Preoperative noninvasive prediction of Rex vs. Warren shunt selection in children with extrahepatic portal vein obstruction: a machine learning model based on serology and ultrasound

Introduction

Extrahepatic portal vein obstruction (EHPVO), or cavernous transformation of the portal vein, represents the most common etiology of pediatric portal hypertension (1), with approximately 10% of affected children succumbing to recurrent upper gastrointestinal bleeding (2,3). While preoperative wedged hepatic venous portography remains the gold standard for evaluating portal venous anatomy and determining surgical approach (4,5), it carries significant limitations including procedural risk, time consumption, radiation exposure, and high cost – underscoring the need for non-invasive alternatives (6).

Primary surgical interventions include the Rex shunt (mesoportal bypass) and selective portosystemic (Warren) shunt (7,8). The Rex shunt, which reconstructs physiological portal flow via autologous venous grafting into the Rex recessus, demonstrates superior efficacy in improving metabolic profiles compared to the Warren shunt (3,7,9,10). Warren shunt is a type of selective portosystemic shunt designed to maintain portal and mesenteric blood perfusion to the liver. This surgical approach includes the classic distal splenorenal shunt, as well as other variations such as the splenorenal shunt. It has proven effective in managing variceal bleeding and hypersplenism, while also demonstrating a comparable long-term outcomes in terms of survival and rebleeding rates (1,11). Nevertheless, this technique fails to restore physiological portal flow to the liver, and the resultant reduction in hepatic portal perfusion may lead to inevitable liver injury. Thus, Warren shunt is recommended as an alternative when the Rex shunt is not feasible due to anatomical constraints (12).

Ultrasound serves as the cornerstone preoperative modality for assessing graft suitability and obstruction severity (1,13-16), with documented statistical differences in biochemical, hematological, and radiographic parameters between shunt outcomes (17,18). This study therefore aims to develop a machine learning model integrating ultrasound biomarkers and serological profiles, providing a non-invasive preoperative assessment tool to guide surgical decision-making in pediatric EHPVO. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-571/rc).

Methods

Study design

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of the Guangzhou Women and Children’s Medical Center (approval No. 2023082916172037) and individual consent for this retrospective analysis was waived. A total of 198 children with EHPVO who underwent surgical intervention at our center from December 2017 to June 2023 were initially enrolled. Of these, 157 and 41 patients received Rex shunt and Warren shunt procedures, respectively. All included patients were diagnosed with EHPVO through wedged hepatic venous portography and subsequently underwent shunt surgery. Patients were excluded (n=95) if they met any of the following criteria: incomplete preoperative ultrasound imaging data, history of multiple shunt surgeries (>2 procedures), incomplete clinical records, or loss to follow-up. Consequently, the final cohort comprised 81 Rex shunt and 22 Warren shunt patients. The study workflow is detailed in Figure 1.

Figure 1 The workflow of this study. EHPVO, extrahepatic portal vein obstruction.

Data collection

Demographic and clinical records included patient age, sex, body weight, and history of prior splenectomy. Comprehensive serological profiling encompassed: liver function parameters [glutamic oxaloacetic transaminase (AST), glutamic pyruvic transaminase (ALT), albumin (ALB), gamma-glutamyl transferase (GGT)], hepatobiliary markers [total bile acids (TBA), total bilirubin (TBIL)], coagulation profile [prothrombin time (PT), international normalized ratio (INR)], and hematological indices [red blood cell count (RBC), white blood cell count (WBC), platelet count (PLT)]. Preoperative ultrasound evaluation featured B-mode measurements of right hepatic lobe oblique diameter (RHLOD), splenic longitudinal/short-axis diameters, and luminal diameters of the left portal vein branch and splenic vein, supplemented by Doppler assessment of flow velocities in the left portal vein and splenic vein.

Ultrasound measurements were performed by two radiologists with over 10 years of experience in the field, utilising advanced ultrasound equipment such as Philips IE33, Philips EPIQ7, GE LOGIQ E11, GE Vivid7 Dimension, and SuperSonic Aixplorer. These devices were equipped with 7–12 MHz convex array and linear array probes. During each ultrasound examination, the diameters of the sagittal left branch of the portal vein and splenic vein measurements were taken from wall to wall along their longitudinal axes and the largest consistent segments. The highest flow velocities in these veins were recorded. However, it should be noted that gas interference may compromise ultrasound imaging at the superior mesenteric vein with the bypass vein, leading to unclear visualization. Spleen thickness is measured as the distance from the splenic hilum to the curved line tangent to the spleen’s opposite edge. Spleen length is defined as the maximum distance between the upper and lower poles of the spleen. The RHLOD is defined as the maximum length measured along the primary oblique plane of the right hepatic lobe, extending from its posterosuperior to anteroinferior borders. Measurement requires precise transducer alignment parallel to the longitudinal axis of the right hepatic lobe. All measurements were conducted by two ultrasound physicians who were unaware of the patients’ specific conditions. Measurements of vein diameter and velocity were taken during breath-holding, if the child was able to cooperate. Infants who were non-cooperative were sedated with chloral hydrate. All children were scanned in the supine position.

The indications for the Rex shunt primarily refer to the surgical management guidelines for EHPVO proposed by Superina et al. in 2006 (19). These guidelines specify the prerequisites for performing a Rex shunt, including confirmation of EHPVO by portal venography, patency of the left portal vein as confirmed by intrahepatic portography or intraoperative exploration, and the absence of underlying liver diseases such as fibrosis or cirrhosis (19). For preoperative evaluation, our institution routinely employs ultrasound (US) and contrast-enhanced computed tomography (CECT) for initial screening of EHPVO. These modalities allow assessment of the vascular anatomy, patency and caliber of potential graft vessels, and required graft length. Although retrograde portography remains the gold standard radiological method for diagnosing EHPVO and is essential preoperatively—particularly for demonstrating patency of the Rex recessus, which is a mandatory criterion for Rex shunt—it is not used as a first-line diagnostic tool due to its invasive nature. Meanwhile, magnetic resonance angiography (MRA), while capable of providing high-resolution vascular images, is not incorporated into our standard preoperative protocol owing to its higher cost, longer examination time, and challenges related to compliance among young pediatric patients.

Feature selection and model development

The cohort was randomly partitioned into training and validation sets at a 7:3 ratio to ensure robust validation. Within the training subset, the least absolute shrinkage and selection operator (LASSO) algorithm (glmnet v4.1-3) screened predictor variables to eliminate multicollinearity and derive optimal features (19). Subsequently, five machine learning algorithms were implemented in R 4.5.0 under default configurations: LASSO regression, support vector machine (SVM; e1071 v1.7-9), random forest (RF; randomForest v4.6-14), extreme gradient boosting (XGBoost; xgboost v1.7.8.1) (20,21), and Naïve Bayes (NB; e1071 v1.7-9). Model performance underwent validation through internal validation dataset, evaluated via comprehensive metrics including accuracy, area under the receiver operating characteristic curve (AUC), Kappa statistic, specificity, sensitivity, Youden’s index, and F1-score. Systematic metric comparison identified the optimal algorithm.

Model interpretation

In order to further understand the importance of the selected variables for the optional model, we employed Shapley Additive Explanations (SHAP) analysis to visualize their contribution to the model’s diagnostic capacity.

Intra-observer and inter-observer reproducibility evaluation

Intra-observer and inter-observer agreement of the measurements were assessed in randomly selected subjects. The same observer (X.Z.) measured the diameter and velocity of the veins twice to evaluate intra-observer reproducibility. A second observer (B.Y.), blinded to the first observer’s results, performed the same measurements to assess inter-observer reproducibility.

Statistical analysis

The analysis of baseline data began with normality tests of quantitative data. Count data were summarized as counts (percentage, %), and statistical analysis was performed using Fisher’s exact test. Continuous data were summarized using the median (Min, Max) and mean [standard deviation (SD)], with group comparisons conducted using the Wilcoxon test. Receiver operating characteristic (ROC) curve analysis was conducted to assess diagnostic performance. The statistically significant cut-off value of two-tailed P value was 0.05.

Results

Baseline characteristics

A total of 103 pediatric patients with extrahepatic portal vein obstruction (EHPVO) undergoing surgical management at our institution were enrolled and randomly allocated to training (n=73) and validation cohorts (n=30) at a 7:3 ratio. Initial comparison revealed no significant differences in baseline characteristics between cohorts across all variables (Table 1). Surgical subgroup analysis demonstrated that compared to Rex shunt recipients, Warren shunt patients exhibited significantly higher preoperative glutamic pyruvic transaminase (ALT; P=0.01) and GGT (P<0.001), reduced red blood cell counts (P=0.006) and albumin levels (P=0.02), along with larger RHLOD (P=0.009) and greater splenic vein internal diameter (P=0.001) on ultrasonography (Table S1).

Feature selection and model performance

A total of 21 features were initially extracted, with LASSO regression analysis identifying six optimal predictors for modeling (Figure S1). These LASSO-selected features were subsequently employed to construct five distinct surgical outcome prediction models: LASSO regression, SVM, NB, RF, and XGBoost. As demonstrated in Figure 2 and Table S2, the XGBoost algorithm significantly outperformed other models in both training and validation sets, achieving AUC values of 0.90 [95% confidence interval (CI): 0.79–0.99] and 0.75 (95% CI: 0.54–0.97), respectively.

Figure 2 Performance comparison of five models on the training cohort (A) and validation cohort (B). AUC, area under the curve; LASSO, least absolute shrinkage and selection operator; NB, Naïve Bayes; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting.

Model interpretation and clinical implementation platform

To elucidate the contribution of LASSO-selected features to diagnostic efficacy, we employed SHAP analysis. As visualized in Figure 3, GGT parameters demonstrated substantial predictive contribution to shunt selection. Building upon the optimal XGBoost model, we developed an interactive web-based clinical calculator (https://rexshunt.shinyapps.io/rexorwarren/, Figure 4) to support personalized therapeutic decision-making for children with extrahepatic portal vein obstruction. Figure 5 demonstrates the workflow of the interactive web-based clinical calculator through an illustrative case. Clinicians input preoperative ultrasound parameters and serological markers, generating real-time surgical approach predictions with probability scores. This platform enables clinicians to input preoperative serological and sonographic parameters to generate surgical approach predictions.

Figure 3 Model interpretation. The importance ranking of different variables according to the Avg (|SHAP|) using the optimal XGBoost model. ALB, albumin; Avg (|SHAP|), Shapley Additive Explanations absolute mean value; GGT, gamma-glutamyl transferase; LPVV, velocity of left branch of portal vein; RBC, red blood cell; RHLOD, right hepatic lobe oblique diameter; SVV, velocity of spleen vein.

Figure 4 Online computing platform presentation of the optimal XGBoost model. ALB, albumin; GGT, gamma-glutamyl transferase; RBC, red blood cell.

Figure 5 Preoperative sonographic and Doppler evaluation in an 8-year-old boy with EHPVO. (A) Right hepatic lobe oblique diameter measurement (RHLOD: 109 mm); (B) Left portal vein velocity (LPVV: 13.3 cm/s); (C) Splenic vein velocity (SVV: 19 cm/s). Laboratory findings: RBC 2.5×10¹²/L, ALB 39.1 g/L, GGT 15 IU/L. The online platform predicted 40% Rex shunt probability. Rex shunt was performed but resulted in complete graft occlusion confirmed by portography and contrast-enhanced CT at 18-month follow-up (see details in Appendix 1). ALB, albumin; CT, computed tomography; EHPVO, extrahepatic portal vein obstruction; GGT, gamma-glutamyl transferase; LPVV, velocity of left branch of portal vein; RBC, red blood cell; RHLOD, right hepatic lobe oblique diameter; SVV, velocity of spleen vein.

Intra- and inter-observer variability

The results of intra- and inter-observer variability for ultrasound indicators are demonstrated that there was a high degree of agreement for all ultrasound indicators. Detailed statistical outcomes are provided in Table S3.

Discussion

EHPVO constitutes the most common cause of pediatric portal hypertension. While the Rex shunt is increasingly recognized as the therapeutic gold standard for restoring physiological portal flow (9), its stringent anatomical prerequisites and high postoperative stenosis rates necessitate alternative approaches such as the Warren shunt—a portosystemic diversion that alleviates portal hypertension symptoms in ineligible patients. Although preoperative wedged hepatic venous portography remains the primary modality for evaluating portal venous anatomy, emerging evidence supports CT-based Rex recessus classification (types I–III) for surgical candidate selection (5,18). However, both CECT and wedged hepatic venous portography raise significant clinical and ethical concerns due to their invasiveness and radiation exposure in pediatric populations.

To address these limitations, we developed a machine learning model that noninvasively integrates ultrasound features and serological markers for preoperative assessment. SHAP-based interpretability analysis identified GGT as the most influential predictor for surgical selection, consistent with its role in evaluating hepatic impairment. Specifically, lower GGT levels indicated relatively preserved hepatocellular integrity and minimal secondary liver injury, highlighting patients in earlier disease stages with milder portal hypertension who are optimal candidates for Rex reconstruction—a procedure that restores physiological hepatopetal flow in those with sufficient functional reserve. In contrast, elevated GGT levels reflected more advanced hepatic dysfunction and possible cholestatic stress, aligning with the profile of Warren shunt candidates who exhibited significantly higher ALT and GGT, reduced RBC counts and albumin, and hepatomegaly, all indicative of greater hepatic compromise and hypersplenism. These individuals benefit instead from Warren shunt due to its effective decompressive effect without relying on compromised parenchymal function. Collectively, these findings reinforce that Rex surgery is favored in earlier disease stages, whereas Warren shunt serves better in cases with significant hepatic injury.

The most significant clinical implication of our model lies in its potential to refine preoperative workflows. As indicated by the accuracy rates of 0.89 in the training cohort and 0.80 in the validation cohort (Table S2), the implementation of our model could potentially allow approximately 80% of EHPVO patients to avoid invasive imaging examinations during preoperative assessment, thereby enabling earlier preparation for surgical intervention. The high sensitivity (81.48%) means the model is highly effective at identifying patients who are likely to be candidates for a Rex shunt. In practice, this could be used to triage patients. Patients predicted as “Rex-positive” by the model (with high probability) could potentially proceed directly to operative planning, thereby avoiding invasive portography for a significant subset. While the moderate specificity (66.67%) and consequent false positive rate (33.33%) indicate that some patients deemed “Rex-positive” by the model (with low probability) may still require standard imaging for final confirmation, the model’s high sensitivity ensures that very few true Rex candidates (false negative rate of 18.52%) would be missed. This triage function could streamline decision-making, reduce procedural risks and costs, and prioritize invasive testing for complex cases.

We explicitly acknowledge that the limited sample size, especially the small cohort of Warren shunt procedures (n=22), constitutes a significant limitation to this study. This constraint not only affects the statistical power during model training but also raises valid concerns regarding the stability and generalizability of the predictive algorithm. Therefore, we emphatically emphasize that external validation through multi-center collaborations is indispensable before large-scale clinical application. Additionally, future refinements of the model should aim to integrate more precise indicators of disease severity, such as endoscopic variceal characteristics and clinical documentation of gastrointestinal bleeding. The current single-center design also necessitates further validation across ethnically and geographically diverse populations to ensure widespread applicability.

Conclusions

This study established a predictive model combining serological markers and ultrasound parameters to preoperatively assess portal venous anatomy in pediatric EHPVO. External validation confirmed the model’s accuracy and stability. The online tool provides a noninvasive, user-friendly solution to guide surgical strategy selection for children with EHPVO.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-571/rc

Data Sharing Statement: Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-571/dss

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-2025-571/prf

Funding: This study was supported by grants from the Science and Technology Plan Project of Guangzhou, China (Nos. 2024A04J4423 and 2024A03J0958).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2025-571/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of the Guangzhou Women and Children’s Medical Center (approval No. 2023082916172037) and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Zhang J, Li L. Rex Shunt for Extra-Hepatic Portal Venous Obstruction in Children. Children (Basel) 2022;9:297. [Crossref] [PubMed]
2. Wang RY, Wang JF, Liu Q, et al. Combined Rex-bypass shunt with pericardial devascularization alleviated prehepatic portal hypertension caused by cavernomatous transformation of portal vein. Postgrad Med 2017;129:768-76. [Crossref] [PubMed]
3. Wang J, Ning Y, Ren H, et al. Medium-to Long-term Outcomes of Rex Shunt in 105 Children With Extrahepatic Portal Vein Obstruction in China. J Pediatr Surg 2025;60:161930. [Crossref] [PubMed]
4. Kaur P, Khanna R, Sood V, et al. Wedged hepatic vein portovenography for assessment of Rex vein patency in children with extrahepatic portal venous obstruction. J Pediatr Gastroenterol Nutr 2024;79:213-21. [Crossref] [PubMed]
5. Chang X, Liu L, Wang J, et al. Effectiveness of Preoperative Intrahepatic Portal Venous Classification System in Guiding Preoperative Surgical Decisions and Predicting Hypotensive Effects After Meso-rex Bypass for Children With EHPVO. J Pediatr Surg 2025;60:161990. [Crossref] [PubMed]
6. Zielsdorf S, Narayanan L, Kantymyr S, et al. Surgical shunts for extrahepatic portal vein obstruction in pediatric patients: a systematic review. HPB (Oxford) 2021;23:656-65. [Crossref] [PubMed]
7. Lautz TB, Keys LA, Melvin JC, et al. Advantages of the meso-Rex bypass compared with portosystemic shunts in the management of extrahepatic portal vein obstruction in children. J Am Coll Surg 2013;216:83-9. [Crossref] [PubMed]
8. de Ville de Goyet J, D'Ambrosio G, Grimaldi C. Surgical management of portal hypertension in children. Semin Pediatr Surg 2012;21:219-32. [Crossref] [PubMed]
9. Zhang JS, Li L. Effectiveness of Rex shunt for improving the abnormal portal hemodynamics and portal venous pathology in EHPVO animal model. Pediatr Surg Int 2023;39:192. [Crossref] [PubMed]
10. Carollo V, Marrone G, Cortis K, et al. Multimodality imaging of the Meso-Rex bypass. Abdom Radiol (NY) 2019;44:1379-94. [Crossref] [PubMed]
11. Lautz TB, Kim ST, Donaldson JS, et al. Outcomes of percutaneous interventions for managing stenosis after meso-Rex bypass for extrahepatic portal vein obstruction. J Vasc Interv Radiol 2012;23:377-83. [Crossref] [PubMed]
12. Tantemsapya N, Laohapensang M. The effectiveness of alternative vessel grafts for meso-rex bypass in the treatment of extrahepatic portal vein obstruction in children. Pediatr Surg Int 2024;41:30. [Crossref] [PubMed]
13. Xie X, Meng Q, Lu Q, et al. Clinical Value of Ultrasound in Evaluating Stent Placement for Managing Graft Stenosis after Meso-rex Bypass. Zhongguo Yi Xue Ke Xue Yuan Xue Bao 2021;43:886-91. [Crossref] [PubMed]
14. Yuldashev RZ, Aliev MM, Shokhaydarov SI, et al. Spleen stiffness measurement as a non-invasive test to evaluate and monitor portal hypertension in children with extrahepatic portal vein obstruction. Pediatr Surg Int 2020;36:637-41. [Crossref] [PubMed]
15. Chaves IJ, Rigsby CK, Schoeneman SE, et al. Pre- and postoperative imaging and interventions for the meso-Rex bypass in children and young adults. Pediatr Radiol 2012;42:220-32; quiz 271-2. [Crossref] [PubMed]
16. Chen W, Rodriguez-Davalos MI, Facciuto ME, et al. Experience with duplex sonographic evaluation of meso-rex bypass in extrahepatic portal vein obstruction. J Ultrasound Med 2011;30:403-9. [Crossref] [PubMed]
17. Wen Z, Wang J, Yang C, et al. Is re-Rex shunt a better choice for patients with failed Rex shunt? Front Pediatr 2023;11:1135059. [Crossref] [PubMed]
18. Wu H, Zhou N, Lu L, et al. Value of preoperative computed tomography for meso-Rex bypass in children with extrahepatic portal vein obstruction. Insights Imaging 2021;12:109. [Crossref] [PubMed]
19. Superina R, Shneider B, Emre S, et al. Surgical guidelines for the management of extra-hepatic portal vein obstruction. Pediatr Transplant 2006;10:908-13. [Crossref] [PubMed]
20. Chen T, Guestrin C. editors. XGBoost: A Scalable Tree Boosting System. San Francisco, California, USA: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; 2016.
21. Breiman L. Random Forests. Machine Learning 2001;45:5-32.