In this study, data from 173 patients diagnosed with cirrhosis between January 2011 and December 2019 were retrospectively analyzed. The patient inclusion criteria were as follows: (1) patients who had recovered from a first episode of EV bleeding and there was no bleeding for at least 5 consecutive days; (2) abdominal computed tomography (CT) scan and HVPG measurement were performed before endoscopic variceal ligation within 2 weeks; and (3) at least 1 year of follow-up after endoscopic therapies. The patient exclusion criteria were as follows: (1) previous therapy including splenectomy, endoscopic variceal ligation, tissue adhesive injection, or usage of non-selective beta blocker to prevent rebleeding; (2) confirmed to have hepatocellular carcinoma based on a histologic examination of the liver; (3) non-sinusoidal portal hypertension (e.g., hepatic cavernoma, Budd-Chiari syndrome); and (4) no contrast-enhanced CT images and HVPG measurement. The recruitment process is shown in Supplementary Figure 1.

The endpoint of the study was EV rebleeding during the 1-year follow-up. EV rebleeding is defined as the occurrence of new esophageal variceal bleeding after a period of 24 h or more from the 24-h point of stable vital signs and hematocrit/hemoglobin following the first episode of EV bleeding (4, 16).

The first episode of EV bleeding was controlled by measures including fluid resuscitation and medication administration (somatostatin and proton pump inhibitors). After recovering from a first episode of EV bleeding, all patients received secondary prevention of EV bleeding according to sixth Baveno Consensus (Baveno VI) (3), namely, the combination of endoscopic variceal ligation and non-selective beta blocker. Moreover, all hepatitis B-related cirrhotic patients received antiviral therapy.

Endoscopic examination was performed by experienced endoscopists. During the examination, the form, location and bleeding signs of varices were noted, and the size of the varices was classified as small, medium or large corresponding to <30, 30–60, or >60%, respectively, of the maximum theoretical size (17).

Contrast-enhanced CT scans were performed using a 320-detector CT scanner (Aquilion ONE, TOSHIBA) and a 64-detector CT scanner (Discovery, GE Healthcare). Non-enhanced CT scans were first acquired, followed by three post-contrast CT scans in three phases: arterial, portal vein and delayed. Arterial phase scanning started ~20–30 s after injection, portal phase scanning was started 30–40 s after the beginning of the arterial phase, and delayed phase scanning was started 40–60 s after the beginning of the portal phase scanning. The following parameters were used: tube voltage, 120 kV; tube current, 150–600 mAs; 80 × 0.5 mm or 64 × 0.625 mm detector collimation, matrix, 512 × 512; slice thickness, 5 mm; and pitch, 1.388 or 0.984. All patients received an intravenous, non-ionic contrast medium (iodine concentration, 370 mg/mL; volume, 1.5–2.0 mL/kg of body weight; Omnipaque 350, GE Healthcare, Shanghai, China) at a rate of 3–5 mL/s. Two imaging-based indexes including diameters of portal vein and spleen vein were assessed.

Regions of interest (ROIs) were drawn around the whole liver and spleen slice-by-slice using 3D-slicer software version 4.10.2 (Boston, USA) by a radiologist (Reader 1, Z.X.Y.) with 12 years of working experience in abdominal imaging. In ROIs of the liver and spleen, each ROI was as close as possible to the margin but excluded large vascular structures and artifacts to avoid adjacent organs such as the gallbladder, intestine, stomach, kidney and mesentery (Figure 1).

After image segmentation, we used Python 3.8.3 based on pyradiomics (version 3.0; https://pyradiomics.readthedocs.io/en/latest/index.html) for feature extraction. A total of 2,264 features were extracted from the liver and spleen ROIs (1,132 from each organ). Furthermore, the images of 173 patients were segmented by another radiologist (Reader 2, W.X.M.) who specialized in abdominal imaging and had 26 years of working experience to evaluate reproducibility. Reader 1 outlined the ROIs again after 1 month to minimize recall bias. The interobserver reproducibility and intraobserver reproducibility of all extracted features were evaluated by intra/interclass correlation coefficients (ICCs). Features with ICCs > 0.75 were considered to have good reproducibility. In the reproducibility analysis, a total of 1,882 features (953 from liver images and 929 features from spleen images) were found to be sufficiently reproducible and stable (ICCs > 0.75).

The radiomics workflow is presented in Figure 2. The training cohort was used for feature selection and model building, while the validation cohort was used to test model performance. To select the best features and avoid overfitting from the training cohort, we used the least absolute shrinkage and selection operator (LASSO) method and conducted 100 iterations of 10-fold cross-validations to develop a Lasso Cox regression model (18, 19). The coefficients of some features were decreased to zero by adding penalty terms through Lasso Cox regression, and the features with non-zero coefficients were then selected. Moreover, the optimal tuning parameter (λ) is the value for which the partial likelihood deviance is the minimum criterion. The significant features were weighted with their coefficients and summed to form the Rad-score (Rad-score = coefficient 1 × feature 1+ coefficient 2 × feature 2…) (20). The Rad-score_Liver, Rad-score_Spleen, and Rad-score_{Liver−Spleen} were calculated by a linear combination of the selected features from the liver, spleen and a combination of both organs that were weighted by their own coefficients in the LASSO Cox regression model.

Harrell's concordance index (C-index) and the hazard ratio (HR) were calculated to evaluate the predictive accuracy of the Rad-score. Kaplan–Meier survival analysis and the log-rank test were used to evaluate the stratification ability of each model. In addition, calibration curves were generated to assess the calibration of the Rad-score. Decision curve analysis (DCA) was performed to analyze the clinical usefulness of the Rad-score by measuring the net benefit at different threshold probabilities.

Statistical analysis was conducted with R software (version 3.6.1; http://www.r-project.org). The following R packages were used: glmnet, for running LASSO Cox; psych, for calculating ICCs; survival, for building the Cox proportional risk model and drawing Kaplan–Meier curves; hmisc, for calculating the C-index; rms, for generating calibration curves; stdca, for plotting DCA results; stats, for Mann–Whitney U and chi-square-tests; survcomp, for comparison of different C-indexes; and SurvProb, for predicting EV rebleeding probabilities. All statistical tests were two-sided, and p-values < 0.05 were considered significant.

A total of 173 patients were divided into a training set and a validation set at a ratio of 7:3 with a random sampling method; 121 patients constituted the training cohort, and the other 52 constituted the validation cohort. There was no significant difference in clinical characteristics between the two cohorts (p = 0.212–0.868; Table 1). During the follow-up periods, rebleeding occurred in 39 of 173 patients (22.5%) within 1 year.

After extracting features from ROIs, we obtained 7, 6, and 10 features with non-zero coefficients as the predictive radiomics features for the liver, spleen and both organs, respectively. The Rad-score formulas are follows:

Figure 3 represents the process of features selected with non-zero coefficients in the LASSO Cox regression model in the training cohort. The ICCs of the selected features are also described in Supplementary Table 1.

Predictive factors for EV rebleeding are summarized in Table 2. In the univariate analysis, portal vein diameter [HR = 1.114; 95% confidence interval (CI) = 1.010–1.220; p = 0.002], Rad-score_Liver (HR = 1.692; 95% CI = 1.132–2.262; p < 0.001), Rad-score_Spleen (HR = 1.368; 95% CI = 1.019–1.741; p < 0.001) and Rad-score_{Liver−Spleen} (HR = 3.025; 95% CI = 2.029–3.961; p < 0.001) showed a significant association with EV rebleeding. In the multivariate analysis, Rad-score_Liver (HR = 1.355; 95% CI = 1.101–1.503; p =0.008), Rad-score_Spleen (HR = 1.148; 95% CI = 1.007–1.396; p =0.034) and Rad-score_{Liver−Spleen} (HR = 2.682; 95% CI = 1.793–3.512; p < 0.001) were independent risk factors for EV rebleeding.

To compare the predictive performance of Rad-score_Liver, Rad-score_Spleen and Rad-score_{Liver−Spleen} for EV rebleeding, the C-index was calculated. Rad-score_{Liver−Spleen} showed significantly better performance than Rad-score_Liver and Rad-score_Spleen, yielding a C-index of 0.853 (95% CI = 0.776–0.904) in the training cohort and 0.822 (95% CI = 0.749–0.875) in the validation cohort (Table 3). The calibration curves of the Rad-score_{Liver−Spleen} at 3, 6, 9, and 12 months showed that the predicted probability was very close to the actual probability (Figures 4A,B). DCA showed that the Rad-score_{Liver−Spleen} yielded more clinical net benefit under almost all threshold probabilities, indicating that the Rad-score_{Liver−Spleen} is more practical than the Rad-score_Liver and Rad-score_Spleen for predicting esophageal variceal rebleeding in cirrhotic patients (Figures 4C,D).

We also used clinical indexes, including HVPG, Child-Pugh score, MELD score and EV size, to predict the probability of rebleeding. Compared with the clinical indexes, the Rad-score_{Liver−Spleen} exhibited a higher C-index in the training and validation cohorts (Table 3).

Based on the cutoff values of the Rad-score_{Liver−Spleen} determined by X-tile software (21), all patients were divided into 3 risk groups to predict rebleeding in the training (low-risk, −0.03–0.30; intermediate-risk, 0.31–0.61; high-risk, 0.62–0.9; Figures 5A,B) and validation (low-risk, −0.03–0.30; intermediate-risk, 0.31–0.61; high-risk, 0.62–0.9; Figures 5D,E) cohorts. The 12-month rebleeding probabilities among the 3 risk groups in the training cohort were 0.090, 0.202, and 0.407. Likewise, significant differences were observed in the validation cohort (12-month rebleeding probability: 0.097 for the low-risk group, 0.218 for the intermediate-risk group, and 0.436 for the high-risk group; Table 4). Kaplan–Meier curves showed that the cumulative incidences of rebleeding in the training and validation cohorts were accurately differentiated by the risk stratification system (Figures 5C,F).

For the subgroup analysis in the training cohort, the Rad-score_{Liver−Spleen} in the HBV group had significantly better performance than that in the non-HBV group (C-index, 0.903 vs. C-index, 0.791; P < 0.001). Significant differences were also observed in the validation cohort (C-index, 0.884 vs. C-index, 0.781; P < 0.001, Table 5).