A differential risk assessment and decision model for Transarterial chemoembolization in hepatocellular carcinoma based on hepatic function

In this multi-center cohort study, eligible HCC patients were evaluated for eligibility from three large-volume university hospitals in South Korea (i.e., Seoul National University Hospital [SNUH], Ewha Womans University Medical Center [EUMC], and Samsung Medical Center [SMC]). We included consecutive patients meeting all the following inclusion criteria: 1) the patients with unresectable HCC of BCLC A or B; 2) the patients who received conventional TACE as their initial treatment; 3) the patients aged ≥18 years. The diagnosis of HCC was based on histological examination or clinicoradiological criteria according to international guidelines [2, 3]. Among a total of 763 consecutive HCC patients from SNUH (between January 2012 and April 2014, n = 542) and EUMC (between January 2011 and December 2015, n = 221) who were considered eligible, 166 patients (n = 121 in SNUH and n = 45 in EUMC) were excluded because of the following reasons: poor performance state (ECOG≥1, n = 39); Child-Pugh class B9 or C (n = 44), current or previous features of decompensation (i.e., uncontrolled ascites, variceal hemorrhage, or hepatic encephalopathy; n = 60), and follow-up loss (n = 23). Among these patients, Thus, 597 patients were enrolled for analysis, and were randomly assigned to either the derivation (n = 419) or the internal validation set (n = 178) stratified by age and sex at a 7:3 ratio (Fig. 1). For an independent cohort for external validation, 750 HCC patients who met the abovementioned criteria were evaluated for eligibility from SMC (between January 2007 and December 2012). Of these, 11 patients were excluded because of history of decompensation or Child-Pugh score of 9 or higher. Finally, 739 patients from SMC were enrolled for external validation.

The present study was approved by the institutional review boards of the three participating institutions and was conducted following the ethical guidelines of the World Medical Association Declaration of Helsinki. Informed patient consent was waived by the institutional review board of each institution owing to the retrospective nature of the study.

TACE procedures were performed using the superselective method by experienced interventional radiologists (over 10 years of experience) in study hospitals [17]. The principles of TACE procedures were largely similar between three institutions in terms of superselectivity, choice of chemotherapeutic agents and treatment schedule as described elsewhere [18, 19]. Briefly, an arterial catheter was inserted into the femoral artery using the Seldinger technique and elective angiography of the celiac axis was performed. Then the catheter was advanced into the desired hepatic artery branch. Tumor-feeding vessels were superselected whenever possible, and a suspension containing 20–60 mg of doxorubicin hydrochloride (ADM, Dong-A Pharmacy, Seoul, Korea) and 2–20 mL of iodized oil (Lipiodol, Guerbet, Aulnay-sous-Bois, France) with absorbable gelatin sponge particles (Gelfoam, Upjohn, Kalamazoo, MI) was infused through a catheter (5-Fr) or a microcatheter (2.8- or 3-Fr) placed in the tumor-feeding arteries. The dosages of doxorubicin and iodized oil and the use of gelatin sponge particles were determined for each patient based on tumor burden, tumor characteristics, and hepatic functional reserve [20, 21]. Repeated TACE treatments were considered if residual or newly developed tumors were detected on dynamic computed tomography (CT) or magnetic resonance imaging (MRI) undertaken 4–6 weeks following each TACE session and were performed on an “on-demand” basis depending on individual tumor response and hepatic functional reserve.

The primary outcome was overall survival (OS), which was measured from the date of HCC diagnosis to the date of death from any cause. Survival data of the enrolled patients were obtained from the national statistical data provided by the Korean Ministry of Government Administration and Home Affairs. The data cut-off date was November 30, 2017. The secondary outcome assessed was tumor response. Tumor responses were evaluated after every TACE session with the modified Response Evaluation Criteria in Solid Tumors criteria [22]. The efficacy of the performed TACE was determined by evaluating the pattern of iodized oil retention in the target lesions as an indicator of tumor necrosis [23]. Iodized oil retention was considered as compact on imaging when the contrast medium was well scattered throughout all viable target lesions; otherwise, it was regarded as non-compact uptake [24]. Patients without residual viable tumor after TACE were followed-up with dynamic CT or MRI every 8–12 weeks. All scans were reviewed by two independent radiologists with > 10 years of experience who were unaware of the group assignment. In cases of discordance, an additional third independent experienced radiologist reviewed the images and a consensus was reached among the three.

Baseline characteristics were presented as mean ± standard deviation for normally distributed continuous variables and median with interquartile ranges (IQRs) for continuous variables with a skewed distribution. Discrete variables were summarized by the number of subjects with percentages. To compare baseline characteristics between groups, we used the Student’s t-test or Mann-Whitney U test, as appropriate. Distribution of categorical variables was compared using the chi-square and Fisher’s exact test. OS was calculated as the time from HCC diagnosis until death from any cause. Survival analysis was performed using the Kaplan-Meier analysis and the log-rank test was used to compare between groups. Tumor- and hepatic function-related prognostic factors for OS were explored using the Cox proportional-hazards regression analysis. Based on the results of the univariate analyses, factors with a significant difference (P < 0.05) were included in the multivariate model for development of a prediction model. A prediction model for OS was developed using relevant parameters identified by forward stepwise selection. A model with the minimum Akaike information criterion value was selected, which rewarded the goodness-of-fit of the model. Basically, we performed a proportional-hazards hypothesis for the selected models. The proportional-hazards hypothesis was checked via the Schoenfeld residuals method. The predictive ability was evaluated by using concordance (c)-statistics for discrimination function and Hosmer-Lemeshow test for calibration function. These were performed after excluding the variables sequentially to develop the simplest model out of candidate models developed. When the result of the Hosmer-Lemeshow test was satisfied (P > 0.05) and the c-statistics was 0.7 or more, the difference in c-statistics between the models was compared and the simplest model was selected if there was no statistically significant difference. To obtain the optimism corrected value of the c-statistics, we applied 100-times bootstraps. In the final model, a relative score of each risk factor was assigned based on the estimated coefficient value, and the risk score for prediction of OS was calculated in each subject. Based on the results of the pairwise log-rank test, the patients were divided into subgroups by the OS score and survival curves were compared between the subgroups. Internal and external validation of the developed prediction model was conducted, and sensitivity analyses were performed under various conditions (e.g., among patients with Child-Pugh class A or B). A two-sided P < 0.05 was considered statistically significant. All statistical analysis was performed using SAS ver. 9.4 software (SAS Institute, Cary, NC).

Table 1 summarizes the baseline characteristics of the overall included patients. The median age was 63 in SNUH, 69 in EUMC, and 63 years in SMC. The proportion of male was 79.1% in SNUH, 80.1% in EUMC, and 78.9% in SMC. The etiology of the underlying liver disease was mostly viral. The number tumors over 3 were found in 72 (17.1%) patients of SNUH, 32 (18.2%) of EUMC, and 120 (16.2%) of SMC. The maximal tumor diameter over 5 cm was found in 113 (26.8%) of SNUH, 57 (32.4%) of EUMC, and 272 (36.8%) of SMC. The median alpha-fetoprotein (AFP) level ≥ 200 ng/mL was found in 89 (21.1%) of SNUH, 49 (28.5%) of EUMC, and 194 (26.3%) of SMC. The patient with Child-Pugh class B was 85 (20.2%) in SMUH, 33 (18.8%) in EUMC, and 101 (13.7%) in SMC. The HAP score D were found in 24 patients (5.7%) in SNUH, 23 (13.1%) in EUMC, and 50 (6.8%) in SMC, respectively. ALBI grades 3 were 17 patients (4.0%) in SNUH, 13 (7.4%) in EUMC, and 15 (2.0%) in SMC (Table 1). The time intervals between the diagnosis of HCC and the initial TACE were 2.0 weeks (1.8–3.0) in SNU cohort and 2.0 weeks (1.0–2.8) in EUMC cohort (P = 0.115), suggesting that there was no significant time delay between the diagnosis and the initial treatment in both cohorts.

There were no significant differences between the two hospital cohorts (SNUH and EUMC) with respect to sex, etiology, Child-Pugh class, tumor number, tumor size, and AFP. However, the patients in the EUMC cohort were significantly older (69 [59–78] vs. 63 [54–71] years) and had higher ALBI grades (grade 2: 67.0% vs. 58.2%), HAP scores (score D: 13.1% vs. 5.7%), and modified HAP scores (score D: 5.1% vs. 1.2%) than those in the SNUH cohort (Supplementary Table 1). Thus, instead of using each hospital cohort as either the derivation or the internal validation set, the two hospital cohorts were divided into two groups in a proportion of 7:3, which were stratified for age and sex to minimize the influence of discrepancies between the two hospital cohorts. Following the division into two sets, the discrepancies in baseline characteristics largely diminished between the derivation and internal validation sets, as shown in Table 2. Baseline characteristics of patients from SMC were also shown in Table 1. The patients of SMC were analyzed as an external validation set.

The Cox proportional hazards analyses were performed in the derivation set. Univariate Cox analysis was performed to select prognostic factors associated with OS, and the following variables were significantly associated with OS: Child-Pugh class B, AFP ≥200 ng/mL, ALBI grade 2, tumor number ≥ 3, maximal tumor diameter ≥ 5 cm, initial TACE response, and HAP score D. Multivariate analysis identified four independent predictors for OS as follows: ALBI grade (adjusted hazard ratio [aHR], 2.299; 95% CI, 1.694–3.119; P < 0.001), maximal tumor diameter (aHR, 1.704; 95% CI, 1.424–2.038; P < 0.001), AFP (aHR, 1.583; 95% CI, 1.164–2.154; P = 0.003), and initial TACE response (aHR, 1.915; 95% CI, 1.431–2.563; P < 0.001) (Table 3).

Through forward stepwise selection, a scoring system was developed including the aforementioned four independent predictors for OS (i.e., ALBI, tumor size, AFP, initial TACE response; abbreviated as “ASAR”) as shown in Table 4. Regression coefficients of the factors were 0.832 (ALBI grade), 0.533 (maximal tumor diameter), 0.460 (AFP), and 0.650 (initial TACE response). The weighted scores (0, 1, and 2) for the ASAR scoring system were assigned for the four covariates based on the regression coefficients that were obtained from the final analysis and the total scores ranged from 0 to 6. We confirmed the simplest model according to the result of the Hosmer-Lemeshow test and c-statistics between models with the abovementioned statistical methodology. C-index for OS was 0.733 (95% CI, 0.570–0.871) in the derivation set, which was maintained at 0.733 (95% CI, 0.703–0.768) with 100-fold bootstrapping. In the internal validation, the c-index was 0.700 (95% CI, 0.445–0.905). Goodness-of-fit for the ASAR model was confirmed in both the derivation and the internal validation sets (P = 0.360 and P = 0.926, respectively, by Hosmer-Lemeshow χ² test). In the external validation set, c-index was maintained at 0.680 (95% CI, 0.652–0.707).

Based on the pairwise log-rank test results, the cut-off value was designated as 4, which presented the greatest difference in the survival curves of the derivation set. In the derivation set, patients with ASAR < 4 had significantly longer OS than ASAR ≥4 (hazard ratio [HR], 0.233; 95% CI, 0.171–0.317; P < 0.001; Fig. 2a). In the internal validation set, patients with ASAR < 4 had significantly longer OS than ASAR ≥4 (HR, 0.287; 95% CI, 0.176–0.468; P < 0.001; Fig. 2b). Thus, we designated ASAR < 4 as the low-risk group and ASAR ≥4 as the high-risk group. Among 178 patients of the internal validation set, the proportion of patients who were not amenable to further treatment after initial TACE was higher in the high-risk group (4 out of 29, 13.8%) than in the low-risk group (5 out of 149, 3.4%) (P = 0.019). Of these nine patients, six patients died after first TACE and three patients received best supportive care. We also compared ASAR with ART score for the decision of TACE repetition, and ASAR showed better performance than ART score in the internal validation set (Supplementary Table 2). In the external validation set, patients with ASAR < 4 had also significantly longer OS than ASAR ≥4 (HR, 0.237; 95% CI, 0.189–0.298; P < 0.001; Fig. 2c).

In the combined derivation and internal validation cohorts (SNUH and EUMC), the low-risk group showed significantly longer OS than that shown by the high-risk group (HR, 0.249; 95% CI, 0.192–0.324; P < 0.001; Fig. 3a). The median OS was 70.2 months in the low-risk group vs. 17.7 months in the high-risk group. The 1-year and 3-year survival rates were 92.1 and 69.7%, respectively, in the low-risk group and 52.4 and 23.7%, respectively, in the high-risk group. Among a subgroup of patients with Child-Pugh class A, the low-risk group showed significantly longer OS than the high-risk group (HR, 0.260; 95% CI, 0.191–0.354; P < 0.001; Fig. 3b). The 1-year and 3-year survival rates were 93.4 and 72.4% in the low-risk group, and 59.4 and 27.0% in the high-risk group, respectively. Among the subgroup of patients with Child-Pugh class B, the low-risk group showed significantly longer OS than that shown by the high-risk group (HR, 0.252; 95% CI, 0.151–0.420; P < 0.001; Fig. 3c). The 1-year and 3-year survival rates were 85.2 and 53.4% in the low-risk group, and 46.3 and 6.6% in the high-risk group, respectively.

Survivals from the subanalysis for BCLC-stage B patients were also comparable to those of the entire patients. The low-risk group showed significantly longer OS than the high-risk group both in the derivation (HR, 0.293; 95% CI, 0.190–0.452; P < 0.001; Supplementary Fig. 1a) and internal validation sets (HR, 0.261; 95% CI, 0.132–0.516; P < 0.001; Supplementary Fig. 1b). The 1-year and 3-year survival rates of derivation set were 88.6 and 64.1% in the low-risk group, and 46.3 and 6.6% in the high-risk group, respectively. The 1-year and 3-year survival rates of validation set were 86.7 and 65.3%, in the low-risk group, and 67.3 and 14.2% in the high-risk group, respectively.

Because the initiation of first TACE is an important issue for patients with Child-Pugh class B, we additionally assessed the predictive performance for those patients using a modified version of the ASAR scoring system, which comprised of only baseline components (ALBI, size, and AFP) excluding initial TACE response, i.e., “ASA(R)”. C-index values were 0.788 (95% CI, 0.702–0.876) in the derivation set, 0.745 (95% CI, 0.646–0.862) in the internal validation set, and 0.670 (95% CI, 0.605–0.725) in the external validation set, respectively. Using the cut-off score of 4 for this modified ASA(R) in Child-Pugh class B patients, OS was also significantly different among the derivation set (ASA(R) < 4 vs. ≥4: HR, 0.317; 95% CI, 0.150–0.669; P < 0.001; Fig. 4a), the internal validation set (ASA(R) < 4 vs. ≥4: HR, 0.225; 95% CI, 0.088–0.575; P < 0.001; Fig. 4b), and the external validation set (ASA(R) < 4 vs. ≥4: HR, 0.261; 95% CI, 0.201–0.339; P < 0.001; Fig. 4c). ASA(R) showed significantly better performance than other models for evaluating initial TACE suitability, such as HAP, modified HAP, and modified HAP II models in the internal validation set (Supplementary Table 2).