Prognostic indicators in hepatocellular carcinoma: a systematic review of 72 studies

Background Although there are many studies of the predictors of death in hepatocellular carcinoma (HCC), most combine patients with and without cirrhosis and many combine those with compensated and decompensated cirrhosis. Objective To perform a systematic review of the literature evaluating the predictors of death in patients with cirrhosis and HCC and to evaluate whether the predictors differ between patients with compensated and decompensated cirrhosis. Methods Inclusion criteria: (i) publication in English, (ii) adult patients, (c) >80% of the patients had cirrhosis, (iv) follow-up >6 months and (v) multivariable analysis. Quality was based on the accepted quality criteria for prognostic studies. Results Of the 1106 references obtained, 947 were excluded because they did not meet the inclusion criteria. A total of 23 968 patients were included in 72 studies (median, 177/study); 77% male, median age 64, 55% Child–Pugh class A. The most robust predictors of death were portal vein thrombosis, tumour size, α-foetoprotein and Child–Pugh class. Sensitivity analysis using only 15 ‘good’ studies and 22 studies in which all patients had cirrhosis yielded the same variables. In the studies including mostly compensated or decompensated patients, the predictors were both liver and tumour related. However, these studies were few and the results were not robust. Conclusions This systematic review of 72 studies shows that the most robust predictors of death in patients with cirrhosis and HCC are tumour related and liver related. Future prognostic studies should include these predictors and should be performed in specific patient populations to determine whether specific prognostic indicators are more relevant at different stages of cirrhosis.


Abstract
Background: Although there are many studies of the predictors of death in hepatocellular carcinoma (HCC), most combine patients with and without cirrhosis and many combine those with compensated and decompensated cirrhosis. Objective: To perform a systematic review of the literature evaluating the predictors of death in patients with cirrhosis and HCC and to evaluate whether the predictors differ between patients with compensated and decompensated cirrhosis. Methods: Inclusion criteria: (i) publication in English, (ii) adult patients, (c) 4 80% of the patients had cirrhosis, (iv) followup 4 6 months and (v) multivariable analysis. Quality was based on the accepted quality criteria for prognostic studies. Results: Of the 1106 references obtained, 947 were excluded because they did not meet the inclusion criteria. A total of 23 968 patients were included in 72 studies (median, 177/study); 77% male, median age 64, 55% Child-Pugh class A. The most robust predictors of death were portal vein thrombosis, tumour size, a-foetoprotein and Child-Pugh class. Sensitivity analysis using only 15 'good' studies and 22 studies in which all patients had cirrhosis yielded the same variables. In the studies including mostly compensated or decompensated patients, the predictors were both liver and tumour related. However, these studies were few and the results were not robust. Conclusions: This systematic review of 72 studies shows that the most robust predictors of death in patients with cirrhosis and HCC are tumour related and liver related. Future prognostic studies should include these predictors and should be performed in specific patient populations to determine whether specific prognostic indicators are more relevant at different stages of cirrhosis.
Hepatocellular carcinoma (HCC) is responsible for significant morbidity and mortality in cirrhosis. It commonly leads to decompensation of cirrhosis and is one of the leading causes of death in cirrhotic patients (1,2). Identifying the accurate prognostic indicators of death for HCC allows the provider to counsel individual patients and also forms the basis of any decision-making process. Most cases of HCC in the western world occur in the setting of cirrhosis and, therefore, prognosis is determined not only by factors related to the tumour but also by factors related to cirrhosis. In fact, current prognostic models for HCC include parameters of liver dysfunction and parameters related to HCC (3)(4)(5). However, studies on the prediction of death in HCC include patients both with and without underlying cirrhosis, and this heterogeneity may impact the clinical utility of current prognostic models.
Additionally, one could hypothesize that, as in cirrhosis not associated with HCC (2), the prognostic factors for HCC would be different in patients with underlying compensated vs. decompensated cirrhosis, with factors related to the tumour having a greater prognostic significance in the former and factors related to both tumour and liver disease being more important in the latter. The purpose of this systematic review is to evaluate the predictors of death in patients with HCC and underlying cirrhosis to determine whether the predictors differ between patients with compensated and decompensated cirrhosis. Studies were selected using the following inclusion criteria: (i) English language, (ii) inclusion of only adult patients, (iii) established diagnosis of HCC, (iv) 4 80% of patients included had cirrhosis -if cirrhosis was not specifically mentioned, it was assumed that the population was comprised of cirrhotic patients if all patients in the study were classified into Child-Pugh classes, (v) survival analysis was reported, (vi) multivariate analysis of prognostic indicators of mortality was performed and (vii) patients in the study were either untreated or the study included more than one type of therapy following a local treatment protocol. Trials were excluded if they had (i) o 80% cirrhosis, (ii) only prediction of short-term survival (o6 months), (iii) a single treatment option was being analysed, (iv) randomized trials comparing two different therapies, (v) the main predictive factor being studied was a tumour histological feature (i.e. tumour staining) and (vi) abstract only or full-text article not available. The rationale for excluding trials with only a single treatment option was that as candidates for particular treatment options may have varying prognoses, this may have resulted in a biased inclusion of patients. Required clinical data and data regarding study validity were predefined and collected for all studies meeting the inclusion criteria.

Methods
Study validity was assessed using predefined quality criteria, as listed in Table 1 (2,(6)(7)(8). Because none of the trials met all the quality criteria (Table 1), 'good quality' studies were arbitrarily defined as those that fulfilled all four of the following major quality criteria: (i) enrollment of consecutive patients, (ii) listing relevant baseline data, (iii) reporting the number of deaths and (iv) the absence of overfitting (ratio of the number of deaths/the number of variables 4 10). All 'good quality' studies documented the number of patients with cirrhosis.
An 'advanced' tumour was arbitrarily defined as any of the following: 'multinodular' tumours or tumours exceeding the Milan criteria (one tumour 4 5 cm or three tumors, with any of them being 4 3 cm) (9). In the analysis aimed at determining whether the prognostic variables differed between patients with advanced vs. non-advanced tumours, studies including mostly advanced tumours were defined as those in which 66% of the patients had advanced tumours and studies including mostly non-advanced tumours were defined as those in which 4 66% of the patients had non-advanced tumours. Decompensated cirrhosis was defined as a patient with Child-Pugh B or C classification. In the analysis aimed at determining whether the prognostic variables differed between patients with compensated vs. decompensated cirrhosis, studies including mostly compensated patients were arbitrarily defined as those in which 4 66% of the patients were Child-Pugh A and studies including mostly decompensated patients were arbitrarily defined as those in which 4 66% of the patients were Child-Pugh B or C.
The 1-, 2-, 3-and 5-year cumulative survivals and the final cumulative survival were recorded if present. The first five significant prognostic indicators in each multivariate analysis were recorded as were the number of studies evaluating each variable. A sensitivity analysis was performed using the 'good quality' studies and using the studies in which 100% of the patients had cirrhosis.

Results
As shown in Figure 1, of the 1106 references identified through a Medline search, 947 were excluded on analysis of the abstract provided because they did not meet the inclusion and exclusion criteria. Of the remaining 159 full-text articles, 88 were excluded, 84 because they did not meet the inclusion criteria (at least 35 had combined cirrhotic and non-cirrhotic patients, with o 80% being cirrhotic and 10 did not even mention whether there was underlying cirrhosis) and four because the  full-text paper was not found, leaving 71 articles (4, 10-79) meeting the criteria for inclusion in the review. One trial (46) was included twice as it had two separate analyses for two different age groups, bringing the total number of studies reviewed to 72.

Description of the prognostic studies
Regarding the aim of the study, 69% (50/72) of the studies were predictive studies, i.e. they assessed the prognostic value of multiple variables without concern on the pathophysiology of the variable, and 31% (22/72) were explanatory, i.e. a specific variable was evaluated according to a biologically plausible hypothesis (Table 1). Regarding the design of the study, 71% were retrospective and only 29% were prospective. Only 57% of the studies enrolled patients consecutively. The least commonly fulfilled quality criteria were the presence of an inception cohort (3%), internal/external validation of results (10%) and reporting of loss to follow-up (18%). No study fulfilled all quality criteria. Fifteen studies met our definition for a 'good quality study' , i.e. they fulfilled the four major predefined quality criteria (4,21,26,31,33,34,38,44,53,56,58,66,68,71,77). The geographical origin of the majority of the studies was Japan or Italy (Table 1). Eighty-two per cent of the articles outlined their diagnostic criteria for HCC. This included various combinations of compatible histology, radiology and a-foetoprotein (AFP) investigations.

Description of patients included in the prognostic studies
Overall, 23 968 patients were included in the 72 studies. Their characteristics are reported in Table 2. The median number of patients per trial was 177, with a range from 30 to 4525. The median age was 66, with 71% males. Fifty-two of the included studies explicitly stated the number of patients with cirrhosis.
Although the other 20 studies did not, they did divide all patients by the Child-Pugh classification which is, by convention, reserved for cirrhotic patients. None of these 20 studies met our predefined quality criteria.

Survival in the prognostic studies
As shown in Table 2, the median follow-up time was 21 months, with a mortality rate of 69%. The median survival time was 16 months. The 1-, 3-year and final cumulative survival were 66, 36 and 24% respectively. The median survival in studies including patients with predominantly decompensated cirrhosis (n = 3 studies with available data) (38, 54, 62) was 8 months as compared with 29 months in those including patients with predominantly compensated cirrhosis (n = 4 studies) (39,51,59,74). Twenty-eight studies reported the causes of death. The most common cause of death was recorded as being related to HCC in the majority (59%), followed by progression of liver disease (28%) ( Table 2).

Prognostic variables
A total of 79 variables were evaluated in these studies ( Table 3). The five most common independent predictors of death in HCC were portal vein thrombosis, tumour size, AFP, Child-Pugh class and bilirubin (Table 4). Of these, the variables that were found to be most significant in a larger number of studies were portal vein thrombosis, tumour size and Child-Pugh class (each of them evaluated in more than 30 studies and found to be significant in 4 50% of them  Table 5 and, of these, lack of therapy and performance status are remarkable because they were found to be predictive of death in over two-thirds of the 15 and eight studies in which  (Table 6), the same most common predictors of death are observed as for the overall analysis, i.e. Child-Pugh class, AFP, portal vein thrombosis, CLIP score and tumour size. The most robust predictors were Child-Pugh class, AFP and portal vein thrombosis. As the majority of the studies (89%) evaluating Child-Pugh status divided patients by class (A, B, C) and not by score, the term Child-Pugh class is utilized in this review.
The number of studies that included mostly compensated or decompensated cirrhotic patients was small. However, as shown in Table 8, in both groups of patients, the predictors of death included both liver-related and tumour-related factors. When patients were separated by advanced or non-advanced tumour status, the most important predictors of death in patients with advanced tumours were portal vein thrombosis, AFP, bilirubin and lack of treatment ( Table 9). The number of studies analysing significant predictors of death in patients with nonadvanced tumours was small.

Discussion
Hepatocellular carcinoma can arise in both non-cirrhotic and cirrhotic livers. In the Orient, where hepatitis B or toxins are the most common underlying causes, HCC commonly arises in the absence of cirrhosis (80,81). In the west, where hepatitis C and alcohol are the most common underlying liver diseases, HCC arises mostly in the setting of cirrhosis (81)(82)(83). Prognosis is predictably worse in patients with underlying cirrhosis (81,84). Even in patients with cirrhosis (without HCC), it has recently been shown that the prognostic factors, survival and causes of death differ significantly between those with compensated and decompensated cirrhosis (2) and that these two entities should be considered separately both in clinical practice and in clinical research (85). In fact, HCC is an independent predictor of death in patients with decompensated cirrhosis (2).
Prognosis is an essential part of the assessment of patients with HCC. It allows the patient to make important decisions,      One of the strengths of this systematic review is that it analysed predictors of death in HCC, specifically in studies in which Z80% of the patients had underlying cirrhosis. Furthermore, it aimed to analyse whether the predictors differed in patients with compensated vs. decompensated cirrhosis.
Despite methodological problems in the evaluated studies (none met all quality criteria and only 21% of them met important quality criteria), this systematic review allowed for the identification of the 'robust' predictors of death. As could have probably been predicted, these were both tumour related (portal vein thrombosis, tumour size and AFP) and cirrhosis related (mainly, the Child-Pugh class). The strength or the robustness of a predictor is given by the ratio between the number of studies in which each variable was significant and the number of studies in which it was assessed. With a larger number of studies, a large ratio is an indirect measure of the validity as each study that confirms the predictive value of a variable provides indirect proof of its validity (2). This robustness, which is independent of the quality of the studies, was particularly true for portal vein thrombosis, tumour size and Child-Pugh class, each of which was evaluated in more than 30 studies with a ratio 4 50%, i.e. more than half of the studies proved them to be among the first five most significant variables on a multivariable analysis. Importantly, the same parameters were the most frequent significant variables when only the studies in which 100% of the patients included had cirrhosis and when 'good' quality studies were analysed. In the analysis of good quality studies, one of the most robust predictors of death was the AFP.
Interestingly, one of the most commonly used HCC staging systems, the CLIP system (4), includes all four of these predictors (portal vein thrombosis, tumour size, Child-Pugh class and AFP) and, in fact, the CLIP staging system itself was identified as the sixth most common predictor of mortality, being among the first five significant variables in 11/15 studies (73%). Six of the 11 studies identifying the CLIP score as an independent predictor were of Italian origin (24,26,36,47,53,72). Another commonly used HCC staging system, and the one recommended in recently published guidelines (81), is the Barcelona Clinic Liver Cancer (BCLC) system (5), which was only evaluated in three studies, but was nevertheless found to be among the first five significant variables in two (67%) of them. This staging system also includes three of the predictors (tumour size, Child-Pugh class and portal vein thrombosis). Although it does not include AFP, the BCLC staging system has the advantage of including other parameters such as performance status (found to be an important predictor in 6/8 or 75% of the studies in which it was assessed) and, importantly, of tailoring other prognostic factors to different tumour stages (early vs. intermediate/advanced) and to Child-Pugh class. For example, portal pressure, as determined by the hepatic venous pressure gradient, which was not directly evaluated in any of the studies in this analysis, was shown to be an independent predictor of death in early stages (patients subjected to resection) (86,87) and is included in the BCLC system but only in Child A patients with resectable tumours (5). In the BCLC multivariable analysis study that was performed to identify the predictors of death in intermediate/advanced stages, an AFP of 4 35 ng/ml was not significant on univariate analysis (23), whereas in the CLIP staging system an AFP of 4 400 ng/dl had an independent  DCP  4  5  80  Tumour size  2  2  100  Bilirubin  4  5  80  AFP  2  2  100  Child-Pugh class  4  7  57  Albumin  2  3  67  AFP  4  11  36  CLIP  3  4  75  Treatment received  3  4  75  Tumour size  3  4  75 AFP, a-foetoprotein; CLIP, Cancer of the Liver Italian Program; DCP, des-gamma-carboxy prothrombin. prognostic value (4). As has been recently discussed, the prognostic utility of AFP is unclear in part due to the use of heterogeneous cut-off levels and also due to its variable sensitivity (88,89). For example, in a recent study of 1158 patients (60% Child-Pugh A), even though the specificity of an AFP level 4 600 ng/ml in predicting survival was 93%, its sensitivity was only 23% (89). Therefore, although the BCLC model has been reported to be superior to alternate models for the prediction of survival in HCC, including the Okuda, CLIP, tumour node metastasis, Japanese Integrated System, Groupe d'Etude de Traitement du Carcinoma Hepatocellulaire and Chinese University Prognostic Index models (61,67), three of which include the AFP, it would be interesting to see whether the AFP (as a continuous variable) adds to the prognostic value of the BCLC. The appropriate cutoff level and the group of patients in which the AFP may be helpful remains to be determined, although our analysis would suggest that it appears to be more useful in advanced tumour stages ( Table 9).
The importance of determining prognosis at different stages of cirrhosis cannot be overemphasized (2). We had hypothesized that in the compensated patients factors related to the tumour would be more important, whereas in the decompensated patient, both liver-and tumour-related factors would be important. However, in a subanalysis of studies including mostly patients with compensated or decompensated cirrhosis (Table 8), both tumour-related and cirrhosis-related predictors appeared to be predictive of death in both groups. However, the number of studies is too small to draw firm conclusions, particularly regarding patients with mostly decompensated cirrhosis and those with mostly non-advanced tumour stage.
Our study is limited by the design and prognostic parameters chosen in each of the studies included in this analysis as well as the heterogeneous group of patients included with both treated and untreated tumours. Furthermore, although the choice of the 66% cut-off for the definition of 'mostly decompensated cirrhosis' or 'mostly non-advanced tumour' was in part arbitrary, it can be justified because even at this cut-off there were too few studies available in the 4 66% non-advanced tumour or 4 66% decompensated cirrhosis categories to obtain definitive results.
Future studies on the prognostic indicators of HCC should include patients either with or without cirrhosis. Studies that will apply to patients with HCC seen in western countries should include only patients with cirrhosis and the prognostic variables should be assessed separately for the different stages of cirrhosis, at a minimum, separating those with compensated and those with decompensated cirrhosis. The study quality should be optimized by incorporating most, if not all, of the criteria listed in Table 1. Additionally, parameters identified by the majority of the studies included in this analysis should be included in such studies.