CLINICAL STUDY
Nomograms for Predicting Outcomes after Chemoembolization in Patients with Nonmetastatic Hepatocellular Carcinoma Yeonjung Ha, MD, Seungbong Han, PhD, Ju Hyun Shim, MD, Gi-Young Ko, MD, Hyun-Ki Yoon, MD, Kyu-Bo Sung, MD, Danbi Lee, MD, Kang Mo Kim, MD, Young-Suk Lim, MD, Young-Hwa Chung, MD, Yung Sang Lee, MD, and Han Chu Lee, MD
ABSTRACT Purpose: To construct prognostic nomograms capable of estimating individual probabilities of tumor progression and overall survival (OS) at specific time points during serial transarterial chemoembolization for hepatocellular carcinoma (HCC). Materials and Methods: The study included 1,181 consecutive patients with nonmetastatic HCC undergoing repeated transarterial chemoembolization at a single tertiary referral center. Patients were assigned to 2 cohorts according to the first transarterial chemoembolization date: derivation (2004–2006; n ¼ 854) and validation (2007; n ¼ 327) sets. Multivariate Cox proportional hazards models were developed based on covariates derived before transarterial chemoembolization and assessed for their association with 5-year OS and 3-year progression-free survival (PFS). The accuracy of the models was internally and externally validated. Results: The 5-year OS of the derivation set was 25.4%, and 3-year PFS was 20.8%. Nomograms for OS and PFS were built into the derivation set incorporating the following factors: log [tumor volume] calculated as 4/3 3.14 (maximum radius of tumor in cm3); tumor number; tumor type (nodular or infiltrative); Child-Pugh class (A or B); (model for end-stage liver disease score/10)2; log [α-fetoprotein]; and portal vein invasion. The models had good discrimination and calibration abilities with C-indexes of 0.80 (5-y survival) and 0.77 (3-y progression). The results of external validation confirmed that these models performed well in terms of discrimination and goodness-of-fit (C-indexes 0.77 for 5-y survival and 0.73 for 3-y progression). Conclusion: Nomograms quantifying the survival and progression outcomes in patients treated with transarterial chemoembolization are useful clinical aids in providing personalized care.
ABBREVIATIONS AFP = α-fetoprotein, CI = confidence interval, HBV = hepatitis B virus, HCC = hepatocellular carcinoma, MELD = model for endstage liver disease, OS = overall survival, PFS = progression-free survival, TTV = total tumor volume From the Department of Gastroenterology (Y.H., J.H.S., D.L., K.M.K., Y.-S.L., Y.-H.C., Y.S.L., H.C.L.), Asan Liver Center, and Department of Radiology (G.-Y.K., H.-K.Y., K.-B.S.), Asan Medical Center, University of Ulsan College of Medicine, 88 Olympic-ro 43-gil, Songpa-gu, Seoul 138-736, Republic of Korea; and Department of Applied Statistics (S.H.), Gachon University, Seongnam-si, Gyeonggi-do, Republic of Korea. Received October 29, 2014; final revision received March 29, 2015; accepted April 12, 2015. Address correspondence to J.H.S., E-mail: [email protected]; or H.C.L., E-mail: [email protected] Y.H. and S.H. contributed equally to this article. None of the authors have identified a conflict of interest. Appendix A and Appendix B are available online at www.jvir.org. & SIR, 2015 J Vasc Interv Radiol 2015; 26:1093–1101 http://dx.doi.org/10.1016/j.jvir.2015.04.010
Hepatocellular carcinoma (HCC) is one of the most common lethal malignancies worldwide (1). Despite increased clinical surveillance, only 30% of cases of HCC are diagnosed in the early stages (2). In other words, 70% are still detected in the late stages, in which curative treatments such as surgical resection, percutaneous ablation, or liver transplantation are not feasible (3). Some patients with early-stage disease are not eligible for curative management because of their general medical condition, the location of the tumor, or the shortage of liver donors. Palliative treatments are often required.
1094
’
Nomograms for Postchemoembolization Outcomes
Transarterial chemoembolization has been shown to be a life-prolonging treatment and has been used in practice for a wide spectrum of HCCs including unresectable, untransplantable, and unablatable tumors at an early stage (4–6). However, because of the heterogeneity of the patient populations that underwent transarterial chemoembolization, variable survival outcomes have been reported, with a 2-year survival rate ranging from 20% to 60% among patients with intermediate-stage HCC for whom transarterial chemoembolization is the current standard recommendation (3,7). In certain subsets of patients who had decompensated liver disease, advanced liver dysfunction, macroscopic vascular invasion, or extrahepatic spread, this procedure was regarded as essentially detrimental (8). Many studies have been performed to identify prognostic factors using these heterogeneous populations, and variables such as liver function, tumor-node-metastasis stage, performance status, presence of hepatitis B virus (HBV), and α-fetoprotein (AFP) have been identified (9–14). However, the factors identified varied considerably from study to study, so that comprehensive predictions of patient outcomes have been difficult to make. To obtain per-patient prognoses and help physicians decide on treatment options by identifying more homogeneous subsets of patients, we constructed and validated comprehensive risk-scoring models in the form of nomograms consisting of easily accessible variables predicting overall survival (OS) and progression-free survival (PFS) in patients treated with transarterial chemoembolization.
MATERIALS AND METHODS
Ha et al
’
JVIR
the time of diagnosis, or (c) had other histologically confirmed malignancies before or within 5 years of the diagnosis of HCC. There were 1,181 patients enrolled, and these patients were divided into a derivation set and a validation set, according to the date of initial transarterial chemoembolization (15,16). The derivation set for the primary analysis consisted of 854 patients (72.3%) who underwent their first transarterial chemoembolization between January 2004 and December 2006, and the temporal validation set consisted of the remaining 327 patients (27.7%) (15,16).
Data Collection At the time of diagnosis, demographic and clinical variables including age, sex, tumor size, tumor number, tumor type (nodular or infiltrative), portal vein invasion, hepatic vein invasion, extrahepatic metastases, presence of cirrhosis, bilobar involvement, Child-Pugh class, model for end-stage liver disease (MELD) score, AFP level, and presence of HBV were collected for each patient from the database of our center. Total tumor volume (TTV), which had been initially proposed by Toso et al (17), was calculated from the sum of tumor volumes (4/3 3.14 [maximum radius of the tumor in cm3]). All of the enrolled cases were confirmed as HCC by liver protocol computed tomography or magnetic resonance imaging or liver biopsy according to current American Association for the Study of Liver Diseases guidelines (3). Before the initial transarterial chemoembolization, chest computed tomography and bone scans were checked on a routine basis, and, if indicated, positron emission tomography was performed to exclude extrahepatic metastases.
Study Design This was a retrospective cohort study undertaken to construct risk-scoring models for patients who receive transarterial chemoembolization secondary to unresectable HCC, in whom the derivation and validation cohorts from the gastroenterology department of a single tertiary referral center were temporarily separated.
Transarterial Chemoembolization Procedure The routine protocol of our hospital has been described elsewhere (18). Details of the procedure are provided in Appendix A (available online at www.jvir.org) (19–23).
Derivation and Validation Sets
Study Endpoints
Consecutive data were collected for patients with a new diagnosis of HCC who were treated with transarterial chemoembolization as initial therapy at a single center between January 2004 and December 2007. The diagnosis of HCC based on the American Association for the Study of Liver Diseases practice guidelines was reconfirmed in all included patients (3). Transarterial chemoembolization was the optimal treatment, and curative surgical or locoregional modalities were impossible or contraindicated at the time of enrollment of these patients. The following patients were excluded: patients who (a) were lost to follow-up after the first transarterial chemoembolization session without evaluation of the treatment response, (b) had extrahepatic metastases at
The primary endpoint was OS, which was defined as the interval between the date of the first transarterial chemoembolization and death or last follow-up visit. During follow-up, data were censored at the time when a treatment strategy of curative intent, such as surgical resection, liver transplantation, or percutaneous ablation, was attempted. Patients lost to follow-up were censored at the time of their last visit, and living patients were censored on December 31, 2012. The secondary endpoint was PFS estimated from the date of the first transarterial chemoembolization session to progression or the last follow-up visit. Progressive disease was confirmed according to the modified Response Evaluation Criteria in Solid Tumors (24). Death and follow-up
Volume 26
’
Number 8
’
August
’
2015
1095
data were fully accessible through the database of our center and collected until December 31, 2012.
Statistical Analysis To develop the prediction model for estimating 5-year OS and 3-year PFS, we used a Cox proportional hazards model as a base model. Based on a literature survey, the following variables were selected for candidate potential predictors: age, sex, calculated tumor volume, tumor number, tumor type (nodular or infiltrative), bilobar involvement, Child-Pugh class, MELD score, AFP value, presence of HBV, presence of cirrhosis, hepatic vein invasion, and portal vein invasion. A random forest model was employed for tentative variable selection (19). A detailed description of variable selection and the final model derivation is provided in Appendix A (available online at www.jvir.org).
was 99.7 ng/mL. At the time of diagnosis, 280 (23.7%) patients had portal vein invasion. Of these cases, 90 (32.2%) patients had HCCs invading main venous branches, and 190 (67.8%) patients had HCCs invading lobar or segmental venous branches (25). The two sets were similar in terms of age, sex, tumor size, tumor volume, tumor type, cirrhosis, Child-Pugh class, MELD score, AFP value, and HBV status. However, there were more patients with more than four tumors in the validation set (P ¼ .007). The proportions of patients with tumors in both lobes and with portal vein invasion also were higher in the validation set (P ¼ .015 and .027 for each). Over the study period, 73 (8.5%) patients in the derivation set and 22 (6.7%) patients in the validation set were lost to follow-up at least after the evaluation of their response to the first session and then censored at the date of their last follow-up examination.
Covariates Selected in the Nomograms
RESULTS Baseline Characteristics of Enrolled Patients The demographic and clinical parameters of the derivation and validation sets are provided in Table 1. The mean age of the 1,181 patients was 55.9 years (standard deviation [SD], 10.1 y). Of the patients, 981 (83.1%) were men, 953 (80.7%) had liver cirrhosis, 931 (78.8%) were in Child-Pugh class A, and 927 (78.5%) had nodular-type tumors. Median tumor volume was 39.0 cm3, mean MELD score was 9.0 (SD, 2.5), and median serum AFP
In the derivation set using the input variables of age, sex, calculated tumor volume, tumor number, tumor type (nodular or infiltrative), bilobar involvement, ChildPugh class, MELD score, AFP value, presence of HBV, presence of cirrhosis, hepatic vein invasion, and portal vein invasion, the seven most important variables for survival and progression were provided by the random forest model. These variables were tumor volume, tumor number, tumor type, Child-Pugh class, MELD score, AFP value, and portal vein invasion (Table 2). The final models are shown in Table 3.
Table 1 . Characteristics of the Derivation and Validation Sets All Patients (N ¼ 1,181)
Derivation Set (n ¼ 854)
Validation Set (n ¼ 327)
55.9 ⫾ 10.1
55.7 ⫾ 10.0
56.4 ⫾ 10.4
.36
981 (83.1)
717 (84.0)
264 (80.7)
.19
39.0 (7.0–321.0)
34.5 (7.0–312.8)
48.0 (8.0–382.0)
.40 .007
1
616 (52.2)
455 (53.3)
161 (49.2)
2 3
219 (18.5) 83 (7.0)
168 (19.7) 59 (6.9)
51 (15.6) 24 (7.3)
Characteristics Age (y), mean ⫾ SD Male, n (%) Tumor volume (cm3), median (IQR) Tumor number, n (%)
4
P Value
29 (2.5)
24 (2.8)
5 (1.5)
234 (19.8)
148 (17.3)
86 (26.3)
927 (78.5)
674 (78.9)
253 (77.4)
Infiltrative Bilobar involvement, n (%)
254 (21.5) 326 (27.6)
180 (21.1) 219 (25.6)
74 (22.6) 107 (32.7)
.015
Child-Pugh class A, n (%)
931 (78.8)
671 (78.6)
260 (79.5)
.72
MELD score, mean ⫾ SD Presence of cirrhosis, n (%)
9.0 ⫾ 2.5 953 (80.7)
9.1 ⫾ 2.7 684 (80.1)
8.8 ⫾ 2.0 269 (82.3)
.18 .40
Portal vein invasion, n (%)
280 (23.7)
188 (22.0)
92 (28.1)
.027
36 (3.0) 99.7 (14.3–1,250.0)
31 (3.6) 111.5 (13.3–1,100.0)
5 (1.5) 73.5 (13.7–1,695.0)
.06 .63
899 (76.1)
647 (75.8)
252 (77.1)
0.64
Z5 Tumor type, n (%) Nodular
Hepatic vein invasion, n (%) Serum AFP level (ng/mL), median (IQR) Presence of HBV
.56
AFP ¼ α-fetoprotein, HBV ¼ hepatitis B virus, IQR ¼ interquartile range, MELD ¼ model for end-stage liver disease, SD ¼ standard deviation.
1096
’
Nomograms for Postchemoembolization Outcomes
Ha et al
’
JVIR
Table 2 . Variables Selected by a Random Forest Model OS Variables
Importance
Tumor number
0.0194
Tumor volume Log [AFP]
0.0130 0.0076
PFS Relative Importance 1 0.6662 0.3903
Importance
Relative Importance
0.0095
0.4305
0.0221 0.0132
1 0.5993
MELD score, per point
0.0086
0.4420
0.0039
0.1757
Portal vein invasion Infiltrative tumor type
0.0086 0.0039
0.4432 0.1997
0.0018 0.0004
0.0828 0.0200
Age
0.0023
0.1205
0.0021
0.0940
0.0005 0.0002
0.0278 0.0110
0.0037 0.0004
0.1691 0.0178
Child-Pugh class B Male
0
0.0017
0.0001
0.0036
HBV Liver cirrhosis
0.0011 0.0012
0.0544 0.0606
0.0007 0.0001
0.0315 0.0320
Bilobar involvement
0.0018
0.0915
0.0037
0.1662
Hepatic vein invasion
AFP ¼ α-fetoprotein, HBV ¼ hepatitis B virus, MELD ¼ model for end-stage liver disease, OS ¼ overall survival, PFS ¼ progressionfree survival.
Table 3 . Final Models for OS and PFS OS
PFS
HR (95% CI)
P Value
HR (95% CI)
Portal vein invasion
1.61 (1.26–2.04)
o .001
1.16 (0.93–1.45)
.19
Tumor number Log [tumor volume]
1.28 (1.21–1.36) 1.19 (1.13–1.24)
o .001 o .001
1.21 (1.15–1.27) 1.16 (1.12–1.21)
o .001 o .001
(MELD score/10)2
0.64 (0.53–0.78)
o .001
0.81 (0.69–0.95)
.009
Infiltrative tumor type Log [AFP]
1.26 (0.99–1.61) 1.08 (1.04–1.11)
.06 o .001
1.20 (0.96–1.51) 1.06 (1.03–1.09)
.11 o .001
Child-Pugh class B
1.42 (1.11–1.83)
.006
1.43 (1.15–1.78)
.001
Variables
P Value
AFP ¼ α-fetoprotein, CI ¼ confidence interval, HR ¼ hazard ratio, MELD ¼ model for end-stage liver disease, OS ¼ overall survival, PFS ¼ progression-free survival.
Internal and External Validation of Constructed Nomograms In the derivation set, the 5-year OS was 25.4% with a median survival time of 28.5 months (95% confidence interval [CI], 25.9–31.0 mo), and the 3-year PFS was 20.8% with a median of 14.3 months (95% CI, 13.0–15.5 mo). Using the analysis described earlier, the nomograms for OS (Fig 1a) and PFS (Fig 1b) were developed into a derivation set by parameterization of the variables and summing the points corresponding to each variable: log [tumor volume] þ tumor number (5 for Z 5) þ tumor type (nodular: 0, infiltrative: 1) þ Child-Pugh class (A: 0, B: 1) þ (MELD score/10)2 þ log [AFP] þ portal vein invasion (no: 0, yes: 1). We assessed the individual contributions of the seven variables selected by the random forest model to discrimination and calibration. The final model had C-indexes for OS of 0.84, 0.82, and 0.80 at 1 year, 3 years, and 5 years; in the PFS analysis, the C-indexes were 0.76, 0.77, and 0.78 at 1 year, 3 years, and 5 years. Calibration curves for the
predicted and actual probabilities of OS (Fig 2a) and PFS (Fig 2b) yielded good calibration abilities based on the results of the Hosmer-Lemeshow test (P ¼ .87 for OS and P ¼ .30 for PFS). The scoring system built for the derivation set with median OS and PFS times of 28.5 months (95% CI, 25.9–31.0 mo) and 14.3 months (95% CI, 13.0–15.5 mo), respectively, was externally validated in the temporal validation cohort. The external validation analysis confirmed that these models performed well in terms of discrimination (C-indexes of 0.78, 0.80, and 0.77 for 1-, 3-, and 5-year OS and 0.72 for 1-year PFS and 0.73 for 3- and 5-year PFS) and of calibration (P values of the Hosmer-Lemeshow test, P ¼ .13 for OS and P ¼ .70 for PFS) (Fig 3a, b).
Examples of the Use of Nomograms Prognostic nomograms were developed for predicting OS and PFS rates at 1, 3, and 5 years (Fig 1a, b). Points were assigned for tumor log [tumor volume], tumor
Volume 26
’
Number 8
’
August
’
2015
1097
Figure 1. Nomogram depicting OS (a) and PFS (b). By drawing a line between each variable and the uppermost component “points,” the appropriate points can be assigned to seven variables. The sum of these seven points can be expressed on the “total points” line. Then t-year survival or progression can be calculated by connecting each point to the t-survival line (t ¼ 1, 3, and 5).
number, tumor type, portal vein invasion, (MELD score/10)2, Child-Pugh class, and log [AFP] by drawing lines upward from the corresponding values to the
“points” line. The sum of these seven points, plotted on the “total points” line, corresponded to the t-year OS or PFS probability (t ¼ 1, 3 or 5). Using these seven
1098
’
Nomograms for Postchemoembolization Outcomes
Ha et al
’
JVIR
Figure 2. Calibration plot for OS (a) and PFS (b) in the derivation set. Calibration curves showing average predicted probability (x-axis) against the Kaplan-Meier probability observed in the derivation set (y-axis). The vertical line represents 95% CIs of the Kaplan-Meier estimates. The dashed line represents the reference line. P values calculated by the Hosmer-Lemeshow goodness-of-fit test indicate that the nomograms fit well (P ¼ .87 for OS and P ¼ .30 for PFS).
Figure 3. Calibration plot for OS (a) and PFS (b) in the validation set. Calibration curves showing average predicted probability (x-axis) against the Kaplan-Meier probability observed in the validation set (y-axis). The vertical line represents 95% CIs of the Kaplan-Meier estimates. The dashed line represents the reference line. P values calculated by the Hosmer-Lemeshow goodness-of-fit test indicate that the nomograms fit well (P ¼ .13 for OS and P ¼ .70 for PFS).
predictors, the nomograms rank the importance of each prognostic factor associated with the outcome variables. For example, the factor of tumor volume contributed about 100 points to the total prognostic score of 260 (Fig 1a). We present a specific scenario: A patient had ChildPugh class A, a MELD score of 10, and an infiltrative
HCC with portal invasion—serum AFP level was 200 ng/mL, and TTV was 70 cm3. This patient had estimated 5-year OS and 3-year PFS probabilities of 7.8% and 12.1%, respectively. The 5-year OS and 3-year PFS rates were 40.0% and 26.1% for patients with two nodular HCCs with TTV of 50 cm3, Child-Pugh class A liver function, MELD scores of 7, and AFP of 110 ng/mL.
Volume 26
’
Number 8
’
August
’
2015
Fully automated calculators for these risk-scoring models are provided electronically in Appendix B (available online at www.jvir.org).
DISCUSSION In the present study, we developed nomograms for predicting OS and PFS in patients undergoing transarterial chemoembolization. Because of the heterogeneity of the patients included in the various studies, the clinical factors described to be of prognostic importance in previous research were quite diverse (9–14), and their predictive value had yet to be firmly established. In the present study, seven clinically and statistically important variables were selected and incorporated into comprehensive numerical scoring models presented as nomograms. These models were built based on patients with a wide spectrum of HCCs including tumors invading gross vessels who were treated with transarterial chemoembolization. The models provided good discrimination, with C-indexes 4 0.75 and almost perfect reliability (calibration curves forming almost diagonal lines). The solidity of these nomograms was stringently confirmed via temporal validation (15,16). Nomograms are often used in clinical decision making in daily practice and have been proven to be more accurate than individual prognostic parameters (26,27). They incorporate diverse prognostic factors into a simple, user-friendly graphic calculator and provide individual expected risks and benefits relating to specific endpoints (eg, survival or progression) on the basis of statistical evidence. Numerical nomograms are receiving increasing attention in the medical field as tools for improving monitoring and management on a perpatient basis (28). In addition, they help in facilitating communication between the physician and patient about the anticipated effectiveness of treatment by providing a visual representation of the prediction process (29). In addition to its use for palliation, transarterial chemoembolization has been shown to be useful as a bridge preventing tumor progression to a transected liver and as an option for downstaging patients for definitive treatments such as liver resection and transplantation (30,31). A recent study by Otto et al (32) showed that the response to transarterial chemoembolization was a reliable predictor of HCC recurrence after transplant, possibly overriding conventional factors such as tumor size and number. In this regard, our prediction models may help clinicians to decide which individual patients to include on high-priority and low-priority transplant waiting lists and which patients should undergo rigorous repeated transarterial chemoembolization to downstage them to within the Milan criteria or to a resectable tumor size. In addition, use of these models should allow patients predicted to respond poorly to transarterial chemoembolization to be allocated in a timely fashion
1099
to clinical trials of transarterial chemoembolization combined with sorafenib, even when there is portal invasion of the HCC. Calculated TTV has been shown to be a predictor of HCC tumor grade and microvascular invasion (33). In addition, a TTV-based staging system developed by Hsu et al (34) was found to be more effective in predicting the outcomes of HCC treated in various ways than traditional size-based staging systems. In line with previous findings, TTV as an indicator of overall tumor burden made a substantial contribution to our models for computing probabilities of survival and progression after repeated transarterial chemoembolization, even though the infiltrative-type tumors were not completely spherical but were irregular (35). In our series, in keeping with previous observations (9,36), we found that prognostic tumor factors such as tumor number and type, portal vein invasion, and serologic value of AFP were crucial determinants of progression and survival outcomes after repeated transarterial chemoembolization, and we included these as key components of our nomograms. Apart from these tumor-related variables, underlying liver function is also an established determinant of long-term outcomes in patients with HCC (37,38). This conclusion is supported by our scoring systems because they include as major prognostic covariates MELD scores and Child-Pugh classes, which reflect hepatic functioning. Numerous patients with HCC invading the portal vein were included in our model development. Current relevant guidelines recommend sorafenib as a standard of care in the treatment of these patients with advancedstage HCC, rather than endovascular therapies including transarterial chemoembolization (3,8). However, before the introduction of sorafenib, studies of the potential role of transarterial chemoembolization in patients with advanced HCC had found considerable survival benefit for transarterial chemoembolization compared with the best supportive care and had shown that the procedure could be guaranteed to be safe at least in patients with well-preserved liver function and adequate collateral circulation around the occluded portal vein (39–43). Our prognostic calculators can serve as clinical aids in selecting patients with advanced nonmetastatic HCC who may obtain benefit from transarterial chemoembolization, particularly among patients who experienced serious adverse effects during previous sorafenib treatment and in whom first-line sorafenib treatment has failed. Apart from the fact that this study was a retrospective review in a single Asian center with a cohort mainly infected with HBV, a major limitation was that all the patients received a single protocol-based transarterial chemoembolization using a single regimen of Gelfoam and cisplatin in a selective or superselective manner. However, outcomes after transarterial chemoembolization regarding OS and PFS (25.4% and 23.2% at 5 y and 41.6% and 20.8% at 3 y) in the present series were similar to outcomes in other cohorts
1100
’
Nomograms for Postchemoembolization Outcomes
using different transarterial chemoembolization methods (3,7). Future validation studies using external sets with heterogeneous protocols, particularly including drug-eluting bead transarterial chemoembolization, which is a competitive alternative to conventional transarterial chemoembolization (44), are needed to extrapolate our models to patients in a wider range of settings. Another consideration is that the present models do not include performance status, which has been found possibly to predict long-term survival in patients with HCC, although not specifically in patients who have received chemoembolization (45). The addition of performance status information, if selected using a random forest, may improve the predictive function of our nomograms in future studies. In conclusion, we constructed and validated simple-touse graphic indicators to compute the probabilities of the expected benefit of transarterial chemoembolization in terms of individual patient-specific progression and survival. We used easily accessible and classically proven clinical variables and what we believe to be the largest set of homogeneous cohort data yet available. These nomograms would offer clinical usefulness to clinicians to decide whether to proceed with surgical strategies or to use combined or solitary chemotherapeutic agents (eg, sorafenib) in patients with HCC tumors that are amenable to chemoembolization and to counsel patients about the treatment plans. Nomograms also would be useful to differentiate target patient populations for clinical trials.
REFERENCES 1. Parkin DM, Bray F, Ferlay J, Pisani P. Estimating the world cancer burden: Globocan 2000. Int J Cancer 2001; 94:153–156. 2. Llovet JM, Burroughs A, Bruix J. Hepatocellular carcinoma. Lancet 2003; 362:1907–1917. 3. Bruix J, Sherman M. American Association for the Study of Liver Disease. Management of hepatocellular carcinoma: an update. Hepatology 2011; 53:1020–1022. 4. Llovet JM, Bruix J. Systematic review of randomized trials for unresectable hepatocellular carcinoma: chemoembolization improves survival. Hepatology 2003; 37:429–442. 5. Llovet JM, Real MI, Montana X, et al. Arterial embolisation or chemoembolisation versus symptomatic treatment in patients with unresectable hepatocellular carcinoma: a randomised controlled trial. Lancet 2002; 359:1734–1739. 6. Camma C, Schepis F, Orlando A, et al. Transarterial chemoembolization for unresectable hepatocellular carcinoma: meta-analysis of randomized controlled trials. Radiology 2002; 224:47–54. 7. Bolondi L, Burroughs A, Dufour JF, et al. Heterogeneity of patients with intermediate (BCLC B) hepatocellular carcinoma: proposal for a subclassification to facilitate treatment decisions. Semin Liver Dis 2012; 32: 348–359. 8. European Association for the Study of the Liver; European Organisation for Research and Treatment of Cancer. EASL-EORTC clinical practice guidelines: management of hepatocellular carcinoma. J Hepatol 2012; 56: 908–943. 9. Takayasu K, Arii S, Ikai I, et al. Prospective cohort study of transarterial chemoembolization for unresectable hepatocellular carcinoma in 8510 patients. Gastroenterology 2006; 131:461–469. 10. Cabibbo G, Genco C, Di Marco V, et al. Predicting survival in patients with hepatocellular carcinoma treated by transarterial chemoembolisation. Aliment Pharmacol Ther 2011; 34:196–204.
Ha et al
’
JVIR
11. Eltawil KM, Berry R, Abdolell M, Molinari M. Analysis of survival predictors in a prospective cohort of patients undergoing transarterial chemoembolization for hepatocellular carcinoma in a single Canadian centre. HPB 2012; 14:162–170. 12. Matsuda M, Omata F, Fuwa S, et al. Prognosis of patients with hepatocellular carcinoma treated solely with transcatheter arterial chemoembolization: risk factors for one-year recurrence and two-year mortality (preliminary data). Intern Med 2013; 52:847–853. 13. Shi M, Chen JA, Lin XJ, et al. Transarterial chemoembolization as initial treatment for unresectable hepatocellular carcinoma in southern China. World J Gastroenterol 2010; 16:264–269. 14. Wang Y, Chen Y, Ge N, et al. Prognostic significance of alphafetoprotein status in the outcome of hepatocellular carcinoma after treatment of transarterial chemoembolization. Ann Surg Oncol 2012; 19: 3540–3546. 15. Steyerberg EW. Clinical Prediction Models: A Practical Approach to Development, Validation, and Updating. New York: Springer; 2008. 16. Justice AC, Covinsky KE, Berlin JA. Assessing the generalizability of prognostic information. Ann Intern Med 1999; 130:515–524. 17. Toso C, Trotter J, Wei A, et al. Total tumor volume predicts risk of recurrence following liver transplantation in patients with hepatocellular carcinoma. Liver Transpl 2008; 14:1107–1115. 18. Shim JH, Lee HC, Won HJ, et al. Maximum number of target lesions required to measure responses to transarterial chemoembolization using the enhancement criteria in patients with intrahepatic hepatocellular carcinoma. J Hepatol 2012; 56:406–411. 19. Ishwaran H, Kogalur UB. Random Survival Forests. R package version 1.6.1; 2015. Available at: http://cran.r-project.org/web/packages/random ForestSRC/index.html. Accessed May 29, 2015. 20. Ishwaran H, Kogalur UB. Consistency of random survival forests. Stat Probab Lett 2010; 80:1056–1064. 21. Lin DY, Wei LJ, Ying Z. Model-checking techniques based on cumulative residuals. Biometrics 2002; 58:1–12. 22. Ambler G, Benner A. mfp: Multivariable Fractional Polynomials. R package version 1.5.1; 2015. Available at: http://cran.r-project.org/web/pack ages/mfp/index.html. Accessed May 29, 2015. 23. Harrell F Jr. rms: Regression Modeling Strategies. R package version 4.31; 2015. Available at: http://cran.r-project.org/web/packages/rms/index. html. Accessed May 29, 2015. 24. Lencioni R, Llovet JM. Modified RECIST (mRECIST) assessment for hepatocellular carcinoma. Semin Liver Dis 2010; 30:52–60. 25. Kim GA, Shim JH, Yoon SM, et al. Comparison of chemoembolization with and without radiation therapy and sorafenib for advanced hepatocellular carcinoma with portal vein tumor thrombosis: a propensity score analysis. J Vasc Interv Radiol 2015; 26:320–329, e6. 26. Tseng JY, Yen MS, Twu NF, et al. Prognostic nomogram for overall survival in stage IIB-IVA cervical cancer patients treated with concurrent chemoradiotherapy. Am J Obstet Gynecol 2010; 202(174):e1–e7. 27. Choi GH, Han S, Shim JH, et al. Prognostic scoring models for patients undergoing sorafenib treatment for advanced stage hepatocellular carcinoma in real-life practice. Am J Clin Oncol In press; available online September 29, 2014. 28. Isariyawongse BK, Kattan MW. Prediction tools in surgical oncology. Surg Oncol Clin North Am 2012; 21:439–447. 29. Kattan MW, Marasco J. What is a real nomogram? Semin Oncol 2010; 37:23–26 30. Lei J, Yan L. Comparison between living donor liver transplantation recipients who met the Milan and UCSF criteria after successful downstaging therapies. J Gastrointest Surg 2012; 16:2120–2125. 31. Yao FY, Kerlan RK Jr, Hirose R, et al. Excellent outcome following down-staging of hepatocellular carcinoma prior to liver transplantation: an intention-to-treat analysis. Hepatology 2008; 48:819–827. 32. Otto G, Schuchmann M, Hoppe-Lotichius M, et al. How to decide about liver transplantation in patients with hepatocellular carcinoma: size and number of lesions or response to TACE?. J Hepatol 2013; 59:279–284. 33. Cucchetti A, Piscaglia F, Grigioni AD, et al. Preoperative prediction of hepatocellular carcinoma tumour grade and micro-vascular invasion by means of artificial neural network: a pilot study. J Hepatol 2010; 52: 880–888. 34. Hsu CY, Huang YH, Hsia CY, et al. A new prognostic model for hepatocellular carcinoma based on total tumor volume: the Taipei Integrated Scoring System. J Hepatol 2010; 53:108–117. 35. Han K, Kim JH, Yoon HM, et al. Transcatheter arterial chemoembolization for infiltrative hepatocellular carcinoma: clinical safety and efficacy and factors influencing patient survival. Korean J Radiol 2014; 15:464–471.
Volume 26
’
Number 8
’
August
’
2015
36. Kang SH, Kim do Y, Jeon SM, et al. Clinical characteristics and prognosis of hepatocellular carcinoma with different sets of serum AFP and PIVKA-II levels. Eur J Gastroenterol Hepatol 2012; 24: 849–856. 37. Sawhney S, Montano-Loza AJ, Salat P, et al. Transarterial chemoembolization in patients with hepatocellular carcinoma: predictors of survival. Can J Gastroenterol 2011; 25:426–432. 38. Bruix J, Sala M, Llovet JM. Chemoembolization for hepatocellular carcinoma. Gastroenterology 2004; 127:S179–S188. 39. Chung GE, Lee JH, Kim HY, et al. Transarterial chemoembolization can be safely performed in patients with hepatocellular carcinoma invading the main portal vein and may improve the overall survival. Radiology 2011; 258:627–634. 40. Georgiades CS, Hong K, D’Angelo M, Geschwind JF. Safety and efficacy of transarterial chemoembolization in patients with unresectable hepatocellular carcinoma and portal vein thrombosis. J Vasc Interv Radiol 2005; 16:1653–1659.
1101
41. Kim KM, Kim JH, Park IS, et al. Reappraisal of repeated transarterial chemoembolization in the treatment of hepatocellular carcinoma with portal vein invasion. J Gastroenterol Hepatol 2009; 24:806–814. 42. Minagawa M, Makuuchi M. Treatment of hepatocellular carcinoma accompanied by portal vein tumor thrombus. World J Gastroenterol 2006; 12:7561–7567. 43. Pinter M, Hucke F, Graziadei I, et al. Advanced-stage hepatocellular carcinoma: transarterial chemoembolization versus sorafenib. Radiology 2012; 263:590–599. 44. Lammer J, Malagari K, Vogl T, et al. Prospective randomized study of doxorubicin-eluting-bead embolization in the treatment of hepatocellular carcinoma: results of the PRECISION V study. Cardiovasc Intervent Radiol 2010; 33:41–52. 45. Hsu CY, Lee YH, Hsia CY, et al. Performance status in patients with hepatocellular carcinoma: determinants, prognostic impact, and ability to improve the Barcelona Clinic Liver Cancer system. Hepatology 2013; 57: 112–119.
1101.e1
’
Nomograms for Postchemoembolization Outcomes
APPENDIX A. TRANSARTERIAL CHEMOEMBOLIZATION PROCEDURE Transarterial chemoembolization was conducted by experienced interventional radiologists under local anesthesia with right femoral access. Angiography of the celiac trunk and superior mesenteric artery was initially performed in all patients to assess the anatomy, tumor burden, and patency of the portal vein. After angiography, cisplatin (2 mg/kg cisplatin in distilled normal saline) was administered for 15 minutes into the right lobar, left lobar, or proper hepatic artery according to the location of the tumor. An emulsion of cisplatin in ethiodized oil (Lipiodol; Laboratoire Guerbet, AulnaySous-Bois, France) in a 1:1 ratio with the adjusted concentration of 20 ng/mL was delivered next, followed by embolization with 1-mm diameter absorbable gelatin sponge particles (Gelfoam; Ethicon, Inc, Somerville, New Jersey) in a selective or, if possible, superselective manner until arterial flow stasis was achieved. The dose of ethiodized oil depended on the tumor size. The cisplatin injection was preceded by intravenous hydration and antiemetics and followed by further intravenous hydration. Dynamic contrast-enhanced computed tomography of the liver was always performed 4–6 weeks after transarterial chemoembolization to assess the effect of each procedure. Transarterial chemoembolization was repeated every 6–8 week if there were residual tumors, as long as new metastasis had not developed and hepatic function was preserved.
Statistical Analysis We used a Cox proportional hazards model to develop the prediction model for estimating 5-year OS and 3-year PFS. The covariates age, sex, calculated tumor volume, tumor number, tumor type (nodular or infiltrative), bilobar involvement, Child-Pugh class, MELD score, AFP value, presence of HBV, presence of cirrhosis, hepatic vein invasion, and portal vein invasion were selected from a survey of the literature and considered as candidate prognostic variables. There were two missing
Ha et al
’
JVIR
values for AFP in the derivation cohort, and the data were replaced by a median value of the overall cohort (99.9 ng/mL). A model-building approach was used in which all the above-mentioned candidate covariates were at first included. A random forest model was used for selecting variables (19). The major advantage of this model is that it can produce an unbiased measure of the importance of a variable even in mixtures of categorical and continuous variables (20). Following this, clinically or statistically less meaningful variables, such as variables with contradictory coefficient signs, were excluded one by one. To avoid a multicollinearity problem among the variables finally selected, we computed a variation inflation factor and assessed multicollinearity (variance inflation factor Z 5). The proportional hazards assumption also was verified using the scaled Schoenfeld residual test. The overall goodness-of-fit of the model was measured using the Cramér-von-Mises test based on cumulative residuals. The Cramér-von-Mises test can be used to examine the linearity assumption of predictors in a multivariate model (21). Based on the multiple fractional polynomial model (22), we employed the log-transformed AFP variable and the transformed MELD variable (MELD score/10)2. Discrimination and calibration powers were examined according to Cindexes and the calibration curves for the derivation and validation cohorts. A Hosmer-Lemeshow–type test was used to assess calibration ability, and the complete-case method was used to handle missing data. To generate user-friendly nomograms, we used the rms package developed by Harrell in R software version 2.15.1 (23). The 1-, 3-, and 5-year OS probabilities and 1-, 3-, and 5-year PFS probabilities also are provided. Points were allocated for each level of prognostic factor according to the scale shown in the nomogram. The total score was determined by summing individual points, and these were used to calculate t-year survival probability and progression-free probability (t ¼ 1, 3 or 5). The function of cph in the rms package produces biascorrected performance measures using the bootstrap resampling method. All analyses were based on a twosided test at a significance level of .05.
Nomograms for Predicting Outcomes after Chemoembolization in Patients with Nonmetastatic Hepatocellular Carcinoma Yeonjung Ha, MD, Seungbong Han, PhD, Ju Hyun Shim, MD, Gi-Young Ko, MD, Hyun-Ki Yoon, MD, Kyu-Bo Sung, MD, Danbi Lee, MD, Kang Mo Kim, MD, Young-Suk Lim, MD, Young-Hwa Chung, MD, Yung Sang Lee, MD, and Han Chu Lee, MD
ABSTRACT Purpose: To construct prognostic nomograms capable of estimating individual probabilities of tumor progression and overall survival (OS) at specific time points during serial transarterial chemoembolization for hepatocellular carcinoma (HCC). Materials and Methods: The study included 1,181 consecutive patients with nonmetastatic HCC undergoing repeated transarterial chemoembolization at a single tertiary referral center. Patients were assigned to 2 cohorts according to the first transarterial chemoembolization date: derivation (2004–2006; n ¼ 854) and validation (2007; n ¼ 327) sets. Multivariate Cox proportional hazards models were developed based on covariates derived before transarterial chemoembolization and assessed for their association with 5-year OS and 3-year progression-free survival (PFS). The accuracy of the models was internally and externally validated. Results: The 5-year OS of the derivation set was 25.4%, and 3-year PFS was 20.8%. Nomograms for OS and PFS were built into the derivation set incorporating the following factors: log [tumor volume] calculated as 4/3 3.14 (maximum radius of tumor in cm3); tumor number; tumor type (nodular or infiltrative); Child-Pugh class (A or B); (model for end-stage liver disease score/10)2; log [α-fetoprotein]; and portal vein invasion. The models had good discrimination and calibration abilities with C-indexes of 0.80 (5-y survival) and 0.77 (3-y progression). The results of external validation confirmed that these models performed well in terms of discrimination and goodness-of-fit (C-indexes 0.77 for 5-y survival and 0.73 for 3-y progression). Conclusion: Nomograms quantifying the survival and progression outcomes in patients treated with transarterial chemoembolization are useful clinical aids in providing personalized care.
ABBREVIATIONS AFP = α-fetoprotein, CI = confidence interval, HBV = hepatitis B virus, HCC = hepatocellular carcinoma, MELD = model for endstage liver disease, OS = overall survival, PFS = progression-free survival, TTV = total tumor volume From the Department of Gastroenterology (Y.H., J.H.S., D.L., K.M.K., Y.-S.L., Y.-H.C., Y.S.L., H.C.L.), Asan Liver Center, and Department of Radiology (G.-Y.K., H.-K.Y., K.-B.S.), Asan Medical Center, University of Ulsan College of Medicine, 88 Olympic-ro 43-gil, Songpa-gu, Seoul 138-736, Republic of Korea; and Department of Applied Statistics (S.H.), Gachon University, Seongnam-si, Gyeonggi-do, Republic of Korea. Received October 29, 2014; final revision received March 29, 2015; accepted April 12, 2015. Address correspondence to J.H.S., E-mail: [email protected]; or H.C.L., E-mail: [email protected] Y.H. and S.H. contributed equally to this article. None of the authors have identified a conflict of interest. Appendix A and Appendix B are available online at www.jvir.org. & SIR, 2015 J Vasc Interv Radiol 2015; 26:1093–1101 http://dx.doi.org/10.1016/j.jvir.2015.04.010
Hepatocellular carcinoma (HCC) is one of the most common lethal malignancies worldwide (1). Despite increased clinical surveillance, only 30% of cases of HCC are diagnosed in the early stages (2). In other words, 70% are still detected in the late stages, in which curative treatments such as surgical resection, percutaneous ablation, or liver transplantation are not feasible (3). Some patients with early-stage disease are not eligible for curative management because of their general medical condition, the location of the tumor, or the shortage of liver donors. Palliative treatments are often required.
1094
’
Nomograms for Postchemoembolization Outcomes
Transarterial chemoembolization has been shown to be a life-prolonging treatment and has been used in practice for a wide spectrum of HCCs including unresectable, untransplantable, and unablatable tumors at an early stage (4–6). However, because of the heterogeneity of the patient populations that underwent transarterial chemoembolization, variable survival outcomes have been reported, with a 2-year survival rate ranging from 20% to 60% among patients with intermediate-stage HCC for whom transarterial chemoembolization is the current standard recommendation (3,7). In certain subsets of patients who had decompensated liver disease, advanced liver dysfunction, macroscopic vascular invasion, or extrahepatic spread, this procedure was regarded as essentially detrimental (8). Many studies have been performed to identify prognostic factors using these heterogeneous populations, and variables such as liver function, tumor-node-metastasis stage, performance status, presence of hepatitis B virus (HBV), and α-fetoprotein (AFP) have been identified (9–14). However, the factors identified varied considerably from study to study, so that comprehensive predictions of patient outcomes have been difficult to make. To obtain per-patient prognoses and help physicians decide on treatment options by identifying more homogeneous subsets of patients, we constructed and validated comprehensive risk-scoring models in the form of nomograms consisting of easily accessible variables predicting overall survival (OS) and progression-free survival (PFS) in patients treated with transarterial chemoembolization.
MATERIALS AND METHODS
Ha et al
’
JVIR
the time of diagnosis, or (c) had other histologically confirmed malignancies before or within 5 years of the diagnosis of HCC. There were 1,181 patients enrolled, and these patients were divided into a derivation set and a validation set, according to the date of initial transarterial chemoembolization (15,16). The derivation set for the primary analysis consisted of 854 patients (72.3%) who underwent their first transarterial chemoembolization between January 2004 and December 2006, and the temporal validation set consisted of the remaining 327 patients (27.7%) (15,16).
Data Collection At the time of diagnosis, demographic and clinical variables including age, sex, tumor size, tumor number, tumor type (nodular or infiltrative), portal vein invasion, hepatic vein invasion, extrahepatic metastases, presence of cirrhosis, bilobar involvement, Child-Pugh class, model for end-stage liver disease (MELD) score, AFP level, and presence of HBV were collected for each patient from the database of our center. Total tumor volume (TTV), which had been initially proposed by Toso et al (17), was calculated from the sum of tumor volumes (4/3 3.14 [maximum radius of the tumor in cm3]). All of the enrolled cases were confirmed as HCC by liver protocol computed tomography or magnetic resonance imaging or liver biopsy according to current American Association for the Study of Liver Diseases guidelines (3). Before the initial transarterial chemoembolization, chest computed tomography and bone scans were checked on a routine basis, and, if indicated, positron emission tomography was performed to exclude extrahepatic metastases.
Study Design This was a retrospective cohort study undertaken to construct risk-scoring models for patients who receive transarterial chemoembolization secondary to unresectable HCC, in whom the derivation and validation cohorts from the gastroenterology department of a single tertiary referral center were temporarily separated.
Transarterial Chemoembolization Procedure The routine protocol of our hospital has been described elsewhere (18). Details of the procedure are provided in Appendix A (available online at www.jvir.org) (19–23).
Derivation and Validation Sets
Study Endpoints
Consecutive data were collected for patients with a new diagnosis of HCC who were treated with transarterial chemoembolization as initial therapy at a single center between January 2004 and December 2007. The diagnosis of HCC based on the American Association for the Study of Liver Diseases practice guidelines was reconfirmed in all included patients (3). Transarterial chemoembolization was the optimal treatment, and curative surgical or locoregional modalities were impossible or contraindicated at the time of enrollment of these patients. The following patients were excluded: patients who (a) were lost to follow-up after the first transarterial chemoembolization session without evaluation of the treatment response, (b) had extrahepatic metastases at
The primary endpoint was OS, which was defined as the interval between the date of the first transarterial chemoembolization and death or last follow-up visit. During follow-up, data were censored at the time when a treatment strategy of curative intent, such as surgical resection, liver transplantation, or percutaneous ablation, was attempted. Patients lost to follow-up were censored at the time of their last visit, and living patients were censored on December 31, 2012. The secondary endpoint was PFS estimated from the date of the first transarterial chemoembolization session to progression or the last follow-up visit. Progressive disease was confirmed according to the modified Response Evaluation Criteria in Solid Tumors (24). Death and follow-up
Volume 26
’
Number 8
’
August
’
2015
1095
data were fully accessible through the database of our center and collected until December 31, 2012.
Statistical Analysis To develop the prediction model for estimating 5-year OS and 3-year PFS, we used a Cox proportional hazards model as a base model. Based on a literature survey, the following variables were selected for candidate potential predictors: age, sex, calculated tumor volume, tumor number, tumor type (nodular or infiltrative), bilobar involvement, Child-Pugh class, MELD score, AFP value, presence of HBV, presence of cirrhosis, hepatic vein invasion, and portal vein invasion. A random forest model was employed for tentative variable selection (19). A detailed description of variable selection and the final model derivation is provided in Appendix A (available online at www.jvir.org).
was 99.7 ng/mL. At the time of diagnosis, 280 (23.7%) patients had portal vein invasion. Of these cases, 90 (32.2%) patients had HCCs invading main venous branches, and 190 (67.8%) patients had HCCs invading lobar or segmental venous branches (25). The two sets were similar in terms of age, sex, tumor size, tumor volume, tumor type, cirrhosis, Child-Pugh class, MELD score, AFP value, and HBV status. However, there were more patients with more than four tumors in the validation set (P ¼ .007). The proportions of patients with tumors in both lobes and with portal vein invasion also were higher in the validation set (P ¼ .015 and .027 for each). Over the study period, 73 (8.5%) patients in the derivation set and 22 (6.7%) patients in the validation set were lost to follow-up at least after the evaluation of their response to the first session and then censored at the date of their last follow-up examination.
Covariates Selected in the Nomograms
RESULTS Baseline Characteristics of Enrolled Patients The demographic and clinical parameters of the derivation and validation sets are provided in Table 1. The mean age of the 1,181 patients was 55.9 years (standard deviation [SD], 10.1 y). Of the patients, 981 (83.1%) were men, 953 (80.7%) had liver cirrhosis, 931 (78.8%) were in Child-Pugh class A, and 927 (78.5%) had nodular-type tumors. Median tumor volume was 39.0 cm3, mean MELD score was 9.0 (SD, 2.5), and median serum AFP
In the derivation set using the input variables of age, sex, calculated tumor volume, tumor number, tumor type (nodular or infiltrative), bilobar involvement, ChildPugh class, MELD score, AFP value, presence of HBV, presence of cirrhosis, hepatic vein invasion, and portal vein invasion, the seven most important variables for survival and progression were provided by the random forest model. These variables were tumor volume, tumor number, tumor type, Child-Pugh class, MELD score, AFP value, and portal vein invasion (Table 2). The final models are shown in Table 3.
Table 1 . Characteristics of the Derivation and Validation Sets All Patients (N ¼ 1,181)
Derivation Set (n ¼ 854)
Validation Set (n ¼ 327)
55.9 ⫾ 10.1
55.7 ⫾ 10.0
56.4 ⫾ 10.4
.36
981 (83.1)
717 (84.0)
264 (80.7)
.19
39.0 (7.0–321.0)
34.5 (7.0–312.8)
48.0 (8.0–382.0)
.40 .007
1
616 (52.2)
455 (53.3)
161 (49.2)
2 3
219 (18.5) 83 (7.0)
168 (19.7) 59 (6.9)
51 (15.6) 24 (7.3)
Characteristics Age (y), mean ⫾ SD Male, n (%) Tumor volume (cm3), median (IQR) Tumor number, n (%)
4
P Value
29 (2.5)
24 (2.8)
5 (1.5)
234 (19.8)
148 (17.3)
86 (26.3)
927 (78.5)
674 (78.9)
253 (77.4)
Infiltrative Bilobar involvement, n (%)
254 (21.5) 326 (27.6)
180 (21.1) 219 (25.6)
74 (22.6) 107 (32.7)
.015
Child-Pugh class A, n (%)
931 (78.8)
671 (78.6)
260 (79.5)
.72
MELD score, mean ⫾ SD Presence of cirrhosis, n (%)
9.0 ⫾ 2.5 953 (80.7)
9.1 ⫾ 2.7 684 (80.1)
8.8 ⫾ 2.0 269 (82.3)
.18 .40
Portal vein invasion, n (%)
280 (23.7)
188 (22.0)
92 (28.1)
.027
36 (3.0) 99.7 (14.3–1,250.0)
31 (3.6) 111.5 (13.3–1,100.0)
5 (1.5) 73.5 (13.7–1,695.0)
.06 .63
899 (76.1)
647 (75.8)
252 (77.1)
0.64
Z5 Tumor type, n (%) Nodular
Hepatic vein invasion, n (%) Serum AFP level (ng/mL), median (IQR) Presence of HBV
.56
AFP ¼ α-fetoprotein, HBV ¼ hepatitis B virus, IQR ¼ interquartile range, MELD ¼ model for end-stage liver disease, SD ¼ standard deviation.
1096
’
Nomograms for Postchemoembolization Outcomes
Ha et al
’
JVIR
Table 2 . Variables Selected by a Random Forest Model OS Variables
Importance
Tumor number
0.0194
Tumor volume Log [AFP]
0.0130 0.0076
PFS Relative Importance 1 0.6662 0.3903
Importance
Relative Importance
0.0095
0.4305
0.0221 0.0132
1 0.5993
MELD score, per point
0.0086
0.4420
0.0039
0.1757
Portal vein invasion Infiltrative tumor type
0.0086 0.0039
0.4432 0.1997
0.0018 0.0004
0.0828 0.0200
Age
0.0023
0.1205
0.0021
0.0940
0.0005 0.0002
0.0278 0.0110
0.0037 0.0004
0.1691 0.0178
Child-Pugh class B Male
0
0.0017
0.0001
0.0036
HBV Liver cirrhosis
0.0011 0.0012
0.0544 0.0606
0.0007 0.0001
0.0315 0.0320
Bilobar involvement
0.0018
0.0915
0.0037
0.1662
Hepatic vein invasion
AFP ¼ α-fetoprotein, HBV ¼ hepatitis B virus, MELD ¼ model for end-stage liver disease, OS ¼ overall survival, PFS ¼ progressionfree survival.
Table 3 . Final Models for OS and PFS OS
PFS
HR (95% CI)
P Value
HR (95% CI)
Portal vein invasion
1.61 (1.26–2.04)
o .001
1.16 (0.93–1.45)
.19
Tumor number Log [tumor volume]
1.28 (1.21–1.36) 1.19 (1.13–1.24)
o .001 o .001
1.21 (1.15–1.27) 1.16 (1.12–1.21)
o .001 o .001
(MELD score/10)2
0.64 (0.53–0.78)
o .001
0.81 (0.69–0.95)
.009
Infiltrative tumor type Log [AFP]
1.26 (0.99–1.61) 1.08 (1.04–1.11)
.06 o .001
1.20 (0.96–1.51) 1.06 (1.03–1.09)
.11 o .001
Child-Pugh class B
1.42 (1.11–1.83)
.006
1.43 (1.15–1.78)
.001
Variables
P Value
AFP ¼ α-fetoprotein, CI ¼ confidence interval, HR ¼ hazard ratio, MELD ¼ model for end-stage liver disease, OS ¼ overall survival, PFS ¼ progression-free survival.
Internal and External Validation of Constructed Nomograms In the derivation set, the 5-year OS was 25.4% with a median survival time of 28.5 months (95% confidence interval [CI], 25.9–31.0 mo), and the 3-year PFS was 20.8% with a median of 14.3 months (95% CI, 13.0–15.5 mo). Using the analysis described earlier, the nomograms for OS (Fig 1a) and PFS (Fig 1b) were developed into a derivation set by parameterization of the variables and summing the points corresponding to each variable: log [tumor volume] þ tumor number (5 for Z 5) þ tumor type (nodular: 0, infiltrative: 1) þ Child-Pugh class (A: 0, B: 1) þ (MELD score/10)2 þ log [AFP] þ portal vein invasion (no: 0, yes: 1). We assessed the individual contributions of the seven variables selected by the random forest model to discrimination and calibration. The final model had C-indexes for OS of 0.84, 0.82, and 0.80 at 1 year, 3 years, and 5 years; in the PFS analysis, the C-indexes were 0.76, 0.77, and 0.78 at 1 year, 3 years, and 5 years. Calibration curves for the
predicted and actual probabilities of OS (Fig 2a) and PFS (Fig 2b) yielded good calibration abilities based on the results of the Hosmer-Lemeshow test (P ¼ .87 for OS and P ¼ .30 for PFS). The scoring system built for the derivation set with median OS and PFS times of 28.5 months (95% CI, 25.9–31.0 mo) and 14.3 months (95% CI, 13.0–15.5 mo), respectively, was externally validated in the temporal validation cohort. The external validation analysis confirmed that these models performed well in terms of discrimination (C-indexes of 0.78, 0.80, and 0.77 for 1-, 3-, and 5-year OS and 0.72 for 1-year PFS and 0.73 for 3- and 5-year PFS) and of calibration (P values of the Hosmer-Lemeshow test, P ¼ .13 for OS and P ¼ .70 for PFS) (Fig 3a, b).
Examples of the Use of Nomograms Prognostic nomograms were developed for predicting OS and PFS rates at 1, 3, and 5 years (Fig 1a, b). Points were assigned for tumor log [tumor volume], tumor
Volume 26
’
Number 8
’
August
’
2015
1097
Figure 1. Nomogram depicting OS (a) and PFS (b). By drawing a line between each variable and the uppermost component “points,” the appropriate points can be assigned to seven variables. The sum of these seven points can be expressed on the “total points” line. Then t-year survival or progression can be calculated by connecting each point to the t-survival line (t ¼ 1, 3, and 5).
number, tumor type, portal vein invasion, (MELD score/10)2, Child-Pugh class, and log [AFP] by drawing lines upward from the corresponding values to the
“points” line. The sum of these seven points, plotted on the “total points” line, corresponded to the t-year OS or PFS probability (t ¼ 1, 3 or 5). Using these seven
1098
’
Nomograms for Postchemoembolization Outcomes
Ha et al
’
JVIR
Figure 2. Calibration plot for OS (a) and PFS (b) in the derivation set. Calibration curves showing average predicted probability (x-axis) against the Kaplan-Meier probability observed in the derivation set (y-axis). The vertical line represents 95% CIs of the Kaplan-Meier estimates. The dashed line represents the reference line. P values calculated by the Hosmer-Lemeshow goodness-of-fit test indicate that the nomograms fit well (P ¼ .87 for OS and P ¼ .30 for PFS).
Figure 3. Calibration plot for OS (a) and PFS (b) in the validation set. Calibration curves showing average predicted probability (x-axis) against the Kaplan-Meier probability observed in the validation set (y-axis). The vertical line represents 95% CIs of the Kaplan-Meier estimates. The dashed line represents the reference line. P values calculated by the Hosmer-Lemeshow goodness-of-fit test indicate that the nomograms fit well (P ¼ .13 for OS and P ¼ .70 for PFS).
predictors, the nomograms rank the importance of each prognostic factor associated with the outcome variables. For example, the factor of tumor volume contributed about 100 points to the total prognostic score of 260 (Fig 1a). We present a specific scenario: A patient had ChildPugh class A, a MELD score of 10, and an infiltrative
HCC with portal invasion—serum AFP level was 200 ng/mL, and TTV was 70 cm3. This patient had estimated 5-year OS and 3-year PFS probabilities of 7.8% and 12.1%, respectively. The 5-year OS and 3-year PFS rates were 40.0% and 26.1% for patients with two nodular HCCs with TTV of 50 cm3, Child-Pugh class A liver function, MELD scores of 7, and AFP of 110 ng/mL.
Volume 26
’
Number 8
’
August
’
2015
Fully automated calculators for these risk-scoring models are provided electronically in Appendix B (available online at www.jvir.org).
DISCUSSION In the present study, we developed nomograms for predicting OS and PFS in patients undergoing transarterial chemoembolization. Because of the heterogeneity of the patients included in the various studies, the clinical factors described to be of prognostic importance in previous research were quite diverse (9–14), and their predictive value had yet to be firmly established. In the present study, seven clinically and statistically important variables were selected and incorporated into comprehensive numerical scoring models presented as nomograms. These models were built based on patients with a wide spectrum of HCCs including tumors invading gross vessels who were treated with transarterial chemoembolization. The models provided good discrimination, with C-indexes 4 0.75 and almost perfect reliability (calibration curves forming almost diagonal lines). The solidity of these nomograms was stringently confirmed via temporal validation (15,16). Nomograms are often used in clinical decision making in daily practice and have been proven to be more accurate than individual prognostic parameters (26,27). They incorporate diverse prognostic factors into a simple, user-friendly graphic calculator and provide individual expected risks and benefits relating to specific endpoints (eg, survival or progression) on the basis of statistical evidence. Numerical nomograms are receiving increasing attention in the medical field as tools for improving monitoring and management on a perpatient basis (28). In addition, they help in facilitating communication between the physician and patient about the anticipated effectiveness of treatment by providing a visual representation of the prediction process (29). In addition to its use for palliation, transarterial chemoembolization has been shown to be useful as a bridge preventing tumor progression to a transected liver and as an option for downstaging patients for definitive treatments such as liver resection and transplantation (30,31). A recent study by Otto et al (32) showed that the response to transarterial chemoembolization was a reliable predictor of HCC recurrence after transplant, possibly overriding conventional factors such as tumor size and number. In this regard, our prediction models may help clinicians to decide which individual patients to include on high-priority and low-priority transplant waiting lists and which patients should undergo rigorous repeated transarterial chemoembolization to downstage them to within the Milan criteria or to a resectable tumor size. In addition, use of these models should allow patients predicted to respond poorly to transarterial chemoembolization to be allocated in a timely fashion
1099
to clinical trials of transarterial chemoembolization combined with sorafenib, even when there is portal invasion of the HCC. Calculated TTV has been shown to be a predictor of HCC tumor grade and microvascular invasion (33). In addition, a TTV-based staging system developed by Hsu et al (34) was found to be more effective in predicting the outcomes of HCC treated in various ways than traditional size-based staging systems. In line with previous findings, TTV as an indicator of overall tumor burden made a substantial contribution to our models for computing probabilities of survival and progression after repeated transarterial chemoembolization, even though the infiltrative-type tumors were not completely spherical but were irregular (35). In our series, in keeping with previous observations (9,36), we found that prognostic tumor factors such as tumor number and type, portal vein invasion, and serologic value of AFP were crucial determinants of progression and survival outcomes after repeated transarterial chemoembolization, and we included these as key components of our nomograms. Apart from these tumor-related variables, underlying liver function is also an established determinant of long-term outcomes in patients with HCC (37,38). This conclusion is supported by our scoring systems because they include as major prognostic covariates MELD scores and Child-Pugh classes, which reflect hepatic functioning. Numerous patients with HCC invading the portal vein were included in our model development. Current relevant guidelines recommend sorafenib as a standard of care in the treatment of these patients with advancedstage HCC, rather than endovascular therapies including transarterial chemoembolization (3,8). However, before the introduction of sorafenib, studies of the potential role of transarterial chemoembolization in patients with advanced HCC had found considerable survival benefit for transarterial chemoembolization compared with the best supportive care and had shown that the procedure could be guaranteed to be safe at least in patients with well-preserved liver function and adequate collateral circulation around the occluded portal vein (39–43). Our prognostic calculators can serve as clinical aids in selecting patients with advanced nonmetastatic HCC who may obtain benefit from transarterial chemoembolization, particularly among patients who experienced serious adverse effects during previous sorafenib treatment and in whom first-line sorafenib treatment has failed. Apart from the fact that this study was a retrospective review in a single Asian center with a cohort mainly infected with HBV, a major limitation was that all the patients received a single protocol-based transarterial chemoembolization using a single regimen of Gelfoam and cisplatin in a selective or superselective manner. However, outcomes after transarterial chemoembolization regarding OS and PFS (25.4% and 23.2% at 5 y and 41.6% and 20.8% at 3 y) in the present series were similar to outcomes in other cohorts
1100
’
Nomograms for Postchemoembolization Outcomes
using different transarterial chemoembolization methods (3,7). Future validation studies using external sets with heterogeneous protocols, particularly including drug-eluting bead transarterial chemoembolization, which is a competitive alternative to conventional transarterial chemoembolization (44), are needed to extrapolate our models to patients in a wider range of settings. Another consideration is that the present models do not include performance status, which has been found possibly to predict long-term survival in patients with HCC, although not specifically in patients who have received chemoembolization (45). The addition of performance status information, if selected using a random forest, may improve the predictive function of our nomograms in future studies. In conclusion, we constructed and validated simple-touse graphic indicators to compute the probabilities of the expected benefit of transarterial chemoembolization in terms of individual patient-specific progression and survival. We used easily accessible and classically proven clinical variables and what we believe to be the largest set of homogeneous cohort data yet available. These nomograms would offer clinical usefulness to clinicians to decide whether to proceed with surgical strategies or to use combined or solitary chemotherapeutic agents (eg, sorafenib) in patients with HCC tumors that are amenable to chemoembolization and to counsel patients about the treatment plans. Nomograms also would be useful to differentiate target patient populations for clinical trials.
REFERENCES 1. Parkin DM, Bray F, Ferlay J, Pisani P. Estimating the world cancer burden: Globocan 2000. Int J Cancer 2001; 94:153–156. 2. Llovet JM, Burroughs A, Bruix J. Hepatocellular carcinoma. Lancet 2003; 362:1907–1917. 3. Bruix J, Sherman M. American Association for the Study of Liver Disease. Management of hepatocellular carcinoma: an update. Hepatology 2011; 53:1020–1022. 4. Llovet JM, Bruix J. Systematic review of randomized trials for unresectable hepatocellular carcinoma: chemoembolization improves survival. Hepatology 2003; 37:429–442. 5. Llovet JM, Real MI, Montana X, et al. Arterial embolisation or chemoembolisation versus symptomatic treatment in patients with unresectable hepatocellular carcinoma: a randomised controlled trial. Lancet 2002; 359:1734–1739. 6. Camma C, Schepis F, Orlando A, et al. Transarterial chemoembolization for unresectable hepatocellular carcinoma: meta-analysis of randomized controlled trials. Radiology 2002; 224:47–54. 7. Bolondi L, Burroughs A, Dufour JF, et al. Heterogeneity of patients with intermediate (BCLC B) hepatocellular carcinoma: proposal for a subclassification to facilitate treatment decisions. Semin Liver Dis 2012; 32: 348–359. 8. European Association for the Study of the Liver; European Organisation for Research and Treatment of Cancer. EASL-EORTC clinical practice guidelines: management of hepatocellular carcinoma. J Hepatol 2012; 56: 908–943. 9. Takayasu K, Arii S, Ikai I, et al. Prospective cohort study of transarterial chemoembolization for unresectable hepatocellular carcinoma in 8510 patients. Gastroenterology 2006; 131:461–469. 10. Cabibbo G, Genco C, Di Marco V, et al. Predicting survival in patients with hepatocellular carcinoma treated by transarterial chemoembolisation. Aliment Pharmacol Ther 2011; 34:196–204.
Ha et al
’
JVIR
11. Eltawil KM, Berry R, Abdolell M, Molinari M. Analysis of survival predictors in a prospective cohort of patients undergoing transarterial chemoembolization for hepatocellular carcinoma in a single Canadian centre. HPB 2012; 14:162–170. 12. Matsuda M, Omata F, Fuwa S, et al. Prognosis of patients with hepatocellular carcinoma treated solely with transcatheter arterial chemoembolization: risk factors for one-year recurrence and two-year mortality (preliminary data). Intern Med 2013; 52:847–853. 13. Shi M, Chen JA, Lin XJ, et al. Transarterial chemoembolization as initial treatment for unresectable hepatocellular carcinoma in southern China. World J Gastroenterol 2010; 16:264–269. 14. Wang Y, Chen Y, Ge N, et al. Prognostic significance of alphafetoprotein status in the outcome of hepatocellular carcinoma after treatment of transarterial chemoembolization. Ann Surg Oncol 2012; 19: 3540–3546. 15. Steyerberg EW. Clinical Prediction Models: A Practical Approach to Development, Validation, and Updating. New York: Springer; 2008. 16. Justice AC, Covinsky KE, Berlin JA. Assessing the generalizability of prognostic information. Ann Intern Med 1999; 130:515–524. 17. Toso C, Trotter J, Wei A, et al. Total tumor volume predicts risk of recurrence following liver transplantation in patients with hepatocellular carcinoma. Liver Transpl 2008; 14:1107–1115. 18. Shim JH, Lee HC, Won HJ, et al. Maximum number of target lesions required to measure responses to transarterial chemoembolization using the enhancement criteria in patients with intrahepatic hepatocellular carcinoma. J Hepatol 2012; 56:406–411. 19. Ishwaran H, Kogalur UB. Random Survival Forests. R package version 1.6.1; 2015. Available at: http://cran.r-project.org/web/packages/random ForestSRC/index.html. Accessed May 29, 2015. 20. Ishwaran H, Kogalur UB. Consistency of random survival forests. Stat Probab Lett 2010; 80:1056–1064. 21. Lin DY, Wei LJ, Ying Z. Model-checking techniques based on cumulative residuals. Biometrics 2002; 58:1–12. 22. Ambler G, Benner A. mfp: Multivariable Fractional Polynomials. R package version 1.5.1; 2015. Available at: http://cran.r-project.org/web/pack ages/mfp/index.html. Accessed May 29, 2015. 23. Harrell F Jr. rms: Regression Modeling Strategies. R package version 4.31; 2015. Available at: http://cran.r-project.org/web/packages/rms/index. html. Accessed May 29, 2015. 24. Lencioni R, Llovet JM. Modified RECIST (mRECIST) assessment for hepatocellular carcinoma. Semin Liver Dis 2010; 30:52–60. 25. Kim GA, Shim JH, Yoon SM, et al. Comparison of chemoembolization with and without radiation therapy and sorafenib for advanced hepatocellular carcinoma with portal vein tumor thrombosis: a propensity score analysis. J Vasc Interv Radiol 2015; 26:320–329, e6. 26. Tseng JY, Yen MS, Twu NF, et al. Prognostic nomogram for overall survival in stage IIB-IVA cervical cancer patients treated with concurrent chemoradiotherapy. Am J Obstet Gynecol 2010; 202(174):e1–e7. 27. Choi GH, Han S, Shim JH, et al. Prognostic scoring models for patients undergoing sorafenib treatment for advanced stage hepatocellular carcinoma in real-life practice. Am J Clin Oncol In press; available online September 29, 2014. 28. Isariyawongse BK, Kattan MW. Prediction tools in surgical oncology. Surg Oncol Clin North Am 2012; 21:439–447. 29. Kattan MW, Marasco J. What is a real nomogram? Semin Oncol 2010; 37:23–26 30. Lei J, Yan L. Comparison between living donor liver transplantation recipients who met the Milan and UCSF criteria after successful downstaging therapies. J Gastrointest Surg 2012; 16:2120–2125. 31. Yao FY, Kerlan RK Jr, Hirose R, et al. Excellent outcome following down-staging of hepatocellular carcinoma prior to liver transplantation: an intention-to-treat analysis. Hepatology 2008; 48:819–827. 32. Otto G, Schuchmann M, Hoppe-Lotichius M, et al. How to decide about liver transplantation in patients with hepatocellular carcinoma: size and number of lesions or response to TACE?. J Hepatol 2013; 59:279–284. 33. Cucchetti A, Piscaglia F, Grigioni AD, et al. Preoperative prediction of hepatocellular carcinoma tumour grade and micro-vascular invasion by means of artificial neural network: a pilot study. J Hepatol 2010; 52: 880–888. 34. Hsu CY, Huang YH, Hsia CY, et al. A new prognostic model for hepatocellular carcinoma based on total tumor volume: the Taipei Integrated Scoring System. J Hepatol 2010; 53:108–117. 35. Han K, Kim JH, Yoon HM, et al. Transcatheter arterial chemoembolization for infiltrative hepatocellular carcinoma: clinical safety and efficacy and factors influencing patient survival. Korean J Radiol 2014; 15:464–471.
Volume 26
’
Number 8
’
August
’
2015
36. Kang SH, Kim do Y, Jeon SM, et al. Clinical characteristics and prognosis of hepatocellular carcinoma with different sets of serum AFP and PIVKA-II levels. Eur J Gastroenterol Hepatol 2012; 24: 849–856. 37. Sawhney S, Montano-Loza AJ, Salat P, et al. Transarterial chemoembolization in patients with hepatocellular carcinoma: predictors of survival. Can J Gastroenterol 2011; 25:426–432. 38. Bruix J, Sala M, Llovet JM. Chemoembolization for hepatocellular carcinoma. Gastroenterology 2004; 127:S179–S188. 39. Chung GE, Lee JH, Kim HY, et al. Transarterial chemoembolization can be safely performed in patients with hepatocellular carcinoma invading the main portal vein and may improve the overall survival. Radiology 2011; 258:627–634. 40. Georgiades CS, Hong K, D’Angelo M, Geschwind JF. Safety and efficacy of transarterial chemoembolization in patients with unresectable hepatocellular carcinoma and portal vein thrombosis. J Vasc Interv Radiol 2005; 16:1653–1659.
1101
41. Kim KM, Kim JH, Park IS, et al. Reappraisal of repeated transarterial chemoembolization in the treatment of hepatocellular carcinoma with portal vein invasion. J Gastroenterol Hepatol 2009; 24:806–814. 42. Minagawa M, Makuuchi M. Treatment of hepatocellular carcinoma accompanied by portal vein tumor thrombus. World J Gastroenterol 2006; 12:7561–7567. 43. Pinter M, Hucke F, Graziadei I, et al. Advanced-stage hepatocellular carcinoma: transarterial chemoembolization versus sorafenib. Radiology 2012; 263:590–599. 44. Lammer J, Malagari K, Vogl T, et al. Prospective randomized study of doxorubicin-eluting-bead embolization in the treatment of hepatocellular carcinoma: results of the PRECISION V study. Cardiovasc Intervent Radiol 2010; 33:41–52. 45. Hsu CY, Lee YH, Hsia CY, et al. Performance status in patients with hepatocellular carcinoma: determinants, prognostic impact, and ability to improve the Barcelona Clinic Liver Cancer system. Hepatology 2013; 57: 112–119.
1101.e1
’
Nomograms for Postchemoembolization Outcomes
APPENDIX A. TRANSARTERIAL CHEMOEMBOLIZATION PROCEDURE Transarterial chemoembolization was conducted by experienced interventional radiologists under local anesthesia with right femoral access. Angiography of the celiac trunk and superior mesenteric artery was initially performed in all patients to assess the anatomy, tumor burden, and patency of the portal vein. After angiography, cisplatin (2 mg/kg cisplatin in distilled normal saline) was administered for 15 minutes into the right lobar, left lobar, or proper hepatic artery according to the location of the tumor. An emulsion of cisplatin in ethiodized oil (Lipiodol; Laboratoire Guerbet, AulnaySous-Bois, France) in a 1:1 ratio with the adjusted concentration of 20 ng/mL was delivered next, followed by embolization with 1-mm diameter absorbable gelatin sponge particles (Gelfoam; Ethicon, Inc, Somerville, New Jersey) in a selective or, if possible, superselective manner until arterial flow stasis was achieved. The dose of ethiodized oil depended on the tumor size. The cisplatin injection was preceded by intravenous hydration and antiemetics and followed by further intravenous hydration. Dynamic contrast-enhanced computed tomography of the liver was always performed 4–6 weeks after transarterial chemoembolization to assess the effect of each procedure. Transarterial chemoembolization was repeated every 6–8 week if there were residual tumors, as long as new metastasis had not developed and hepatic function was preserved.
Statistical Analysis We used a Cox proportional hazards model to develop the prediction model for estimating 5-year OS and 3-year PFS. The covariates age, sex, calculated tumor volume, tumor number, tumor type (nodular or infiltrative), bilobar involvement, Child-Pugh class, MELD score, AFP value, presence of HBV, presence of cirrhosis, hepatic vein invasion, and portal vein invasion were selected from a survey of the literature and considered as candidate prognostic variables. There were two missing
Ha et al
’
JVIR
values for AFP in the derivation cohort, and the data were replaced by a median value of the overall cohort (99.9 ng/mL). A model-building approach was used in which all the above-mentioned candidate covariates were at first included. A random forest model was used for selecting variables (19). The major advantage of this model is that it can produce an unbiased measure of the importance of a variable even in mixtures of categorical and continuous variables (20). Following this, clinically or statistically less meaningful variables, such as variables with contradictory coefficient signs, were excluded one by one. To avoid a multicollinearity problem among the variables finally selected, we computed a variation inflation factor and assessed multicollinearity (variance inflation factor Z 5). The proportional hazards assumption also was verified using the scaled Schoenfeld residual test. The overall goodness-of-fit of the model was measured using the Cramér-von-Mises test based on cumulative residuals. The Cramér-von-Mises test can be used to examine the linearity assumption of predictors in a multivariate model (21). Based on the multiple fractional polynomial model (22), we employed the log-transformed AFP variable and the transformed MELD variable (MELD score/10)2. Discrimination and calibration powers were examined according to Cindexes and the calibration curves for the derivation and validation cohorts. A Hosmer-Lemeshow–type test was used to assess calibration ability, and the complete-case method was used to handle missing data. To generate user-friendly nomograms, we used the rms package developed by Harrell in R software version 2.15.1 (23). The 1-, 3-, and 5-year OS probabilities and 1-, 3-, and 5-year PFS probabilities also are provided. Points were allocated for each level of prognostic factor according to the scale shown in the nomogram. The total score was determined by summing individual points, and these were used to calculate t-year survival probability and progression-free probability (t ¼ 1, 3 or 5). The function of cph in the rms package produces biascorrected performance measures using the bootstrap resampling method. All analyses were based on a twosided test at a significance level of .05.
