Tree-based classification system incorporating the HVTT-PVTT score for personalized management of hepatocellular carcinoma patients with macroscopic vascular invasion

Baseline characteristics

In total, 869 eligible HCC patients with vascular invasion were included in this study. The median age was 49 years (range, 13-80 years). All patients had macroscopic vascular tumor invasion, including PVTT (82%), HVTT (7.1%), or both (10.8%). Approximately one-third (273/869) of the patients had extrahepatic metastases. In terms of the initial therapy reported for the cases included in this study, the first-line treatment strategies included transarterial chemoembolization (TACE) (58.6%), liver resection (17.6%), sorafenib (8.1%), supportive care (14.8%), and radiotherapy (0.9%). During follow-up, 642 (73.9%) patients died. The median OS was 7.0 months (95% confidence interval (CI), 6.37-7.62 months). The median follow-up time was 51.67 months (95 % CI, 44.06-57.11 months). The 869 patients were randomly divided into the training set (426 patients) and validation set (443 patients). The baseline characteristics of the training and validation sets are summarized in Table 1. In summary, the patients in the training and validation sets shared similar characteristics.

Development of HVTT-PVTT scoring system

The training set was used to build the prognostic classification system. In the training set, patients who had both PVTT and HVTT had significantly poorer outcome than others who had only one type of tumor thrombus (4.63 vs. 6.87 months, P = 0.025). We hope to build a new classification including both HVTT and PVTT. Therefore, thirteen groups with different types of vascular invasion were generated (Table 2). To separate patients with different prognosis and surgical proportions, we ranked the thirteen groups according to their surgical proportions and listed their median OS in Figure 1. The top four subgroups had significantly larger surgical proportions and better prognosis than others. In addition, according to the regression coefficients (B) of PVTT and HVTT in the Cox regression model and clinical judgement, different points were given and an equation was built as follows. The HVTT-PVTT score = PVTT (none = 0, segmental branches of portal vein or above = 1, right/left portal vein = 2, and main portal vein or superior mesenteric vein = 3) + HVTT (none = 0, hepatic vein = 1, and inferior vena cava or right atrium = 3). Because inferior vena cava and right atrium tumor thrombus were also extrahepatic vascular invasion like main trunk PVTT, they were also given three points in this equation. We found that the top four groups all had scores < 3 points and other groups had scores ≥ 3 points. Therefore, we classified patients into vascular invasion stage I (HVTT-PVTT < 3 points) and II (HVTT-PVTT ≥ 3 points) with different survival outcomes (7.4 vs 4.6 months, P < 0.001) and surgical proportions (24.4% vs 3.6%, P < 0.001) by using HVTT-PVTT score. Patients with vascular invasion stage I (HVTT-PVTT < 3 points) are more likely to be candidates for surgery and have better survival outcomes than patients with vascular invasion stage II (HVTT-PVTT ≥ 3 points).

Development of a visualized tree-based classification system

Using the Cox regression model, Child-Pugh (stage B vs. stage A) (hazard ratio (HR) with 95% CI=1.383 (1.077-1.777), P < 0.05), Tumor burden (high vs. low) (HR with 95% CI= 1.264 (1.005-1.590), P < 0.05), Extrahepatic metastases (presence vs. absence) (HR with 95% CI= 1.38 (1.085-1.755), P < 0.05), and HVTT-PVTT scores (≥ 3 vs. < 3) (HR with 95% CI= 1.434 (1.130-1.819), P < 0.05) were identified as independent prognostic factors of the survival outcome (Table 3). Based on these four prognostic factors and the survival data, the CART algorithm automatically separated patients into subgroups with different prognosis. Subgroups with a similar median OS were merged into larger groups according to clinical judgement to generate the final classification system (Figure 2). Finally, the patients in the training set were classified into three stages. The median OS for these stages was 10.3, 6.1, and 3.3 months, the 1-year survival rates were 48%, 31%, and 18%, and the 3-year survival rates were 22%, 13%, and 9%, respectively (Table 4).

Performance assessment

After the staging system was built, the validation set was used to assess its performance. The proposed classification scheme could significantly discriminate patients with different prognosis in the validation set (Table 4, Figure 3). The median OS for the different stages was 12.3, 6.9, and 5.6 months, the 1-year survival rates were 52%, 37%, and 28%, and the 3-year survival rates were 24%, 15%, and 5%, respectively. In the training set, the C-index values for this model, Cheng’s portal vein staging system [10], and the Hong Kong Liver Cancer (HKLC) system [17] were 0.63, 0.57, and 0.58, respectively. In the validation set, the C-index values for this model, Cheng’s portal vein staging system, and the HKLC system were 0.60, 0.58, and 0.58, respectively. This prognostic model had better discriminatory ability than the previous portal vein staging system (all P values < 0.001) [17].

Treatment guidance

For patients in each subgroup, the first-line treatment strategies included liver resection (48.4%, 10.8%, and 5%), TACE (34.6%, 61.3%, and 67.7%), sorafenib (5.4%, 11.8%, and 19.2%), and supportive care (11.2%, 14.8% and 6.9%). In BCLC-C1, 34.6% (108/312) of the patients received radical surgery; the patients who received liver resection had better survival outcomes than those who received other treatments (median OS: 24.2 months vs 8.5 months; P < 0.001) (Figure 4A). In BCLC-C2, the median OS of patients who received radical surgery, sorafenib, radiotherapy, TACE alone, and supportive treatment was 18.3, 13.7, 17.9, 5.6, and 2.8 months, respectively. No significant differences in survival were observed between patients who underwent liver resection and patients who received sorafenib or radiotherapy (P > 0.05) (Figure 4B). In BCLC-C3, less than 5% of the patients were eligible for liver resection; the median OS of the patients who received sorafenib, TACE alone, and supportive treatment was 6.1, 4.5, and 2.7 months, respectively (P > 0.05) (Figure 4C).

Study design

This study was approved by the institutional review board of the Sun Yat-sen University Cancer Center. Between December 2008 and December 2014, 8599 patients were diagnosed with HCC and treated at Sun Yat-sen University Cancer Center in Guangzhou, including 1228 patients classified into BCLC-C stage. Patients who had no macroscopic vascular invasion or missing medical imaging data, Eastern Cooperative Oncology Group performance status (ECOG PS) ≥ 2, or received treatment previously were excluded. Finally, 869 HCC patients were included into this study. All characteristics were collected, including demographic data, clinical history, baseline laboratory results, Child-Pugh grade, tumor number, tumor size, extent of vascular invasion, extrahepatic metastases, treatment strategies, and survival data.

Diagnosis

HCC was diagnosed based on AASLD criteria either histologically or through a clinical evaluation based on a typical radiologic appearance in dynamic computed tomography (CT) or magnetic resonance imaging (MRI) and an elevated alpha-fetoprotein (AFP) level [25]. Vascular invasion was defined as a portal vein or hepatic vein tumor thrombus, which was distinguished from a thrombus based on the presence of arterial enhancement and venous expansion or the development of a new thrombus directly contiguous with the tumor. The tumor burden was a composite factor of the size of the largest tumor and the number of nodules in the liver. For simplified definitions of the tumor burden, low tumor burden was defined as a solitary nodule or multiple nodules with no nodule larger than 5 cm, and high tumor burden was defined as multiple nodules with sizes larger than 5 cm [17, 23]. Extrahepatic metastases included lymph nodal or distant organs metastases. Metastases to distant organs was assessed by chest X-ray, bone scan, positron emission tomography (PET)-CT, and chest CT and was diagnosed based on radiographic evidence of a mass in distant organs [26]. Nodal metastasis was defined when fulfilled at least two following criteria: (1) the short axis was larger than 10 mm; (2) in the arterial phase the lymph node was observed contrast-enhanced; (3) the size increased during the follow-up [27, 28]. The medical imaging data from all cases were evaluated by one of the two authors (C.F. and S.L.J.). All patients were followed up until death or loss to follow-up. Overall survival (OS) was defined as the time from diagnosis of HCC to death or the date of last follow-up.

Subdivision workup of vascular invasion

All cases included in this study were randomly allocated in a 1:1 ratio to build training and validation sets. Then, 426 and 443 patients were allocated into the training and validation sets, respectively. The PVTT were divided according to Cheng’s PVTT classification system into the following categories: type I, tumor thrombus involve the segmental branches of the portal vein or above; type II, the right/left portal vein; type III, the main portal vein; and type IV the superior mesenteric vein [29]. HVTT was categorized into tumor thrombus in the hepatic veins, the inferior vena cava or right atrium. Additionally, for patients who had both HVTT and PVTT. Groups describing different types of vascular invasion were built by considering both the PVTT and HVTT [14, 16, 30]. The surgery proportion and median OS of each groups were calculated. Because many studies have reported that patients who were candidates for surgery had significantly better prognosis than others [13, 14], to identify patients who underwent surgery and also had favorable prognosis, a new HVTT-PVTT score was built according to the regression coefficients (B) of PVTT and HVTT in the Cox regression model and clinical judgement [31].

Statistical analysis

The characteristics of patients in the training and validation sets were compared. Pearson’s chi-square test and Fisher’s exact test were used to compare categorical variables, and the Mann–Whitney U test was used to compare continuous variables. The Kaplan–Meier method and log-rank test were applied for time-to-event data. Univariate and multivariable analyses were performed using Cox proportional hazards regression. All independent prognostic factors were included in these analyses.

CART algorithm was used to separate the patients into subgroups with similar prognosis using suitable factors. All relevant parameters were imported into the R package “rpart” to automatically separate the patients into different subgroups with different prognosis [32]. Similar subgroups were merged into larger groups manually to simplify and generate the final staging system. The concordance index (C-index) was used to compare the new staging system with Cheng’s and HKLC staging systems, which could both separated these patients into five subgroups [20, 33]. Survival benefits of different treatments in each group were compared using the Kaplan–Meier method and log-rank test [34]. An abstract treatment guideline for the new classification system was proposed. The R (R Core Team, version 3.5.1, Vienna, Austria) and SPSS version 20.0 software (IBM, Armonk, NY) were used for the statistical analysis. P values less than 0.05 were considered statistically significant.