Construction of the Chinese University Prognostic Index for hepatocellular carcinoma and comparison with the TNM staging system, the Okuda staging system, and the Cancer of the Liver Italian Program staging system
The current TNM staging system for patients with hepatocellular carcinoma (HCC) does not include liver function parameters and does not provide a precise prognosis for patients in different risk groups. The objectives of this study were to construct a new prognostic index for patients with hepatocellular carcinoma, the Chinese University Prognostic Index (CUPI), and to compare it with existing staging systems in terms of their ability to classify patients into different risk group.
From 1996 to 1998, 926 ethnic Chinese patients who were diagnosed with HCC (mainly hepatitis B-associated) at a single institution were recruited prospectively into this study. A multivariate analysis on 19 patient characteristics was performed using a Cox regression model to identify independent prognostic factors. Weights were derived from the regression coefficients of various factors to construct the CUPI. Patients were classified according to different staging systems. Survival curves were plotted with the Kaplan–Meier method and were compared by using a log-rank test.
Both the TNM staging system and the Okuda staging system had prognostic significance, but the significance was lower for the Cancer of the Liver Italian Program (CLIP) prognostic score among the patients in the study population. The CUPI was constructed by adding the following factors into the TNM staging system: total bilirubin, ascites, alkaline phosphatase, α fetoprotein, and asymptomatic disease on presentation. The new CUPI characterized three risk groups with highly significant differences in survival during the whole period of follow-up (P < 0.00001) and was more discriminant than the other systems.
Hepatocellular carcinoma (HCC) was the fourth leading cause of global cancer deaths in 1998.1 More than half of the worldwide cases occur in China, and there is also a rising incidence both in Europe and in the United States.2, 3 HCC, in general, is an aggressive and rapidly fatal malignancy, but the course of the disease can still vary depending on many disease characteristics. The tumor-lymph node-metastasis (TNM)4 staging system is widely recognized as the standard approach for prognostication in most solid tumor systems. The current TNM staging system for HCC uses size of tumor, presence of vascular invasion, and distribution of tumor(s) as criteria for T staging. The criteria for N and M staging are similar to other tumor systems. There are no other nontumor-related factors included in the present TNM staging system. It is generally accepted that the prognosis for patients with HCC, unlike other tumors, does not depend only on the size and extent of tumor alone but also on the functional state of the nontumorous liver.5, 6 This is due to concomitant, chronic liver disease that is present in most patients. Thus, patients with HCC, in effect, have two diseases with independent natural histories. The inclusion of nontumor-related factors into a staging system for HCC, mainly liver function status, appears logical and may be expected to permit more accurate prognostication. The International Union Against Cancer (UICC) also recommends the inclusion of patient-related factors, so that the prediction of prognoses at different stages can be improved.5
One of the early attempts to incorporate liver function status into the staging of patients with HCC was proposed by Okuda et al. in 1985 and was based on a study of the natural history of 850 patients with HCC.6 This later was called the Okuda staging system and has found wide acceptance as an improved classification system for HCC. The Okuda staging system was based on three measures of liver function (ascites, albumin, and bilirubin levels) and an estimate of the percentage of primary tumor involvement in the liver. Two more Japanese studies7, 8 and one Spanish study9 that used similar approaches by incorporating nontumor-related factors, but with fewer patients, also reported different formulations of a prognostic index for HCC. However, clinical use of these prognostic indexes is limited due to the complexity of the mathematical models and inclusion of nondiscrete and subjective variables. More recently, another prognostic index was derived by the Cancer of the Liver Italian Program (CLIP) investigators (the CLIP score).10 This was based on 435 patients from 16 centers, and those investigators found that Child–Pugh grading, distribution of tumor(s), α-fetoprotein (AFP) level, and portal vein thrombosis were independent predictive factors for survival. The CLIP score also was validated prospectively in 196 patients, and it had greater predictive power for survival compared with the Okuda staging system.11 These studies all were done in countries outside Southeast Asia, and, to date, there has been no similar study in our area of the world, where there is a high incidence of hepatitis B virus (HBV)-related HCC.
Therefore, the objective of this article was to report on large numbers of patients with HCC from a single center in Hong Kong and to derive a new prognostic index called the Chinese University Prognostic Index (CUPI). Another objective was to compare the CUPI with other staging or prognostic scoring systems (TNM, Okuda, and CLIP score) and to find the most suitable system for general use in prognostication.
MATERIALS AND METHODS
All consecutive adult patients (age ≥ 18 years) who were diagnosed with HCC and registered in the Joint Hepatoma Clinic at the Prince of Wales Hospital from 1996 to 1998 were included in this study. There were 1035 patients identified, including 926 patients (89%) with complete survival data and a confirmed diagnosis. These 926 patients became the study population. Eleven percent of these patients (109 of 1035 patients) were excluded due to incomplete survival data, but they had no systemic differences compared with the study cohort. All patients in the sample population were ethnic Chinese. Patients were diagnosed with HCC on the basis of either histologic examination of tumor tissue or serum AFP ≥ 500 ng/mL with radiologic evidence of space-occupying lesion(s) in the liver. Survival was measured from the date of diagnosis to the date of death or last contact for surviving patients. The study population was censored on September 30, 1999. The study sample was split randomly once into two sets: a training set (75% of the population) and a confirmatory set (25% of the population). The training set was used to construct the new CUPI, and the confirmatory set was used to validate the index.
All patients were assessed by a group of experienced surgeons, oncologists, and physicians in the Joint Hepatoma Clinic. Surgical resection was the primary treatment if the disease was considered resectable. When this was not considered possible, patients underwent some form of locoregional therapy (intra-arterial chemotherapy, embolization or internal radiation, and percutaneous ethanol injection) if the disease was confined to the liver or entered into a Phase II study of investigational systemic chemotherapy. Patients in overt liver failure at the time of presentation were not given any anticancer therapy and received best supportive care only.
The accuracy of data entry was evaluated by double entry of a random sample of 100 records, which resulted in an item-to-item agreement of 98.9%. Digital and logical checks further guaranteed the accuracy of the data. Nineteen clinical characteristics that were reported previously as significant prognostic variables6–11 and that were available at the time of diagnosis were chosen for the study. The definitions and cut-off values of respective continuous variables are shown in Table 1. Disease was classified as asymptomatic when patients had HCC that was discovered incidentally and had no symptoms attributable to HCC. Alcohol intake was classified as heavy when the daily ethanol intake was > 80 g for 10 years or more. The diagnosis of cirrhosis was based either on radiologic findings or histologic findings. These study variables were evaluated first by univariate analysis using the log-rank test on survival outcome. The prognostic value of the study factors was assessed further by multivariate analysis. TNM staging was entered first into the Cox regression model.12 The remaining 18 variables were then added to the model containing TNM staging, which was not allowed to be removed, by a forward, stepwise approach. Nested models were tested by the likelihood-ratio method using SPSS software (version 8.0; (SPSS Inc., Chicago, IL).
Table 1. Nineteen Study Variables in 926 Patients with Hepatocellular Carcinoma
No. of patients (%)
HBsAg: hepatitis B surface antigen; AFP: α-fetoprotein.
Age group (yrs)
Asymptomatic disease on presentation
History of heavy alcohol intake
Stage I and II
Stage IIIa and IIIb
Stage IVa and IVb
Total bilirubin (μmol/L)
Alkaline phosphatase (IU/L)
Prothrombin time (seconds)
Construction of the CUPI
The ratio of regression coefficients of the final model was determined and was rounded to whole digits for convenience of calculation. This gave the weights of each prognostic factor. The CUPI score for an individual patient was the sum of the weights of the relevant prognostic factors. The probability of a patient surviving for 3 months was estimated by using the CUPI in a logistic regression model. The high-risk group was defined as patients with a probability > 70% of dying within 3 months. Patients with a probability < 30% of dying within 3 months were classified as the low-risk group, and the intermediate-risk group had a probability between 30% and 70% of dying within 3 months.
Validation of the CUPI
The final model identified in the training set was fitted into the independent confirmatory data set to evaluate the predictive power of the CUPI. The directions and significance of each prognostic factor were examined. Similarly, patients were classified into three risk groups on the basis of their CUPI scores, as determined previously in the training set. Survival functions were then compared with the training set by log-rank test with respect to the three risk groups.
Comparison with the Okuda Stating System, the TNM Staging System, and the CLIP Score
All 926 patients were classified into different prognostic groups according to both the Okuda staging system (Stages I–III) and the TNM classification system (Stages I–II, Stages IIIa–IIIb, and Stages IVa–IVb). Survival curves were plotted by the Kaplan–Meier method and were compared by log-rank test at the 5% significance level with respect to different groups of each staging system. The change of likelihood ratio in the Cox regression model was evaluated by the goodness-of-fit test to compare predictive power of survival by the CUPI and the CLIP prognostic score.
There were 769 male patients and 157 female patients in the study cohort with a mean age of 58.5 years (range, 22–88 years). Positive hepatitis B and C serology was found in 79% and 3.3% of male and female patients, respectively. The average proportion of missing data was 1.9% for the 19 study variables, and the missing data were random. Surgical resection for the primary tumor was performed in 96 patients (10.4%), and nonsurgical interventions were administered to 289 patients (31.2%). The remaining patients received only supportive care (58.4%). The mean follow-up was 33 weeks. Two hundred fourteen patients were censored, and 712 patients had died as of September 30, 1999. The overall median survival was 19 weeks (95% confidence interval [95% CI], 17–21). Survival rates for 3 months, 6 months, 1 year, 2 years, and 3 years were 58.9%, 40%, 23.4%, 6.2%, and 2.2%, respectively.
The assumptions of the Cox proportional hazards model were met for all 19 study variables. It was found that TNM staging was a highly significant predictor (P = 0.0001) for survival, as shown in Model 1 in Table 2. According to the likelihood test on the two nested models (Model 1 and Model 2), the addition of asymptomatic disease on presentation, AFP, total bilirubin (TB), alkaline phosphatase (ALP), and ascites to TNM staging significantly improved the estimation (P < 0.00001) (Table 2).
Table 2. Prognostic Factors in Model 1 and Model 2 Derived from the Training Set (713 patients) and Prognostic Factors for the Confirmatory Set (213 patients) in Model 3a
Likelihood ratio test: Model 1, −2LL = 5376.97; Model 2, −2LL = 5091.06, where −2LL = (Model 1 − Model 2) (chi-square test with 6 degrees of freedom, 285.911; P < 0.00001.
IVa and IVb (reference)
I and II
IIIa and IIIb
AFP ≥ 500 ng/mL
< 34 (reference)
ALP ≥ 200 IU/L
Model 2 in Table 2 shows the regression coefficients (β) and hazard ratios (HR) for each prognostic factor in the training data set. The weights of each prognostic factor are summarized in Table 3. The 6 prognostic factors were TB 34–51 μmol/L (weight, 3) or TB ≥ 52 μmol/L (weight, 4); presence of ascites (weight, 3); ALP ≥ 200 IU/L (weight, 3); AFP ≥ 500 ng/mL (weight, 2); TNM Stage I–II (weight, −3), Stage IIIa–IIIb (weight, −1), or Stage IVa–IVb (weight, 0); and asymptomatic disease on presentation (weight, −4). The sum of the weights of the six prognostic factors gave the prognostic index scores, ranging from −7 to 12. The cut-off scores associated with probabilities of survival at 3 months were from −7 to 1 for the low-risk group, from 2 to 7 for the intermediate-risk group, and from 8 to 12 for the high-risk group. The 3-month survival rates were 85.7% for the low-risk group, 56.4% for the intermediate-risk group, and 20.2% for the high-risk group. The median survival for the low-risk, intermediate-risk, and high-risk groups were 10.1 months (95% CI, 7.7–12.5), 3.7 months (95% CI, 3.1–4.3), and 1.4 months (95% CI, 1.1–1.7), respectively.
Table 3. Weight of the Six Prognostic Factors in the Chinese University Prognostic Index
CUPI: Chinese University Prognostic Index; AFP: α-fetoprotein; TB. total bilirubin; ALP: alkaline phosphatase.
CUPI scores: Summation of the weights of TNM staging + asymptomatic disease on presentation + ascites + AFP + TB + ALp (low-risk group, CUPI score ≤ 1; intermediate risk group, CUPI score = 2–7; high-risk group, CUPI score ≥ 8). For instance, the estimated survival for a patient diagnosed with TNM stage IIIb hepatocellular carcinoma with ascites, AFP 10,000 ng/mL, TB 20 μmol/L and ALP 100 IU/L is calculated as follows. (−1) + (0) + (3) + (2) + (0) + (0) = 4. The patient belongs to the intermediate risk group, with a median survival of 3.7 months (95% confidence interval, 3.1–4.3 months).
I and II
IIIa and IIIb
IVa and IVb (reference)
Asymptomatic disease on presentation
AFP ≥ 500 ng/mL
< 34 (reference)
ALP ≥ 200 IU/L
The final model was then fitted to the confirmatory set, as shown with Model 3 in Table 2, to validate the new CUPI, and all six prognostic factors were significant (P < 0.05). Furthermore, the directions of the coefficients in Model 3 were the same as in the training set, with close magnitudes of HR. Figure 1 shows the corresponding survival curves of differing risk groups for both the training set and the confirmatory set. There was no significant difference in survival for each pair of risk groups by log-rank tests (low-risk group, P = 0.35; intermediate-risk group, P = 0.87; high-risk group, P = 0.63)
The difference in survival among different risk groups classified by the CUPI (P < 0.00001), the TNM staging system (P < 0.00001), the Okuda staging system (P < 0.00001), and the CLIP score (P < 0.00001) were highly significant, as shown in Figure 2a–d, respectively. However, patients identified with Okuda Stage I and II disease were not significantly different with regard to survival (P = 0.27). Unlike the two other staging systems, the CUPI was highly significant during the whole period of follow-up. Survival rates for different risk groups are shown in Table 4. It is noteworthy that the CUPI was more sensitive in differentiating groups of patients from the beginning of follow-up compared with the TNM staging system, in which the difference between stages was not observed before 30 weeks of survival (P = 0.15). With the goodness-of-fit test, it was found that the CUPI (P < 0.001) was more predictive of survival compared with the CLIP prognostic score (P < 0.001).
Table 4. Survival Rates for Different Risk groups According to the Chinese University Prognostic Index
Risk group (no.)
Survival rate (%)
Low risk (315)
Intermediate risk (346)
High risk (118)
Treatment for patients with HCC largely has been unsuccessful in the past. Recently, new modalities of treatment, such as liver transplantation, radiofrequency ablation, new systemic drugs or drug combinations, intra-arterial chemotherapy with or without embolization, and internal and external radiotherapy, are emerging. However, there is still no convincing evidence that overall survival can be improved with these treatments. There is wide recognition that the conventional TNM staging system has its limitation in stratifying patients with HCC into different risk groups, because nontumor-related factors, such as liver function status, although they are important in prognosis, are not included. A better staging system for patients with HCC should stratify patients more precisely according to their prognosis and should be reliable and useful for comparing treatment results among different modalities. A reliable staging system also could be used for patient stratification before entry into prospective, randomized, Phase III clinical trials to permit better delineation of treatment effects among different risk groups.
Our joint hepatoma clinic is formed by a team of oncologists, surgeons, and interventional radiologists with an interest in HCC. All patients with HCC, irrespective of their disease stage, are reviewed together by the team, and uniform treatment policies and protocols are carried out. Thus, patients in the current study formed an ideal cohort for development of a prognostic index.
The anatomic extent of a malignant neoplasm, without doubt, is the most important disease characteristic in terms of prognosis for patients with malignant disease in general. Therefore, Fielding et al.5 urged that the TNM staging system, which was established by Pierr Denoix in 1943, should remain as a reference classification system. The system also is easy to understand, remember, and apply in daily practice. The UICC also suggested that the inclusion of other disease-related factors should be encouraged, so that the assessment of prognosis by TNM staging can be improved.5 Therefore, we found that the most appropriate way was to identify independent prognostic variables from a large patient data base and add them to the conventional TNM staging system by applying appropriate statistical methodology, such as the Cox method.
In the current study, which was based on a large, homogenous cohort of 926 patients, we adopted such an approach to derive a new prognostic index, the CUPI, by refining the current TNM system. We have constructed a simple scoring system using variables that were objective measures and that were easily available at the time of diagnosis. Short-term survival at 3 months was chosen as the reference to determine the chance of survival, because this often is considered the minimum expected survival when selecting patients for treatment or entering into clinical trials. Selection bias has been minimized by including consecutive patients (apart from the 109 patients who were excluded due to incomplete data) over the study period, all of whom were analyzed for the identification of prognostic factors. This sample population was collected prospectively from a single center in Hong Kong, where patient investigation procedures and management are uniform. Together with the comprehensive inclusion of study variables, quality control of data, and satisfactory validation confirmed by an independent data set (the confirmatory set), this is probably the largest sample of patients for the study of prognostic factors in HCC in the literature to date.
Our study found that the CUPI is more discriminative over the first few months after presentation (Fig. 1) than TNM staging alone. This is important, because patients with HCC have a high mortality rate soon after diagnosis. However, the predictive power of TNM staging for survival is only obvious after 30 weeks. The CUPI also resulted in better overall discrimination compared with the Okuda staging system, the TNM stating system, or the CLIP score, as shown from the respective survival curves. The overall median survival of 19.0 weeks in the current study has improved compared with earlier studies from Hong Kong, in which median survivals of 3.5 weeks and 8.0 weeks were reported in 198113 and 1990,14 respectively. Advances in diagnosis and treatment may have contributed to the prolongation of survival in patients with HCC, but lead-time bias associated with an increasing number of patients with HCC detected in the asymptomatic phase also may have played a role. Our finding that asymptomatic disease on presentation was a significant independent prognostic factor also supports the latter explanation. However, the survival of our patients still seems to be shorter compared with other reported series on non-Asian populations. A North American study15 of 314 patients with HCC reported a median survival of 10 months, and another Italian study of patients with HCC reported a median survival of 11 months.11 The difference in survival between Southeast Asian and non-Asian populations with HCC may be due to the different etiology of the disease or more severe underlying liver disease.
The other significant variables identified in the current study—serum levels of TB,9, 15, 16 AFP,8, 15, 17 and ALP16 and the presence of ascites9—also were identified as significant variables in other studies. Some variables in the current study, such as abdominal pain and weight loss, are nondiscrete and subjective and are less useful as prognostic indicators. We did not include treatment modality as a variable, because the primary intention of this study was to construct a prognostic index that predicts survival based on information available at the time of diagnosis. We acknowledge that treatment may have a confounding effect on the calculation of the prognostic index. However, our group of patients was nonselected, and only 10.4% of them received treatment with curative intent, which was surgery. Other treatments may have an effect, but this should not be large, because, to date, none of these treatments has been shown convincingly to affect survival. Furthermore, there is great variation in the choice of treatment and treatment protocol across different centers. The inclusion of variations within a treatment regimen or its interaction with clinical characteristics would make the model difficult to apply routinely.
In the current study, vascular invasion also was not one of the studied variables, because its impact on survival already was reflected in the conventional TNM staging system for patients with HCC. This parameter is still valid in the new CUPI, which was built on the conventional TNM staging system. In managing the continuous variables, such as AFP and ALP levels, we manually set the cut-off at a value above what was considered clinically significant.
Eleven percent of patients (109 of 1035 patients) were excluded from the study because they had incomplete survival data. Most of these patients had advanced stage disease and returned to mainland China after they were seen once. The possible bias of the CUPI due to this group of patients with missing data is unlikely to be significant, because we could not detect any systemic differences between this group and the patients who remained in the study.
There are a number of other prognostic classification systems for patients with HCC other than the TNM staging system reported in the literature. The Okuda staging system6 proposed in 1985 has been used widely because of its simplicity. It was the first to incorporate liver function into the staging of patients with HCC. However, the estimation of liver involvement (> 50% or < 50% of the whole liver) is subjective and tedious, even with an imaging modality. When the Okuda staging system was applied to our cohort of patients, it became less discriminant with regard to survival for patients with Stage I and II disease (Fig. 2c). Concerning the two Japanese studies7, 8 and one Spanish study,9 regression coefficients of significant variables were used directly to estimate survival, but the complicated calculation that they required may limit their application in a clinical setting. Rigorous validation of the prognostic index using independent data sets has been reported only by Calvet et al.9 and Kakio et al.8 The studies by both Akashi et al.7 (90 patients) and Calvet et al.9 (206 patients) included an extensive list of patient characteristics with limited sample size. Another prospective study from France that included 761 patients also found that TB, AFP, ALP, portal vein obstruction, and Karnofsky performance score were significant prognostic variables.18 Those findings are similar to ours, except that TNM staging was not included in that study, and there was no comparison with other existing staging systems. More recently, the CLIP investigators studied 435 patients with HCC from 16 Italian institutions and identified 4 independent predictors of survival, including Child–Pugh stage of cirrhosis, tumor extent in the liver, AFP level, and the presence of portal venous thrombosis.10 The CLIP scoring system was able to classify patients into different risk groups better than the Child–Pugh and Okuda staging systems. However, the CLIP system was less discriminant for patients with early stage disease. Although it has been validated by a prospective study,11 the CLIP investigators also acknowledged that estimating the extent tumor is highly subjective. Furthermore, scores were assigned arbitrarily to the variables without considering the prognostic value for survival based on an objective statistical approach. When the CLIP scoring system was applied to our patients, we found that survival curves for different risk groups were overlapping (Fig. 2d). In contrast, our CUPI was more discriminant. Therefore, the CUPI may be a better prognostic system for patients with HBV-related HCC in our part of the world.
In summary, a new CUPI was constructed for patients with HCC based on large numbers of patients. The new CUPI is more discriminant in predicting survival compared with the conventional TNM staging system and with other reported systems. However, almost all of the patients in this study had HBV-related HCC (79%). Whether this is applicable to other patient cohorts with different etiologic factors is not known. The new CUPI also must be validated prospectively before it can be accepted for general use. This validation process is underway currently in our center.