HCC in Older Patients

Transcription

1 Dig Dis Sci (2010) 55: DOI /s ORIGINAL ARTICLE HCC in Older Patients Brian I. Carr Petr Pancoska Robert A. Branch Received: 2 December 2009 / Accepted: 19 February 2010 / Published online: 18 March 2010 Ó Springer Science+Business Media, LLC 2010 Abstract Background Hepatocellular carcinoma (HCC) is a heterogeneous disease, with many poorly-defined prognostic patient subsets. Identification of discrete subsets will aid rational patient and treatment selection. Methods A database with 778 biopsy-proven, unresectable and untransplantable HCC patients who were followed from diagnosis till death was interrogated. Using a moving average algorithm, patients were ordered by survival and then survival cohorts were analyzed according to standard liver function and CT characteristic parameters. Results We found characteristic age clustering groupings by survival. In the older age patients, two survival sub-groups were identified, with days in one and 330 1,250 days survival in the other group. The longer surviving group had the lowest serum bilirubin and AFP levels and the lowest tumor mass. Remarkably, the trends for both AFP and bilirubin were similar, suggesting that they were not independent variables. This idea was supported by the similar correlation of typical AFP with GGTP, ALKP and SGOT levels. Conclusions A large HCC cohort showed significant age clustering characteristics for survival, especially in older patients. AFP, bilirubin and age were found to be interrelated factors for HCC severity and survival. B. I. Carr (&) Department of Medical Oncology, Liver Tumor Program, Kimmel Cancer Center, Thomas Jefferson University, Bluemle Building, Room 1002, 233 S.10th Street, Philadelphia, PA 19107, USA brian.carr@jefferson.edu P. Pancoska R. A. Branch Center for Clinical Pharmacology, University of Pittsburgh, Pittsburgh, PA, USA Keywords Hepatoma AFP Survival Elderly Abbreviations HCC Hepatocellular carcinoma AFP Alpha-fetoprotein GGTP Gamma glutamyl transpeptidase ALKP Alkaline phosphatase CT Computerized axial tomography scan Introduction Meaningful analysis of clinical practice data is in general a complicated task. In this study we present a special case of clinical practice data of a large cohort of HCC patients where the unique design, together with a tailored data analysis approach, provided novel insight into the factors behind the HCC as well as tools for survival prognosis. Most hepatocellular carcinoma (HCC) arises on the basis of chronic hepatitis or cirrhosis or both. The prognosis is considered to depend on tumor factors such as serum alpha-fetoprotein (AFP) levels and tumor size, as well as liver factors such as serum bilirubin, gamma glutamyl transpeptidase (GGTP), alkaline phosphatase (ALKP) and transaminase levels. It is likely that there is even an interaction between these processes, since cirrhosis is a pre-malignant condition, leading to either liver failure, HCC or both. This dual set of influences was initially reflected in the staging system of Okuda et al. [1] and subsequently by many other systems in which greater complexity was taken into account in an effort to identify prognostic subsets for survival [2 7]. AFP is included amongst many of these HCC scoring and classification

2 Dig Dis Sci (2010) 55: systems, as its higher levels have been shown in multiple published series to represent worse prognosis [8 12]. Here we show quantitative evidence that while HCC is a heterogeneous disease, differences between its subcategories can be clearly articulated with the help of newly observed coherences between the levels of several clinical parameters. Our findings are that patients older than 55 years with non-resectable HCC can be categorized into multiple, clinically diverse, but internally homogeneous groupings. These groupings exhibit two unique and very different survival ranges, one being long ( days) and the other short (45 80 days). Methods Patients Seven hundred seventy-eight patients with unresectable and biopsy-proven HCC, who had a complete clinical dataset, were followed from diagnosis till death. They were unresectable due either to the presence of multiple and bilobar liver lesions, the presence of portal venous thrombosis or to abnormal liver function parameters and were treated with cisplatin-based chemoembolization. Therefore, no censoring was needed in the data processing. Their survival time was recorded. All patients had baseline liver function tests, including GGTP (gamma glutamyl transpeptidase) levels, AFP (alpha-fetoprotein) serum tumor marker levels and baseline helical CT (computerized axial tomography) scans performed. Analysis Strategy Our processing started from visualization of all possible correlations of survival-ordered serum parameter value pairs. This revealed that the uncertainty of the actual disease status of patients that come for treatment in random stages of the disease obscures details of the underlying data structures. We therefore used a moving average processing method [13 15] with the optimized size of the averaging window (the number of patients with the most similar survival times in one group, whose clinical parameters were averaged to characterize the mean survival of any such group). Our optimization algorithm [16] resulted in a size of 101 patients in these subcohorts. Consequently, we were able to characterize 678 survival times. Each of these central survivals was characterized by individual clinical data obtained within a 101-patient subgroup, with 50 patients having shorter survival, closest to the central survival, and 50 patients having longer survival, closest to the central one (50? 1? 50 = 101). The complete data processed in this way covered the interval of actual survivals ranging from 30 to 1,250 days. We systematically checked by standard statistical methods that the survivaldependent means for all parameters were not influenced by biases and problems caused by presence of outliers or discontinuities in the variances of the data as we moved along this survival interval. With this averaging, we lost to a certain degree the relevance of our results for an individual patient. We therefore included a final step in our analysis in which we took advantage of the fact that in all projections, all 678 subcohorts of 101 patients were identical. This allowed us to compute for each of the 101-subcohorts the exact percentages of patients with actual (i.e., non-averaged), normal and abnormal levels of all clinical serum parameters. In this way, we could have the best of both typical characterization of the disease (revealing more information about the underlying biology or pathophysiology of HCC), as well as an individual patient s characterization of the survival category in terms of probabilities of the correct prognosis (described in detail below). A multi-parameter coherence observed as significant in one or more twoparameter projections provides the tool for starting to build up the steps of a multi-parameter survival classification scheme. The significance of results for individual patients is such that we designed our filtering algorithm to preserve the exact identity of all 101 patients in all 678 characterized survivals in all pair-wise parameter projections. This allowed us to provide probabilistic characterization of clinical state for patients in every subcohort [17]. We used the 101 patients in all 678 subcohorts to estimate the probability that out of 100 individual patients of age 55 and older presenting at baseline with HCC with the actual outcome within either the 350 1,250 day survival category or the day survival category, that each will have a certain survival-category unique (and thus recognizable with confidence) combination of their clinical parameter levels. For this purpose, we first dichotomized all parameters into normal and elevated categories by coding their actual baseline parameter levels as normal (code 1) or abnormal (code 0), using the generally accepted serum parameter thresholds. These dichotomizing parameter codes that were determined from individual data for each patient in the cohort were then ordered according to patient actual survival and processed by a moving average window of ±50 patients. We then multiplied the resulting mean value of the parameter factor for every survival by 101 to recover the number N of patients with normal levels of any parameter in each 101 patient subcohort that was used to characterize survival. The number of patients with elevated levels of each parameter in each subcohort was therefore E = [101-N]. We then computed the ratio of normal to elevated patients at each subcohort R = [N/E]. This ratio

3 3586 Dig Dis Sci (2010) 55: was corrected for bias of these patient parameter levels in the whole cohort. To this effect, we determined X being % of patients with normal parameter levels and Y being % of patients with elevated parameter levels in the total cohort. Then, we computed the adjusted normal/elevated ratio AR = [Y/X] 9 R. AR[50 means that in the 101 cohort used to calculate typical values of each parameter levels for each survival, there were more patients with normal parameter levels. AR \50.0 means that there were more patients with elevated parameter levels in that subcohort. Results Age Clustering and Survival Figure 1a shows the correlation between the survival (logarithmic survival scale, re-labeled in actual survival) with typical serum total bilirubin levels. It shows that our filtering quantifies the well-described adverse prognostic effects of elevated serum bilirubin levels for survival, with higher bilirubin levels correlating with poorer survival. Our strategy in interpreting these results was to discover additional coherences in the data by overlaying the values of an additional parameter on the pair-wise correlations of typical parameters. We applied this strategy in Fig. 1 to show Fig. 1 Correlation between survival and bilirubin (a) or AFP (b) showing age clusterings that, in addition to coherence between the bilirubin and survival, there was also coherence to patient age in the subcohorts, indicated by a thick color and numbering that show two different sub-groups of patients, with typical ages above 55 years (47% of the total) showing systematic and linearly correlated increase of typical bilirubin levels with decreasing survival. These two groups are also clearly separated by the length of their survival, with days in the shorter surviving group and 330 1,250 days in the longer surviving group. Thus, if age was not an important factor for survival categorization, we would expect typical ages to be randomly distributed and not clustering without interruption, as we found. Thus, Fig. 1 indicates the importance of age as an essential part of our categorization scheme that identifies HCC subtypes. Given the prognostic importance of serum AFP levels that has been shown in many other studies, we took the same approach with serum AFP levels as we did with bilirubin (Fig. 1b). Although the AFP, in contrast to bilirubin, is considered an indicator of tumor growth, we found a very similar linear correlation in trends between serum AFP levels and survival for these two older patient groups. Interestingly, these trends between both bilirubin and AFP and survival are observably different for the subcohort of younger patients. Tumor Mass and Survival We subsequently concentrated exclusively on the older patient subgroups, because they show two completely different survival ranges, but a similar trend between survival and serum parameters. We examined the tumor burden for older patients within the long and short survival subgroups. We used the total tumor amount for this purpose, calculated as the product of the average tumor size and the total tumor number in each patient group. Figure 2a shows the relationship of total tumor burden to survival, with the longest survivors having the lowest tumor burden. The oldest patients, [80 years, were in this longest surviving subset. This estimate of total tumor mass was then plotted against typical serum AFP levels, for each survival-characteristic patient group, as shown in Fig. 2b. Because of the linear relationship between the typical AFP levels and survival (Fig. 1b), ordering of typical AFP levels on the X-axis of Fig. 2a also reflects an ordering according to survival. Thus, patients with small tumor masses and low typical AFP levels have survival on the 1,250 days end of the long-survival interval, while patients with AFP levels close to 250 and relative tumor mass of about 40 have a survival of about a year, which is the short end of the long survival interval. A similar dependence holds within the 1,000 1,000,000 ng/ml interval of AFP levels within the short (45 80 days) survival interval. We then plotted the total tumor mass against typical bilirubin levels

4 Dig Dis Sci (2010) 55: bilirubin and AFP levels follow synchronously and linearly the increase of the tumor patient tumor mass with decreasing survival (Fig. 2b). Relationships Between Tumor Factors and Liver Factors Fig. 2 Correlation of trends between tumor mass and survival (a), AFP (b) and bilirubin levels (c) for the two groupings of older patients with typical ages [55 years (Fig. 2c). This also showed a linear correlation between total tumor mass and bilirubin levels, within both the short and long survival groups, with one exception. For the subgroup of oldest patients (typically [80 years, about 10% of all patients, yellow age-color code), where the increase of the tumor mass indicates linear decrease of survival (see Figure 2), this trend is not reflected in the correlated change of bilirubin levels, which remains under the normal threshold of 1.2 mg/dl. By contrast, we showed in Fig. 2b, that for this oldest patient population, typical AFP levels increase synchronously with the increasing tumor size and decreasing survival at the longer end of the long-survival category. For the remaining patients (typical age years), both The similarity between the trends in survival for both AFP and bilirubin levels shown in Fig. 1 for the two survival groups of older patients led us to consider the possibility that tumor factors and liver factors were not completely independent, but might be somehow connected. To investigate this possibility, we plotted several serum parameters against typical AFP, including serum bilrubin, ALKP, SGOT and GGTP (Fig. 3). This identified additional coherences in the typical serum bilirubin data, shown for these two age-defined subgroups, as indicated. Again, because of the linear relationships between AFP and survival (Fig. 1), which are characteristic for this age group of patients, the ordering by AFP on the X-axis of the plots in Fig. 3 reproduces the survival ordering of patients, from the longest survival at the lowest AFP levels to the shortest survival at the highest AFP levels. This ordering is valid both globally and internally, within both survival intervals of 330 1,250 and days. Figure 3 shows a significant increase in the typical levels of all parameters between the long (330 1,250 days) and short (45 80 days) survival groups. There is a discontinuity in the AFP (=survival) versus parameter trends at the boundary between the long (left) and short (right) panels (Fig. 3). The plot shows a correlation between bilirubin and AFP values in both older age categories. The correlation is linear, with higher AFP values correlating with higher bilirubin values in each subgroup. When we overlaid patient ages onto this distinct correlation between the parameters indicating liver failure and tumor growth, the complete set of subcohorts identified in the survivalbilirubin relationship of the HCC-multidimensionality projection was transferred without exception or interruption into the bilirubin-afp projection of the HCC multidimensional characteristics consistently for both survival groups. These near identities of these different parameter relationships which we show in Figs. 1 and 3a are striking. This shows not only that bilirubin and age are coherently related to the HCC survival, but also that AFP and bilirubin with age are interrelated as factors underlying HCC severity and survival. Figure 3b d provides similar results for serum ALKP, SGOT and GGTP levels, when plotted against AFP values. These panels indicate that the coherence between all the considered serum-based parameters with age and survival is a general feature of HCC in these two older age categories, and clearly identifies the two internally homogeneous and mutually heterogeneous

5 3588 Dig Dis Sci (2010) 55: their baseline clinical characteristics that could be useful in prognosis. These are, first, the linear relationships between survival and typical levels of baseline parameters, and second, the coherence between trends between the various baseline parameters. Discussion Fig. 3 Correlation of trends of AFP to total bilirubin (a), ALKP (b), SGOT (c) and GGTP (d) levels for two groupings of HCC patients [55 years within the two survival categories subtypes of HCC that continuously combine into a resulting final HCC status that additionally reflects the better or worse survival chance within the two older age categories. This approach revealed two aspects of these patients and This study was based on an analysis of clinical practice data from 778 unresectable HCC patients presenting at a single institution, who were followed till death, without exception. This permitted use of the hard endpoint of overall survival. In order to reduce the stochasticity inherent to this type of data, we used a minimal amount of averaging to allow us to see the heterogeneity within the patient clinical parameters. We concentrated on coherences within the trends of multiple clinical parameters, so as to identify clinically relevant factors behind the HCC heterogeneity that might influence patient survival. For this approach, we started by ordering patients according to their actual survival as shown in Fig. 1. We plotted serum bilirubin and AFP levels against survival, since they are important parameters of liver failure and tumor aggressiveness, respectively. Surprisingly, the trends were remarkably similar for the distinct survival intervals, with each coherency being typical for different age groupings. The trends between survival and each parameter were simpler for the two groups of older patients [55 years, on which we subsequently focused. The two older age groups represented survival-related phenotypes of patients with days survival for ages 55 70, and 330 1,250 days for patients aged [55 years, including the oldest patients [80 years with survival of 500 1,250 days. These same age coherences were subsequently also seen when we plotted clinical parameters against each other. The trends for these two older patient categories of shorter and longer survival were then presented in one combined plot that was separated into two parts according to survival (Figs. 2 and 3). This allowed us to present their similarities and differences for each parameter under study. To better characterize the survival heterogeneity of these older patients, we examined their tumor characteristics as a function of the CT scan parameters of tumor number and size. We estimated total tumor burden as a product of baseline tumor number and median tumor size and plotted typical values against survival, serum AFP and serum bilirubin levels (Fig. 2). The shorter survival phenotype was clearly associated with the highest tumor burden. It was also associated with the highest serum AFP and bilirubin levels (Fig. 2a, b) as expected. By contrast, in the longer survival group, the serum AFP continued to be inversely proportional to survival. However, for bilirubin

6 Dig Dis Sci (2010) 55: levels, there was a difference in the trends for the longer survivors between 55 and 70 years old and [80 years old. The former displayed inverse proportionality, while the latter had a constant normal bilirubin level despite increasing tumor mass and AFP, with decreasing survival (Fig. 2c, bottom left). In order to examine why there were such similar trends in Fig. 1 for AFP and bilirubin in relation to survival, despite their representing the apparently different processes of liver damage and tumor aggressiveness, we then plotted each of the additional clinical serum parameters (ALKP, SGOT and GGTP) directly against AFP levels (Fig. 3). These plots showed two extreme types of behavior. In one, as shown in Fig. 3d, there is a linear relationship between AFP and GGTP levels that continued throughout both survival categories, without discontinuity. In the other, as shown in Fig. 3a, there are two discontinuities in the AFP and bilirubin plot. The first discontinuity occurs at AFP levels of 25 ng/ml and the second discontinuity occurs at the boundary of the longer and shorter survival groups. ALKP and SGOT plots (Fig. 3b, c) are intermediate between these two extreme patterns and do not differentiate the oldest age group in the way that the bilirubin plot does. These plots suggest to us that our simple notion of two largely independent processes of liver damage and tumor aggressiveness might need modification. This is because for each liver parameter that we plotted against AFP, our surrogate for tumor aggressiveness, we found a coherent trend between AFP levels and the liver parameter levels. This would not be expected if they were indeed independent processes. One example where these two processes are independent however, is the relationship of serum bilirubin to AFP levels in the oldest and longest surviving patients with[600 days, as depicted in Fig. 3a (extreme left end of the plot). In contrast, as serum AFP levels rise beyond 25 ng/ml, there is a linear relationship between AFP and bilirubin levels, showing their inter-dependency (rest of Fig. 3a). In this respect, bilirubin is unique, since as shown in Fig. 3b d, neither ALKP, GGTP nor SGOT levels showed the transition from independency to inter-dependency. These three parameters are thus never independent of tumor aggressiveness, as judged by AFP levels. Our dataset is unique in several aspects. First, there were no subjects who were censored or lost to follow-up. All patients were followed by a single clinician till death. Second, we have published previously an overview [18] of clinically relevant features of the raw data. For the purpose of this paper, we corrected for all un-informative biases of the data by mathematical transformations (conversion into log-transformed data, use of relative units for better comparability of parameter levels, etc.) before attempting the final data processing described in this paper. We searched for evidence for well-defined heterogeneous structure of the HCC data in this clinical study with each subtype being associated with very different survival. We characterize this heterogeneity in HCC using the combination of two complementary approaches. The first was that clinical parameter values at baseline represent a snapshot of multivariate pathophysiological factors, which operate in concert and in their unity determine the progression of the disease, quantified by multiplicities in the times to survival. In the second, we hypothesized that the projection of these complex multi-variable inter-dependencies into simpler two-parameter correlations of data, ordered systematically according to the survival of individual patients, would facilitate clear identification of the boundaries between the heterogeneous HCC subcategories. Ordering all data according to the survival before actual processing, and keeping this survival-defined order unchanged throughout the complete analysis, is an essential feature of our approach. Thus, any coherences observed for clinical parameters in the data are validating the clinical relevance of such observations, since without functional correlation between such coherently behaving parameters and the survival outcome, we can only expect random distributions of the other parameter values in the survival-ordered data vectors. References 1. Okuda K, Ohtsuki T, Obata H, et al. Natural history of hepatocellular carcinoma and prognosis in relation to treatment. Study of 850 patients. Cancer. 1985;56: [No authors listed]. A new prognostic system for hepatocellular carcinoma: a retrospective study of 435 patients: the Cancer of the Liver Italian Program (CLIP) investigators. Hepatology. 1998;28: Georgiades CS, Liapi E, Frangakis C, et al. Prognostic accuracy of 12 liver staging systems in patients with unresectable hepatocellular carcinoma treated with transarterial chemoembolization. J Vasc Interv Radiol. 2006;17: Kondo K, Chijiiwa K, Nagano M, et al. Comparison of seven prognostic staging systems in patients who undergo hepatectomy for hepatocellular carcinoma. Hepatogastroenterology. 2007;54: Leung TW, Tang AM, Zee B, et al. 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: a study based on 926 patients. Cancer. 2002;94: Llovet JM, Bru C, Bruix J. Prognosis of hepatocellular carcinoma: the BCLC staging classification. Semin Liver Dis. 1999; 19: Minagawa M, Ikai I, Matsuyama Y, Yamaoka Y, Makuuchi M. Staging of hepatocellular carcinoma: assessment of the Japanese TNM and AJCC/UICC TNM systems in a cohort of 13,772 patients in Japan. Ann Surg. 2007;245: Furihata T, Sawada T, Kita J, et al. Serum alpha-fetoprotein level per tumor volume reflects prognosis in patients with hepatocellular

7 3590 Dig Dis Sci (2010) 55: carcinoma after curative hepatectomy. Hepatogastroenterology. 2008;55(86 87): Yen YH, Changchien CS, Wang JH, et al. A modified TNMbased Japan Integrated Score combined with AFP level may serve as a better staging system for early-stage predominant hepatocellular carcinoma patients. Dig Liver Dis. 2009;41(6): Nomura F, Ohnishi K, Tanabe Y. Clinical features and prognosis of hepatocellular carcinoma with reference to serum alpha-fetoprotein levels. Analysis of 606 patients. Cancer. 1989;64(8): Dvorchik I, Carr BI. A simple prognostic scoring system for patients with unresectable hepatocellular carcinoma treated by chemo-embolization. Cancer Detect Prev. 2007;31(2): Peng SY, Chen WJ, Lai PL, Jeng YM, Sheu JC, Hsu HC. High alpha-fetoprotein level correlates with high stage, early recurrence and poor prognosis of hepatocellular carcinoma: significance of hepatitis virus infection, age, p53 and beta-catenin mutations. Int J Cancer. 2004;112(1): Steyerberg EW. Clinical Prediction Models: A Practical Approach to Development, Validation, and Updating: SBH/Statistics for Biology and Health. New York: Springer; Rizzo R, Maria L. Statistical computing. London: Chapman & Hall/CRC Press; Everitt BS, Hothorn T. A Handbook of Statistical Analyses Using R, 2nd ed. London: Chapman & Hall/CRC Press; Chaps. 5 7, Witten IH, Eibe F. Data Mining: Practical Machine Learning Tools and Techniques, 2nd ed. San Francisco: Frank Morgan Kaufmann Publishers; Chaps. 4, 6 and Maplesoft, in Statistics with Maple. Maple-edition 12. Waterloo, Ontario, Canada; Carr BI, Buch SC, Kondragunta V, Pancoska P, Branch RA. Tumor and liver determinants of prognosis in unresectable hepatocellular carcinoma: a case cohort study. J Gastroenterol Hepatol. 2008;23(8 Pt 1):