The selection procedure and study design are shown in Figure 1 . A total of 682 HCC patients who underwent liver resection in the First Affiliated Hospital of Xi’an Jiaotong University between January 2015 and December 2019 were included in this study. The inclusion criteria are listed as follows (1): patients with an accurate pathological diagnosis of HCC (2); patients who underwent radical hepatic resection and did not receive any preoperative antitumor therapies, such as trans-arterial chemoembolization (TACE), radiofrequency ablation, and antineoplastic drugs (3); patients with complete clinicopathologic, laboratory, imaging, and follow-up data. The exclusion criteria are listed as follows (1): the presence of perioperative death (2); the presence of other malignant tumors in the patient’s medical history (3); the presence of signs of macrovascular invasion before hepatectomy, such as portal vein invasion or hepatic vein invasion (4); patients who died without HCC recurrence within 2 years after surgery (5); the presence of extrahepatic metastasis within 2 years after surgery (6); patients who lost to the follow-up after surgery.

Finally, 481 patients were included. The included patients were assigned to the following two groups at a ratio of 7:3: the training cohort (n = 337) used to construct the model and the validation cohort (n = 144) used to assess the performance and verify the reliability of the model. (The randomized code was generated by the Random Number Generators of the SPSS statistical software). This retrospective study used anonymous data and was approved by the First Affiliated Hospital of Xi’an Jiaotong University.

Demographic data of patients included the patient’s age and sex. All patients received routine laboratory examinations before curative resection, such as blood tests, biochemical examinations, and tumor marker examinations. The research parameters in this study included serum AFP, aspartate aminotransferase (AST), alanine aminotransaminase (ALT), bilirubin, albumin, GGT, ALP, the count of neutrophils, platelets, and lymphocytes.

All pathological specimens were independently reassessed by two pathologists. The pathological sampling process was carried out in agreement with the practice guidelines for handling surgical specimens (23), and the sampling process included the sites, distances, and volumes for tissue sampling from HCC surgical specimens to make more accurate pathological diagnoses. Pathological features for the tumor included: tumor size, tumor differentiation, MVI grade, microscopic portal vein thrombus, intratumor necrosis, intratumor hemorrhage, bile duct invasion, capsular invasion, satellite nodules, and surgical margin. The typical pathological characteristics of HCC patients after surgery were shown in Figures S1A–F .

All patients received an upper abdominal contrast-enhanced CT scan preoperatively. Arterial-phase CT scans were obtained at 30 s, portal-phase CT scans were obtained at 60 s, and late-phase CT scans were obtained at 90–120 s. The scanning range and imaging parameters as described in our previous study (21). Peritumoral enhancement, peritumoral boundary, tumor shape, intratumoral artery, and multiple tumors in contrast-enhanced CT images were recorded and factored into the analysis. The typical imaging characteristics of HCC patients before surgery were shown in Figures S1G–I . The CT images were analyzed by two senior radiologists. In the case of inconsistency in assessment, they reached a consensus through reviewing and discussion, and the final analysis was based on their consensus.

Early HCC recurrence was defined as tumor recurrence within 2 years after liver resection. After tumor resection, all patients received regular follow-up according to clinical guidelines (24). Follow-up consisted of laboratory tests (e.g. serum AFP levels, liver function tests) and imaging examinations (ultrasound, CT, MRI) was performed every three months after surgery. A diagnosis of HCC recurrence was based on two or more examination methods including ultrasound, contrast-enhanced CT, MRI, or hepatic artery angiography with or without a raised serum AFP level. The typical imaging features of the recurrent HCC patient were shown in Figure S2 . For patients without any evidence of recurrence, the last follow-up date was December 31, 2019.

Whole blood samples from all enrolled patients were collected before hepatectomy and centrifuged at 3000 rpm for 10 min, and serum samples were immediately stored at -80°C until use. Serum concentrations of VEGF-A were measured with a VEGF ELISA Kit (R & D Systems China Co. Ltd., Shanghai, China), according to the manufacturer’s protocol. All ELISA assays were performed in duplicates.

Statistical analysis were performed with SPSS 27.0 software (IBM, New York, USA) and R software (version 3.4.4, R Project for Statistical Computing, http://www.r-project.org). Continuous variables were presented as means ± standard deviations (SD). Categorical variables were presented as percentages. The independent-sample t test was used for continuous variables. The Chi-squared test or Fisher’s exact test was used for categorical variables. X-tile software was applied to establish the optimal cutoff value of VEGF-A (25). The RFS rates within two years were calculated using the Kaplan–Meier method, and the log-rank test was used to compare differences among different groups. Univariate and multivariate analysis with Cox proportional hazards regression were used to determine the independent risk factors related to early recurrence in HCC patients in the training cohort. Variables that reached statistical significance (P < 0.05) in the univariate analysis were considered for the multivariate Cox regression model.

The nomogram for predicting 6-, 12-, and 24-month RFS was established based on proportionally converting each regression coefficient in multivariate Cox proportional hazards regression to a 0- to 100-point scale. The effect of the variable with the highest β coefficient (absolute value) was assigned 100 points. The points are added across independent variables to derive total points, which are converted to predicted probabilities. In addition, the best cut-off value for the total points was obtained using the X-tile software. Then, the enrolled patients were divided into low-, middle-, and high-risk subgroups to create a risk classification system, stratifying the recurrence risk of all HCC patients after tumor resection. The difference in RFS between the three subgroups was obtained using the Kaplan-Meier method. To quantify the predictive accuracy of the proposed nomogram, Harrell’s concordance index (C-index) was evaluated. The receiver operating characteristic (ROC) curves were constructed to estimate the performance of the model in the training cohort and internal validation cohort. All statistical tests were two-sided, and P < 0.05 indicated statistical difference.

A total of 682 HCC patients underwent liver resection in our center between January 2015 and December 2019. Finally, 481 patients were enrolled in this study according to the inclusion and exclusion criteria ( Figure 1 ). The enrolled HCC patients were assigned to the following two groups at a ratio of 7:3: the training cohort (n = 337) and the validation cohort (n = 144).

The demographic, clinical, pathological, and imaging characteristics of HCC patients in the training cohort and validation cohort are shown in Table 1 . The patients included 398 men and 83 women with an average age of 54.9 years, ranging from 15 to 79 years. Based on the follow-up data and diagnostic criteria, 182 (37.8%) patients experienced early tumor recurrence, while 299 (62.2%) patients had no recurrence evidence at the time of the last follow-up. 274 (54.9%) of patients were pathologically diagnosed as MVI-positive. 64% of HCC patients presented low AFP levels (AFP < 400 ng/mL). The mean serum VEGF-A concentrations in the training cohort and the validation cohort were 164.93 ± 100.66 pg/mL and 162.92 ± 100.57 pg/mL, respectively. There were no significant differences in the characteristics between the training and validation cohorts ( Table 1 ).

The early recurrence patterns were assessed from three aspects (1): type of recurrent tumors: local recurrence at the surgical margins (recurrent tumors located within 20 mm of surgical margin), distant intrahepatic recurrence to the surgical margins (recurrent tumors located more than 20 mm of surgical margin) or both (2); number of recurrent tumors: solitary, two or ≥ 3 (3); time to recurrence (TTR) after surgery. The patterns of recurrences were shown in Supplementary Table 1 . Among the 182 patients who developed tumor recurrence, 148 patients (81.32%) shown tumors distant to the surgical margin, 23 (12.64%) patients had recurrent tumors located within 20 mm of the surgical margin, the rest 11 patients had both distant and near margin recurrence ( Figure 2A ). 114 (62.64%) patients had solitary nodule. Whereas, 68 (37.36%) patients developed multinodular (2 or ≥3) recurrence ( Figure 2B ). Group differences in time to recurrence were examined with the log-rank test. There were no significant differences in TTR related to recurrence patterns ( Figures 2C, D ).

In that, VEGF-A was associated with MVI as shown in previous studies (21), we assessed VEGF-A expression in the present study. The patients in the training cohort were classified into the low-concentrations group (VEGF-A < 127.8 pg/mL, n = 181), middle-concentrations group (VEGF-A among 127.8 to 240.3 pg/mL, n = 145) and high-concentrations group (VEGF-A > 240.3 pg/mL, n = 63). The optimal cut-off values of VEGF-A for predicting RFS were determined by the X-tile bio-informatics software ( Figure 3A ). Log-rank test shown significant differences of RFS between low-, middle-, and high-concentrations groups (P = 0.001, Figure 3B ). For MVI subgroups, the two-year cumulative recurrence rate in M0, M1, and M2 groups was 19.59%, 42.86%, and 58.1%, respectively (P < 0.001, Figure 3C ). In order to determine the predictive ability of MVI and VEGF-A for early recurrence, the ROC curve was drawn by the survival ROC package of the R software. The areas under the ROC curve of MVI and VEGF-A in the training cohort were 0.670 (95%CI: 0.612-0.728) and 0.692 (95%CI: 0.634-0.750), respectively ( Figure 3D ). The result indicated MVI and VEGF-A had the good predictive ability for HCC recurrence.

Univariate and multivariate Cox proportional hazards regression of risk factors for RFS were shown in Table 2 . Factors, such as AFP ≥ 400 ng/mL, GGT > 60 U/L, VEGF-A among 127.8 to 240.3 pg/mL, VEGF-A > 240.3 pg/mL, peritumoral enhancement, unclear peritumoral boundary, irregular tumor shape, tumor size > 50 mm, MVI-positive, microscopic portal vein thrombus, intratumor necrosis, capsular invasion, surgical margin among 5.0 to 10.0 mm and surgical margin < 5.0 mm were found to be the possible predictors for recurrence of HCC by univariate analysis. By multivariable Cox regression analysis, AFP ≥ 400 ng/mL (P = 0.008), VEGF-A among 127.8 to 240.3 pg/mL (P = 0.012), VEGF-A > 240.3 pg/mL (P < 0.001), M1 subgroup of MVI (P = 0.002), M2 subgroup of MVI (P < 0.001), intratumor necrosis (P = 0.011), surgical margin among 5.0 to 10.0 mm (P = 0.043) and surgical margin < 5.0 mm (P = 0.012) were determined to be the independent risk factors associated with early recurrence.

A Forest plot was drawn to show the effect of different risk factors on patient RFS ( Figure 4 ). M2 subgroup of MVI had the strongest predictive power for early recurrence (HR: 3.120; 95%CI: 1.977-4.925).

The nomogram was based on proportionally converting each regression coefficient in multivariate Cox regression analysis to a 0- to 100- point scale. The effect of the variable with the highest β coefficient was assigned 100 points ( Figure 5A ). A nomogram was developed to predict 6−, 12− and 24−month RFS for patients with HCC after radical hepatic resection based on the significant risk factors identified in the Cox regression analysis ( Figure 5B ). The M2 subgroup of MVI had the highest score (100 points). The scores for the other variables were: 44.6 points (serum AFP ≥ 400 ng/mL), 50.7 points (serum VEGF-A among 127.8 to 240.3 pg/mL), 82.3 points (serum VEGF-A > 240.3 pg/mL), 70.1 points (M1 subgroup of MVI), 44.8 points (intratumor necrosis), 41.4 points (surgical margin among 5.0 to 10.0 mm), and 51.1 points (surgical margin of < 5.0 mm).

The C-indexes for RFS prediction were 0.744 (95%CI: 0.701-0.787) and 0.736 (95%CI: 0.672-0.801) for the training and validation cohorts, respectively. To evaluate and compare the discriminatory power of the nomogram model, we plotted the ROC curve and calculated the area under the ROC curve (AUC) with 95% confidence intervals (CI) in the training cohort and validation cohort, respectively. The area under the ROC curve of the established nomogram in the training cohort and validation cohort was 0.781 (95% CI: 0.729-0.832) and 0.808 (95% CI: 0.731-0.886), respectively ( Figure 6A ). At the individual level, the risk scores were calculated for each patient based on the developed risk model. According to the risk scores of patients, we acquired two optimal cutoff points using the X-tile software: 156.1 and 210.2 ( Figure 6B ). Then, we obtained the RFS rates of low-, middle-, and high-risk groups were 80.71% (159 of 197), 48.05% (37 of 77), and 23.81% (15 of 63), respectively. Intergroup comparison shown a statistically significant difference in 2-year RFS using Kaplan–Meier survival analysis in the training cohort (P < 0.001, Figure 6C ). Likewise, consistent findings were also presented in the validation cohort (P < 0.001, Figure 6D ).