G Model YDLD-2849; No. of Pages 7
ARTICLE IN PRESS Digestive and Liver Disease xxx (2015) xxx–xxx
Contents lists available at ScienceDirect
Digestive and Liver Disease journal homepage: www.elsevier.com/locate/dld
Liver, Pancreas and Biliary Tract
Evaluation of the red cell distribution width as a biomarker of early mortality in hepatocellular carcinoma Carlo Smirne a , Glenda Grossi a , David J. Pinato b , Michela E. Burlone a,f , Francesco A. Mauri c , Adam Januszewski d , Alberto Oldani e , Rosalba Minisini a , Rohini Sharma b , Mario Pirisi a,∗ a
Department of Translational Medicine, Università del Piemonte Orientale, Novara, Italy Division of Experimental Medicine, Hammersmith Campus of Imperial College London, London, UK c Department of Histopathology, Hammersmith Campus of Imperial College London, London, UK d Department of Oncology, Hammersmith Campus of Imperial College London, London, UK e Department of Health Sciences, Università del Piemonte Orientale, Novara, Italy f CRRF Monsignor Luigi Novarese, Moncrivello, Italy b
a r t i c l e
i n f o
Article history: Received 3 November 2014 Accepted 11 March 2015 Available online xxx Keywords: Inflammation Liver cancer Prognosis Red cell distribution width
a b s t r a c t Background: The red cell distribution width is a biomarker of early mortality across various disease states. Aim: To verify whether it may refine estimates of survival in hepatocellular carcinoma. Methods: The red cell distribution width measured at diagnosis was analyzed in relationship to mortality by any cause both in a retrospective training cohort (N = 208), and in an independent prospectively collected validation cohort (N = 106) of patients with hepatocellular carcinoma. Based on Cox proportional hazards modelling, a prognostic index was validated. Results: In the training and the validation cohort, median survival time was respectively 1026 and 868 days in patients with red cell distribution width ≤14.6%, vs. 282 and 340 days in patients with red cell distribution width >14.6%; the corresponding hazard ratios were 0.43 (95% CI: 0.31–0.60), p < 0.0001 and 0.28 (95% CI: 0.17–0.47), p < 0.0001. At multivariate analysis, the red cell distribution width remained an independent predictor of survival (p < 0.001) in a Cox model including other widely accepted prognostic factors. Applying to the validation dataset the prognostic index derived from the training dataset, the ability of the model to discriminate the survival probabilities of patients was confirmed (Harrell’s C = 0.769). Conclusions: The red cell distribution width is a novel, reproducible, prospectively validated predictor of survival in patients with hepatocellular carcinoma. © 2015 Editrice Gastroenterologica Italiana S.r.l. Published by Elsevier Ltd. All rights reserved.
1. Introduction Hepatocellular carcinoma (HCC), the most frequent primary liver malignancy, ranks fifth in men among the most commonly diagnosed types of cancer, being however the second most lethal solid tumour. In females, HCC ranks seventh by incidence and sixth by mortality, respectively [1]. The usual precursors of HCC are fibrosis and cirrhosis [2], both the result of chronic inflammation in the liver parenchyma [3]. In industrialised countries, therefore, the prognosis of HCC is largely affected by the degree
∗ Corresponding author at: Dipartimento di Medicina Traslazionale, Via Solaroli 17, 28100 Novara, Italy. Tel.: +39 03213733847; fax: +39 03213733600. E-mail address: [email protected] (M. Pirisi).
of liver dysfunction secondary to cirrhosis. Equally, substantial evidence now suggests that cancer progression is dependent on the complex interaction between the tumour and the host inflammatory response [4], and HCC makes no exception to this rule [5–8]. The red cell distribution width (RDW) is one of the parameters in the full blood count generated automatically by cell counters to quantitate anisocytosis (i.e. the variability of the size of the circulating erythrocytes). Recently, RDW has emerged as a consistent and strong predictor of overall and disease-specific mortality in middle-age and older adults [9], probably due to the ability of RDW to reflect the systemic release of cytokines, such as IL-6, TNF-alpha and hepcidin across a wide range of disease states [10,11]. A second major factor leading to significant variation in erythrocyte size is oxidative stress [12].
http://dx.doi.org/10.1016/j.dld.2015.03.011 1590-8658/© 2015 Editrice Gastroenterologica Italiana S.r.l. Published by Elsevier Ltd. All rights reserved.
Please cite this article in press as: Smirne C, et al. Evaluation of the red cell distribution width as a biomarker of early mortality in hepatocellular carcinoma. Dig Liver Dis (2015), http://dx.doi.org/10.1016/j.dld.2015.03.011
G Model YDLD-2849; No. of Pages 7
ARTICLE IN PRESS C. Smirne et al. / Digestive and Liver Disease xxx (2015) xxx–xxx
2
In the Barcelona Clinic Liver Cancer (BCLC) staging system of HCC, that has been adopted by major scientific societies clinical practice guidelines to stratify liver cancer patients [13,14], the prognostic value of systemic inflammation is reflected by constitutional symptoms, the evaluation of which is unfortunately largely subjective and more apparent in advanced stage disease. Our aim was therefore to evaluate whether adding RDW to factors of demonstrated prognostic significance in HCC may objectively, reproducibly and inexpensively improve our ability to estimate survival of these patients at the time of diagnosis. 2. Materials and methods 2.1. Patients By scanning an electronic database including all patients aged ≥18 years who were consecutively diagnosed with HCC at an academic centre in Northern Italy (Novara) from November 15, 2003 to September 19, 2013 we retrieved data from 238 patient records; 198 patients satisfied either (a) the radiological or combined criteria for the diagnosis of HCC as indicated in the document summarising the conclusions of the Barcelona-2000 EASL conference [15] or (b) the criteria later recommended by the European Association for the Study of the Liver (EASL)/European Organisation for Research and Treatment of Cancer clinical practice guideline [14]. For all these patients, biopsy confirmation was not required. Pathological conformation of diagnosis, based on the definitions of the International Consensus Group for Hepatocellular Neoplasia [16], was obtained in the remaining 40 patients. Thirty patients whose RDW value was not available at tumour diagnosis were excluded. Therefore, the training cohort consisted of 208 patients. The primary treatment modalities chosen for these patients were curative for 81 patients (39%; N = 3 liver transplantation, N = 20 surgical resection and N = 58 radiofrequency ablation), non-curative for 69 patients (33%; N = 34 transarterial chemoembolization, N = 4 transarterial radioembolization, and N = 31 sorafenib), best supportive care for 56 patients (27%), and unknown in 2 patients (1%). Further 11 patients (5%) received a liver transplant after having been treated initially by a different treatment modality. The status of all these patients was known and documented up to (and including) the censor date of February 22, 2015. For external validation, survival analyses were then performed in an independent cohort of 106 consecutive patients aged ≥18 years with similar clinico-pathological characteristics prospectively recruited with the same inclusion criteria by another academic centre (Hammersmith Hospital, London, UK) from December 10th 2010 to January 8th 2014. None of these patients received a liver transplant. In this latter population, the censor date was February 15, 2015. Written informed consent was obtained by all participants to the study, which was conducted in strict adherence to the principles of the Declaration of Helsinki.
within one month from diagnostic imaging in ISO-certified central Laboratories. Cirrhosis was either diagnosed histologically or – when a liver biopsy was not available – according to clinical, laboratory, radiological and/or endoscopic criteria. The Child–Pugh–Turcotte score was recorded to describe liver function. Tumour staging at diagnosis followed the BCLC criteria; patients were thus stratified into four classes: A, very early or early stage; B, intermediate stage; C, advanced stage; D, end-stage [17].
2.3. Statistical analysis Statistical analysis was performed using MedCalc version 14.12.0 (MedCalc Software, Mariakerke, Belgium) and Stata version 13.1 (StataCorp LP, College Station, Texas, United States of America). For continuous variables, the variability of data around the central value was presented as medians (interquartile range), whilst categorical variables were presented as frequencies (percentage of the total). A univariate screen to identify potentially significant predictors of overall survival in HCC patients was conducted in the training cohort. Differences in survival probabilities among groups were analysed by the log-rank test or (when appropriate) the log-rank test for trend; hazard ratios (HR) and 95% confidence intervals were also calculated. The selection of the optimal cut-off value(s) to identify groups was based on biological plausibility [18] and was preceded by inspection of the survival curves obtained by generating a group variable with four groups based on quartiles. In the case of RDW, the cut-off value chosen (>14.6%) coincided with the upper limit of normal reference range [19] and the median value observed in the training cohort. To analyse the effects of multiple covariates on survival, the Cox proportional hazard regression model was used. To avoid to incur in the multicollinearity phenomenon, as expected being the Child–Pugh–Turcotte class part of the criteria on which the BCLC staging system is based, the Child–Pugh–Turcotte score, not the Child–Pugh–Turcotte class, was included among the predictor variables. Eight variables (3 categorical: gender, BCLC stage, and Child–Pugh–Turcotte score; 5 continuous: age at diagnosis, major tumour diameter, alpha-fetoprotein, haemoglobin, and RDW) were entered in the model backwards. For the variables retained in the model, regression coefficients, hazard ratios with their 95% CI and respective p values are presented. The baseline cumulative hazard H0 (t) was used to calculate the survival probability S(t) for any case at time t, by applying the regression coefficient b obtained in the training cohort to individual values the corresponding variable in the validation cohort: S(t) = exp(−H0 (t) × PI)
(1)
where PI is a prognostic index: PI = x1 b1 + x2 b2 + x3 b3 + · · · + xk bk
(2)
2.2. Collection of data In both studied cohorts, overall survival was calculated from the time of radiological diagnosis to the time of death or last follow-up, with the date of HCC diagnosis referring to the date of the diagnostic imaging procedure, even when pathological confirmation was obtained at a later date. Radiological tumour characteristics in both patient cohorts were derived from the diagnostic computed tomography (CT) scan or magnetic resonance imaging (MRI), evaluated by an experienced radiologist. All the clinico-pathological variables derived from the full blood count recorded in this study, including RDW, were performed
The prognostic index was applied without any adjustment of the coefficients to the derivation dataset. Since no missing data were allowed, 96/106 patients in the validation cohort were included in the analysis. To evaluate the discriminatory power and the predictive accuracy of the models, we calculated the Harrell’s C coefficients for (a) the model on the derivation dataset in which the independent variable was the prognostic index and (b) the models on the training plus the derivation datasets in which the independent variables were either the BCLC stage alone, or the BCLC stage plus RDW. The level of statistical significance chosen was 14.6%. In the validation cohort, the median RDW value was 15.3% (interquartile range, 13.9–16.7); 42/106 patients (39.6%) had a RDW ≤14.6% (group C), and 64/106 (60.4%) a RDW >14.6% (group D). A RDW >14.6% predicted patients’ survival both in the training and the validation cohort (Fig. 1, panels B and C). The median survival in the training cohort was 1026 days in the group with RDW ≤14.6% (95% CI: 740–1337) vs. 282 days (95% CI: 212–465) in the group with RDW >14.6%, HR = 0.43 (95% CI: 0.31–0.60), p < 0.0001. Moreover, median survival declined progressively at each RDW quartile, being 1376 days (95% CI: 921–1975; reference) for patients belonging to the first quartile, 671 days (95% CI: 504–1127; HR = 1.73, 95% CI: 1.15–2.58) for patients belonging to the second quartile, 381
Fig. 1. Kaplan–Meier estimates of survival probabilities are presented in panel A for the entire training (continuous line) and validation cohorts (dashed line); in panel B, for the training cohort, dichotomized according to individual values of red cell distribution width coefficient of variation ≤14.6% (continuous line) or >14.6% (dashed line) measured at diagnosis; and in panel C, for the validation cohort, dichotomized according to individual values of red cell distribution width coefficient of variation) ≤14.6% (continuous line) or >14.6% (dashed line) measured at diagnosis.
Please cite this article in press as: Smirne C, et al. Evaluation of the red cell distribution width as a biomarker of early mortality in hepatocellular carcinoma. Dig Liver Dis (2015), http://dx.doi.org/10.1016/j.dld.2015.03.011
G Model YDLD-2849; No. of Pages 7
ARTICLE IN PRESS C. Smirne et al. / Digestive and Liver Disease xxx (2015) xxx–xxx
4
Table 2 Univariate analysis of factors associated with survival in the training cohort (N = 208). p values refer to log rank test. Variable Age, years ≤70 (N = 110, 53%) >70 (N = 98, 47%) Child–Pugh–Turcotte class A (N = 147, 71%) B (N = 48, 23%) C (N = 13, 6%) Barcelona clinic liver cancer stage A (N = 95, 46%) B (N = 55, 26%) C (N = 45, 22%) D (N = 13, 6%) Tumour size, cm ≤4 (N = 125, 60%) >4 (N = 83, 40%) Alpha-fetoprotein, g/L ≤20 (N = 111, 53%) 20–100 (N = 42, 20%) >100 (N = 55, 26%) Haemoglobin, g/L ≥120 (N = 152, 73%) 120 g/L), Child–Pugh–Turcotte class, BCLC stage, initial treatment modality, maximum diameter of the major nodule (>4 cm) and serum AFP concentration (100 g/L) were all significantly associated with survival. The univariate analysis of the same variables in the validation cohort is presented in Table 3. Table 4 presents a summary of the Cox proportional hazard analysis in the subset of patients in the training cohort who did not undergo liver transplantation (N = 194). RDW was confirmed
to be an independent predictor of survival, together with age at diagnosis, BCLC stage, Child–Pugh–Turcotte score, tumour size and serum AFP. Haemoglobin and gender were not retained in the model, meaning that they did not contribute significantly to it. The same analysis was repeated in the validation cohort: RDW was confirmed to be an independent predictor of survival (HR 1.39, 95% CI 1.20–1.60; p < 0.0001), together with BCLC stage, Child–Pugh–Turcotte class and tumour size. Age at diagnosis, AFP, haemoglobin and gender were not retained in this second model. Then, for each subject in the validation cohort we calculated a prognostic index by applying the regression coefficients shown in Table 4 to each corresponding variable, according to equation [2] (see Section 2). Fig. 2, panel A presents the Kaplan–Meier estimates for the validation cohort, with patients grouped according to prognostic index tertiles. Compared to the first tertile (reference), the HR was 3.18 (95% CI: 1.74–5.81) in the second and 7.27 (95% CI: 3.50–15.1) in the third tertile (p < 0.0001 at the log-rank test for trend). Fig. 2, panel B presents a comparison between the observed (Kaplan–Meier) and predicted survival by each tertile of the prognostic index. The Harrell’s C coefficient of the Cox proportional regression analysis conducted on the validation cohort and including the prognostic index as the only independent variable was 0.769. Finally, after merging of the two datasets (i.e., those from the training and validation cohort) used to develop the Cox model, the Harrell’s C coefficient increased from 0.709 using the BCLC stage as the only independent variable to 0.754 using both the BCLC stage and RDW. At a cut-off value of 14.6%, the ability to stratify patients with different survival probabilities at the log-rank test remained through BCLC stages A (p = 0.021), B (p = 0.005) and C (p = 0.029) but not D (p = 0.637).
4. Discussion Our study shows that RDW at the time of diagnosis is an independent and reproducible predictor of patient survival in HCC. These results need to be interpreted in light of our current knowledge on the RDW as a prognostic factor and the relationship between systemic inflammation and the clinical progression of liver cancer. In clinical practice, RDW has been proposed to differentiate iron deficiency anaemia from a thalassemic trait [18,19], although it has been objected that, to this aim, RDW is not sufficiently specific to
Please cite this article in press as: Smirne C, et al. Evaluation of the red cell distribution width as a biomarker of early mortality in hepatocellular carcinoma. Dig Liver Dis (2015), http://dx.doi.org/10.1016/j.dld.2015.03.011
G Model YDLD-2849; No. of Pages 7
ARTICLE IN PRESS C. Smirne et al. / Digestive and Liver Disease xxx (2015) xxx–xxx
5
Table 3 Univariate analysis of factors associated with survival in the validation cohort (N = 106). p values refer to log rank test. Variable Age, years ≤70 (N = 66, 62%) >70 (N = 40, 38%) Child–Pugh–Turcotte classa A (N = 56, 54%) B (N = 40, 39%) C (N = 7, 7%) Barcelona clinic liver cancer stageb A (N = 25, 25%) B (N = 31, 31%) C (N = 34, 34%) D (N = 10, 10%) Tumour size, cmc ≤4 (N = 45, 44%) >4 (N = 57, 56%) Alpha-fetoprotein, g/Ld ≤20 (N = 45, 45%) 20–100 (N = 16, 16%) >100 (N = 40, 40%) Haemoglobin, g/L ≥120 (N = 61, 58%)
ARTICLE IN PRESS Digestive and Liver Disease xxx (2015) xxx–xxx
Contents lists available at ScienceDirect
Digestive and Liver Disease journal homepage: www.elsevier.com/locate/dld
Liver, Pancreas and Biliary Tract
Evaluation of the red cell distribution width as a biomarker of early mortality in hepatocellular carcinoma Carlo Smirne a , Glenda Grossi a , David J. Pinato b , Michela E. Burlone a,f , Francesco A. Mauri c , Adam Januszewski d , Alberto Oldani e , Rosalba Minisini a , Rohini Sharma b , Mario Pirisi a,∗ a
Department of Translational Medicine, Università del Piemonte Orientale, Novara, Italy Division of Experimental Medicine, Hammersmith Campus of Imperial College London, London, UK c Department of Histopathology, Hammersmith Campus of Imperial College London, London, UK d Department of Oncology, Hammersmith Campus of Imperial College London, London, UK e Department of Health Sciences, Università del Piemonte Orientale, Novara, Italy f CRRF Monsignor Luigi Novarese, Moncrivello, Italy b
a r t i c l e
i n f o
Article history: Received 3 November 2014 Accepted 11 March 2015 Available online xxx Keywords: Inflammation Liver cancer Prognosis Red cell distribution width
a b s t r a c t Background: The red cell distribution width is a biomarker of early mortality across various disease states. Aim: To verify whether it may refine estimates of survival in hepatocellular carcinoma. Methods: The red cell distribution width measured at diagnosis was analyzed in relationship to mortality by any cause both in a retrospective training cohort (N = 208), and in an independent prospectively collected validation cohort (N = 106) of patients with hepatocellular carcinoma. Based on Cox proportional hazards modelling, a prognostic index was validated. Results: In the training and the validation cohort, median survival time was respectively 1026 and 868 days in patients with red cell distribution width ≤14.6%, vs. 282 and 340 days in patients with red cell distribution width >14.6%; the corresponding hazard ratios were 0.43 (95% CI: 0.31–0.60), p < 0.0001 and 0.28 (95% CI: 0.17–0.47), p < 0.0001. At multivariate analysis, the red cell distribution width remained an independent predictor of survival (p < 0.001) in a Cox model including other widely accepted prognostic factors. Applying to the validation dataset the prognostic index derived from the training dataset, the ability of the model to discriminate the survival probabilities of patients was confirmed (Harrell’s C = 0.769). Conclusions: The red cell distribution width is a novel, reproducible, prospectively validated predictor of survival in patients with hepatocellular carcinoma. © 2015 Editrice Gastroenterologica Italiana S.r.l. Published by Elsevier Ltd. All rights reserved.
1. Introduction Hepatocellular carcinoma (HCC), the most frequent primary liver malignancy, ranks fifth in men among the most commonly diagnosed types of cancer, being however the second most lethal solid tumour. In females, HCC ranks seventh by incidence and sixth by mortality, respectively [1]. The usual precursors of HCC are fibrosis and cirrhosis [2], both the result of chronic inflammation in the liver parenchyma [3]. In industrialised countries, therefore, the prognosis of HCC is largely affected by the degree
∗ Corresponding author at: Dipartimento di Medicina Traslazionale, Via Solaroli 17, 28100 Novara, Italy. Tel.: +39 03213733847; fax: +39 03213733600. E-mail address: [email protected] (M. Pirisi).
of liver dysfunction secondary to cirrhosis. Equally, substantial evidence now suggests that cancer progression is dependent on the complex interaction between the tumour and the host inflammatory response [4], and HCC makes no exception to this rule [5–8]. The red cell distribution width (RDW) is one of the parameters in the full blood count generated automatically by cell counters to quantitate anisocytosis (i.e. the variability of the size of the circulating erythrocytes). Recently, RDW has emerged as a consistent and strong predictor of overall and disease-specific mortality in middle-age and older adults [9], probably due to the ability of RDW to reflect the systemic release of cytokines, such as IL-6, TNF-alpha and hepcidin across a wide range of disease states [10,11]. A second major factor leading to significant variation in erythrocyte size is oxidative stress [12].
http://dx.doi.org/10.1016/j.dld.2015.03.011 1590-8658/© 2015 Editrice Gastroenterologica Italiana S.r.l. Published by Elsevier Ltd. All rights reserved.
Please cite this article in press as: Smirne C, et al. Evaluation of the red cell distribution width as a biomarker of early mortality in hepatocellular carcinoma. Dig Liver Dis (2015), http://dx.doi.org/10.1016/j.dld.2015.03.011
G Model YDLD-2849; No. of Pages 7
ARTICLE IN PRESS C. Smirne et al. / Digestive and Liver Disease xxx (2015) xxx–xxx
2
In the Barcelona Clinic Liver Cancer (BCLC) staging system of HCC, that has been adopted by major scientific societies clinical practice guidelines to stratify liver cancer patients [13,14], the prognostic value of systemic inflammation is reflected by constitutional symptoms, the evaluation of which is unfortunately largely subjective and more apparent in advanced stage disease. Our aim was therefore to evaluate whether adding RDW to factors of demonstrated prognostic significance in HCC may objectively, reproducibly and inexpensively improve our ability to estimate survival of these patients at the time of diagnosis. 2. Materials and methods 2.1. Patients By scanning an electronic database including all patients aged ≥18 years who were consecutively diagnosed with HCC at an academic centre in Northern Italy (Novara) from November 15, 2003 to September 19, 2013 we retrieved data from 238 patient records; 198 patients satisfied either (a) the radiological or combined criteria for the diagnosis of HCC as indicated in the document summarising the conclusions of the Barcelona-2000 EASL conference [15] or (b) the criteria later recommended by the European Association for the Study of the Liver (EASL)/European Organisation for Research and Treatment of Cancer clinical practice guideline [14]. For all these patients, biopsy confirmation was not required. Pathological conformation of diagnosis, based on the definitions of the International Consensus Group for Hepatocellular Neoplasia [16], was obtained in the remaining 40 patients. Thirty patients whose RDW value was not available at tumour diagnosis were excluded. Therefore, the training cohort consisted of 208 patients. The primary treatment modalities chosen for these patients were curative for 81 patients (39%; N = 3 liver transplantation, N = 20 surgical resection and N = 58 radiofrequency ablation), non-curative for 69 patients (33%; N = 34 transarterial chemoembolization, N = 4 transarterial radioembolization, and N = 31 sorafenib), best supportive care for 56 patients (27%), and unknown in 2 patients (1%). Further 11 patients (5%) received a liver transplant after having been treated initially by a different treatment modality. The status of all these patients was known and documented up to (and including) the censor date of February 22, 2015. For external validation, survival analyses were then performed in an independent cohort of 106 consecutive patients aged ≥18 years with similar clinico-pathological characteristics prospectively recruited with the same inclusion criteria by another academic centre (Hammersmith Hospital, London, UK) from December 10th 2010 to January 8th 2014. None of these patients received a liver transplant. In this latter population, the censor date was February 15, 2015. Written informed consent was obtained by all participants to the study, which was conducted in strict adherence to the principles of the Declaration of Helsinki.
within one month from diagnostic imaging in ISO-certified central Laboratories. Cirrhosis was either diagnosed histologically or – when a liver biopsy was not available – according to clinical, laboratory, radiological and/or endoscopic criteria. The Child–Pugh–Turcotte score was recorded to describe liver function. Tumour staging at diagnosis followed the BCLC criteria; patients were thus stratified into four classes: A, very early or early stage; B, intermediate stage; C, advanced stage; D, end-stage [17].
2.3. Statistical analysis Statistical analysis was performed using MedCalc version 14.12.0 (MedCalc Software, Mariakerke, Belgium) and Stata version 13.1 (StataCorp LP, College Station, Texas, United States of America). For continuous variables, the variability of data around the central value was presented as medians (interquartile range), whilst categorical variables were presented as frequencies (percentage of the total). A univariate screen to identify potentially significant predictors of overall survival in HCC patients was conducted in the training cohort. Differences in survival probabilities among groups were analysed by the log-rank test or (when appropriate) the log-rank test for trend; hazard ratios (HR) and 95% confidence intervals were also calculated. The selection of the optimal cut-off value(s) to identify groups was based on biological plausibility [18] and was preceded by inspection of the survival curves obtained by generating a group variable with four groups based on quartiles. In the case of RDW, the cut-off value chosen (>14.6%) coincided with the upper limit of normal reference range [19] and the median value observed in the training cohort. To analyse the effects of multiple covariates on survival, the Cox proportional hazard regression model was used. To avoid to incur in the multicollinearity phenomenon, as expected being the Child–Pugh–Turcotte class part of the criteria on which the BCLC staging system is based, the Child–Pugh–Turcotte score, not the Child–Pugh–Turcotte class, was included among the predictor variables. Eight variables (3 categorical: gender, BCLC stage, and Child–Pugh–Turcotte score; 5 continuous: age at diagnosis, major tumour diameter, alpha-fetoprotein, haemoglobin, and RDW) were entered in the model backwards. For the variables retained in the model, regression coefficients, hazard ratios with their 95% CI and respective p values are presented. The baseline cumulative hazard H0 (t) was used to calculate the survival probability S(t) for any case at time t, by applying the regression coefficient b obtained in the training cohort to individual values the corresponding variable in the validation cohort: S(t) = exp(−H0 (t) × PI)
(1)
where PI is a prognostic index: PI = x1 b1 + x2 b2 + x3 b3 + · · · + xk bk
(2)
2.2. Collection of data In both studied cohorts, overall survival was calculated from the time of radiological diagnosis to the time of death or last follow-up, with the date of HCC diagnosis referring to the date of the diagnostic imaging procedure, even when pathological confirmation was obtained at a later date. Radiological tumour characteristics in both patient cohorts were derived from the diagnostic computed tomography (CT) scan or magnetic resonance imaging (MRI), evaluated by an experienced radiologist. All the clinico-pathological variables derived from the full blood count recorded in this study, including RDW, were performed
The prognostic index was applied without any adjustment of the coefficients to the derivation dataset. Since no missing data were allowed, 96/106 patients in the validation cohort were included in the analysis. To evaluate the discriminatory power and the predictive accuracy of the models, we calculated the Harrell’s C coefficients for (a) the model on the derivation dataset in which the independent variable was the prognostic index and (b) the models on the training plus the derivation datasets in which the independent variables were either the BCLC stage alone, or the BCLC stage plus RDW. The level of statistical significance chosen was 14.6%. In the validation cohort, the median RDW value was 15.3% (interquartile range, 13.9–16.7); 42/106 patients (39.6%) had a RDW ≤14.6% (group C), and 64/106 (60.4%) a RDW >14.6% (group D). A RDW >14.6% predicted patients’ survival both in the training and the validation cohort (Fig. 1, panels B and C). The median survival in the training cohort was 1026 days in the group with RDW ≤14.6% (95% CI: 740–1337) vs. 282 days (95% CI: 212–465) in the group with RDW >14.6%, HR = 0.43 (95% CI: 0.31–0.60), p < 0.0001. Moreover, median survival declined progressively at each RDW quartile, being 1376 days (95% CI: 921–1975; reference) for patients belonging to the first quartile, 671 days (95% CI: 504–1127; HR = 1.73, 95% CI: 1.15–2.58) for patients belonging to the second quartile, 381
Fig. 1. Kaplan–Meier estimates of survival probabilities are presented in panel A for the entire training (continuous line) and validation cohorts (dashed line); in panel B, for the training cohort, dichotomized according to individual values of red cell distribution width coefficient of variation ≤14.6% (continuous line) or >14.6% (dashed line) measured at diagnosis; and in panel C, for the validation cohort, dichotomized according to individual values of red cell distribution width coefficient of variation) ≤14.6% (continuous line) or >14.6% (dashed line) measured at diagnosis.
Please cite this article in press as: Smirne C, et al. Evaluation of the red cell distribution width as a biomarker of early mortality in hepatocellular carcinoma. Dig Liver Dis (2015), http://dx.doi.org/10.1016/j.dld.2015.03.011
G Model YDLD-2849; No. of Pages 7
ARTICLE IN PRESS C. Smirne et al. / Digestive and Liver Disease xxx (2015) xxx–xxx
4
Table 2 Univariate analysis of factors associated with survival in the training cohort (N = 208). p values refer to log rank test. Variable Age, years ≤70 (N = 110, 53%) >70 (N = 98, 47%) Child–Pugh–Turcotte class A (N = 147, 71%) B (N = 48, 23%) C (N = 13, 6%) Barcelona clinic liver cancer stage A (N = 95, 46%) B (N = 55, 26%) C (N = 45, 22%) D (N = 13, 6%) Tumour size, cm ≤4 (N = 125, 60%) >4 (N = 83, 40%) Alpha-fetoprotein, g/L ≤20 (N = 111, 53%) 20–100 (N = 42, 20%) >100 (N = 55, 26%) Haemoglobin, g/L ≥120 (N = 152, 73%) 120 g/L), Child–Pugh–Turcotte class, BCLC stage, initial treatment modality, maximum diameter of the major nodule (>4 cm) and serum AFP concentration (100 g/L) were all significantly associated with survival. The univariate analysis of the same variables in the validation cohort is presented in Table 3. Table 4 presents a summary of the Cox proportional hazard analysis in the subset of patients in the training cohort who did not undergo liver transplantation (N = 194). RDW was confirmed
to be an independent predictor of survival, together with age at diagnosis, BCLC stage, Child–Pugh–Turcotte score, tumour size and serum AFP. Haemoglobin and gender were not retained in the model, meaning that they did not contribute significantly to it. The same analysis was repeated in the validation cohort: RDW was confirmed to be an independent predictor of survival (HR 1.39, 95% CI 1.20–1.60; p < 0.0001), together with BCLC stage, Child–Pugh–Turcotte class and tumour size. Age at diagnosis, AFP, haemoglobin and gender were not retained in this second model. Then, for each subject in the validation cohort we calculated a prognostic index by applying the regression coefficients shown in Table 4 to each corresponding variable, according to equation [2] (see Section 2). Fig. 2, panel A presents the Kaplan–Meier estimates for the validation cohort, with patients grouped according to prognostic index tertiles. Compared to the first tertile (reference), the HR was 3.18 (95% CI: 1.74–5.81) in the second and 7.27 (95% CI: 3.50–15.1) in the third tertile (p < 0.0001 at the log-rank test for trend). Fig. 2, panel B presents a comparison between the observed (Kaplan–Meier) and predicted survival by each tertile of the prognostic index. The Harrell’s C coefficient of the Cox proportional regression analysis conducted on the validation cohort and including the prognostic index as the only independent variable was 0.769. Finally, after merging of the two datasets (i.e., those from the training and validation cohort) used to develop the Cox model, the Harrell’s C coefficient increased from 0.709 using the BCLC stage as the only independent variable to 0.754 using both the BCLC stage and RDW. At a cut-off value of 14.6%, the ability to stratify patients with different survival probabilities at the log-rank test remained through BCLC stages A (p = 0.021), B (p = 0.005) and C (p = 0.029) but not D (p = 0.637).
4. Discussion Our study shows that RDW at the time of diagnosis is an independent and reproducible predictor of patient survival in HCC. These results need to be interpreted in light of our current knowledge on the RDW as a prognostic factor and the relationship between systemic inflammation and the clinical progression of liver cancer. In clinical practice, RDW has been proposed to differentiate iron deficiency anaemia from a thalassemic trait [18,19], although it has been objected that, to this aim, RDW is not sufficiently specific to
Please cite this article in press as: Smirne C, et al. Evaluation of the red cell distribution width as a biomarker of early mortality in hepatocellular carcinoma. Dig Liver Dis (2015), http://dx.doi.org/10.1016/j.dld.2015.03.011
G Model YDLD-2849; No. of Pages 7
ARTICLE IN PRESS C. Smirne et al. / Digestive and Liver Disease xxx (2015) xxx–xxx
5
Table 3 Univariate analysis of factors associated with survival in the validation cohort (N = 106). p values refer to log rank test. Variable Age, years ≤70 (N = 66, 62%) >70 (N = 40, 38%) Child–Pugh–Turcotte classa A (N = 56, 54%) B (N = 40, 39%) C (N = 7, 7%) Barcelona clinic liver cancer stageb A (N = 25, 25%) B (N = 31, 31%) C (N = 34, 34%) D (N = 10, 10%) Tumour size, cmc ≤4 (N = 45, 44%) >4 (N = 57, 56%) Alpha-fetoprotein, g/Ld ≤20 (N = 45, 45%) 20–100 (N = 16, 16%) >100 (N = 40, 40%) Haemoglobin, g/L ≥120 (N = 61, 58%)
