL). When we used 104 copies/mL as the cutoff point,29 50 patients (61%) had an elevated viral load. The genotype could be determined in 59 patients. Twenty-six patients (44.1%) had genotype B virus, and 33 patients (55.9%) had genotype C virus. The presence of pre-S mutants did not correlate with viral load (P = .623). There was no significant difference in viral load between genotype B and genotype C (P = .374). The correlation between pre-S mutation status and genotype was not statistically significant (P = .798).
Correlation Between HBsAg Expression Patterns and HBV Serum Profiles
Higher expression scores for weak-staining hepatocytes (P = .033; results not shown), the sporadic type I GGH pattern (P < .001), the clustered type I GGH pattern (P = .002), and the type II GGH pattern (P < .001) were associated significantly with an elevated viral load. The sporadic type I GGH pattern and the type II GGH pattern were associated significantly with genotype C (P = .007 and P = .009, respectively). There was no significant correlation between GGH patterns and serum pre-S mutation status. Weak-staining hepatocytes were not associated with serum genotype or pre-S mutation status.
Prognostic Significance of GGH Patterns or HBsAg Immunostaining, HBV Serum Profiles, and Clinicopathologic Indicators
Patients who had locally recurrent HCC had significantly higher expression scores for the type II GGH pattern in their surgical specimens (P = .008) (Fig. 1Bd). The pattern or score of weak-staining hepatocytes (P = .976) (Fig. 1Ba), the sporadic type I GGH pattern (P = .341) (Fig. 1Bb), and the clustered type I GGH pattern (P = .151) (Fig. 1Bc) revealed no significant correlation with local recurrence.
According to a Kaplan-Meier analysis, the sporadic type I GGH pattern (P = .017), the clustered type I GGH pattern (P = .004), and the type II GGH pattern (P < .001) were associated significantly with decreased LRFS (Fig. 2A-C). The type II GGH pattern also was correlated significantly with decreased OS (P = .005) (Fig. 2D). Weak-staining hepatocytes, HBV genotype, and pre-S mutation in serum did not correlate with LRFS (P = .676, P = .449, and P = .884, respectively) or OS (P = .981, P = .784, and P = .522, respectively) (Table 2). The HBsAg staining patterns in HCC tissues, however, did not correlate with LRFS or OS (P = .296 and P = .666, respectively; data not shown). The use of anti-HBV drugs did not correlate with LRFS (P = .991) or OS (P = .625) (Table 2). To further exclude the possible influence of antiviral treatment on disease recurrence or survival, we further analyzed the 65 patients who received no anti-HBV drug treatment. We observed that the sporadic type I GGH pattern (P = .028, data not shown) and the type II GGH pattern (P = .005) (Fig. 2E) were correlated significantly with decreased LRFS. The type II GGH pattern was correlated significantly with decreased OS (P = .025) (Fig. 2F). In the 17 patients who received anti-HBV drug treatment, the type II GGH pattern also was correlated significantly with decreased LRFS (P = .045; data not shown).
|Nontumor liver pathology|
|Knodell inflammation score||≤5/≥6||.907||≤5/≥6||.471|
|Large cell changes||−/+||.218||−/+||.564|
|Small cell changes||−/+||.045a||NS||−/+||.145||.016a|
|Weak staining score||0-1/2-4||.676||0/1/2/3/4||.981|
|Sporadic type I GGH score||0/1-4||.017a||NS||0/1/2/3/4||.137||NS|
|Clustered type I GGH score||0-1/2-3/4||.004a||NS||0-1/2-3/4||.538|
|Type II GGH score||0-1/2-4||<.001a||<.001a||0-1/2-4||.005a||.003a|
|Viral load, copies/mL||<104/≥104||.027a||NS||<104/≥104||.627|
|Pre-S1 and S2 mutation||−/+||.475||−/+||.638|
|Anti-HBV drug treatment||−/+||.991||−/+||.625|
To compare the prognostic value of the GGH pattern with other recognized parameters, univariate analysis using the Cox regression method was adopted. The data revealed that multifocal tumors (P = .022), satellite nodules (P = .005), small cell changes (P = .045), the sporadic type I GGH pattern (P = .017), the clustered type I GGH pattern (P = .004), the type II GGH pattern (P < .001), and viral load (P = .027) were associated significantly with decreased LRFS (Table 2). Multivariate analysis also revealed that satellite nodules (P<.001; odds ratio [OR], 4.110; 95% confidence interval [CI], 1.895-8.911) and the type II GGH pattern (P<.001; OR, 4.048; 95% CI, 1.985-8.253) were independent variables associated with decreased LRFS. In univariate analysis, significant risk factors that were associated with decreased OS were satellite nodules (P < .001), vascular invasion (P = .025), Cancer of the Liver Italian Program (CLIP) score (P = .003), Barcelona Clinic liver cancer stage (P = .006), and the type II GGH pattern (P = .005) (Table 2). Multivariate analysis revealed that satellite nodules (P < .001; OR, 8.395; 95% CI, 3.363-20.959), CLIP score (for a CLIP score of 4: P = .004; OR, 79.036; 95% CI, 6.579-949.507), small cell changes (P = .016; OR, 3.072; 95% CI, 1.230-7.671), and the type II GGH pattern (P = .003; OR, 3.886; 95% CI, 1.605-9.412) were independent variables associated with decreased OS (Table 2).
We also analyzed the prognostic factors associated with early or late local tumor recurrence (Table 3). In univariate analysis, significant risk factors that were associated with early recurrence within 1 year were satellite nodules (P = .001), the clustered type I GGH pattern (P = .010), and the type II GGH pattern (P = .032) (Table 3). Multivariate analysis revealed that the presence of satellite nodules (P = .002; OR, 5.538; 95% CI, 1.805-15.126) was an independent variable associated with early recurrence. For recurrences after 1 year, significant risk factors in univariate analysis were multifocal tumors (P = .009), small cell changes (P = .011), and the type II GGH pattern (P = .006) (Table 3). Multivariate analysis revealed that multifocal tumors (P = .004; OR, 10.053; 95% CI, 2.125-47.552) and the type II GGH pattern (P = .004; OR, 3.407; 95% CI, 1.490-7.789) were independent variables associated with late recurrence (Table 3).
|LRFS <1 Year||LRFS >1 Year|
|Factor||Categories||Univariate P||Multivariate P||Univariate P||Multivariate P|
|Nontumor liver pathology|
|Knodell inflammation score||≤5/≥6||.083||NS||.148||NS|
|Large cell changes||−/+||.314||.419|
|Small cell changes||−/+||.963||.011a||NS|
|Weak staining score||0-1/2-4||.774||.758|
|Sporadic type I GGH score||0/1-4||.122||NS||.068||NS|
|Clustered type I GGH score||0-1/2-3/4||.010a||.056||.262|
|Type II GGH score||0-1/2-4||.032a||NS||.006a||.004a|
|Viral load, copies/mL||<104/≥104||.117||NS||.111||NS|
|Pre-S1 and S2 mutation||−/+||.142||NS||.831|
|Anti-HBV drug treatment||−/+||.190||.341|
In the current study, we have demonstrated that the clustered patterns of GGHs or HBsAg immunostaining in nontumorous liver tissues were correlated significantly with late HCC recurrence after surgical resection, thereby providing an easily recognized, new marker to predict the recurrence of HBV-related HCC. These results also support our previous proposal that GGHs may represent preneoplastic lesions that may accelerate, grow, and progress to HCC after surgical resection.11 In addition to the GGH pattern, multifocal tumors (indicating a field effect) and small cell changes in hepatocytes also were associated significantly with decreased LRFS, suggesting that the underlying nontumorous liver condition is as important as the primary tumor factors in terms of HCC recurrence.
Consistent with previous reports,8, 13, 30 we also observed that satellite nodules and vascular invasion were associated with HCC recurrence and patient survival. However, distinct from the early recurrence (usually within 1 year) associated with primary tumor factors (such as satellite spreading, vascular invasion, or inadequate resection margin), the nontumorous liver condition (such as GGHs or small cell dysplastic changes) generally tend to be associated with late recurrences, usually more than 1 or 2 years after surgical resection (Table 3). After patients undergo partial hepatectomy, a series of cytokines, such as tumor necrosis factor and interleukin-6, will be up-regulated and, the hepatocytes that are primed by these agents will respond to growth factors (eg, hepatocyte growth factor and epidermal growth factor), accelerating regeneration of the remaining liver.31 This assumption is supported by the results reported from 1 study that the gene profiles associated with inflammation (but not with the inflammation associated with tumor signaling) were highly correlated with late HCC recurrence or survival.13 It has been observed that GGHs exhibit enhanced expression of VEGF-A and growth factors with activation of Akt/mTOR signals, primarily induced by the pre-S mutant proteins through the endoplasmic reticulum stress response.20 In addition to the endoplasmic reticulum-dependent stress response, type II GGHs that harbor pre-S2 deletion mutants also can induce endoplasmic reticulum stress-independent, Jun-activating binding protein/p27/retinoblastoma cyclin signals to accelerate cell cycle progression.32 Therefore, virus-associated changes in nontumorous liver tissues also are important for predicting the recurrence of HBV-related HCC.
Unexpectedly, HBV genotype and antiviral therapy revealed no prognostic significance in terms of HCC recurrence in the current study. A larger sample size and longer follow-up may be needed for the association to reach significance, because only 17 patients received the treatment in the current study. Although it has been demonstrated that the presence of serum pre-S deletion mutants indicates a high risk for developing HCC,9 we did not demonstrate that serum pre-S mutants were associated significantly with HCC recurrence or survival. Thus, the direct study of GGH patterns or HBsAg immunostaining in the nontumorous liver may provide more reliable information. Our data also demonstrated that the status of serum pre-S mutants did not correlate perfectly with the GGH pattern. The large number of tumors that had mixed patterns of expression may be a potential confounding factor to explain this discrepancy. Further studies may be needed to clarify the discrepancy and significance of pre-S mutants in serum and in hepatocytes.
CONFLICT OF INTEREST DISCLOSURES
This study was supported by grants from the National Health Research Institutes and the National Science Council (to I.-J.S. and W.H.) and from National Cheng Kung University Hospital, Taiwan (NCKUH Research Grant-9803005 to H.-W.T.).
- 1Department of Health, Taiwan. Health and National Health Insurance Annual Statistics Information Service. http://www.doh.gov.tw/statistic/. Accessed August 18, 2006.