Machine learning prediction of hepatic encephalopathy for long-term survival after transjugular intrahepatic portosystemic shunt in acute variceal bleeding

Abstract

BACKGROUND

Transjugular intrahepatic portosystemic shunt (TIPS) is an effective intervention for managing complications of portal hypertension, particularly acute variceal bleeding (AVB). While effective in reducing portal pressure and preventing rebleeding, TIPS is associated with a considerable risk of overt hepatic encephalopathy (OHE), a complication that significantly elevates mortality rates.

AIM

To develop a machine learning (ML) model to predict OHE occurrence post-TIPS in patients with AVB using a 5-year dataset.

METHODS

This retrospective single-center study included 218 patients with AVB who underwent TIPS. The dataset was divided into training (70%) and testing (30%) sets. Critical features were identified using embedded methods and recursive feature elimination. Three ML algorithms-random forest, extreme gradient boosting, and logistic regression-were validated via 10-fold cross-validation. SHapley Additive exPlanations analysis was employed to interpret the model’s predictions. Survival analysis was conducted using Kaplan-Meier curves and stepwise Cox regression analysis to compare overall survival (OS) between patients with and without OHE.

RESULTS

The median OS of the study cohort was 47.83 ± 22.95 months. Among the models evaluated, logistic regression demonstrated the highest performance with an area under the curve (AUC) of 0.825. Key predictors identified were Child-Pugh score, age, and portal vein thrombosis. Kaplan-Meier analysis revealed that patients without OHE had a significantly longer OS (P = 0.005). The 5-year survival rate was 78.4%, with an OHE incidence of 15.1%. Both actual OHE status and predicted OHE value were significant predictors in each Cox model, with model-predicted OHE achieving an AUC of 88.1 in survival prediction.

CONCLUSION

The ML model accurately predicts post-TIPS OHE and outperforms traditional models, supporting its use in improving outcomes in patients with AVB.

Key Words: Transjugular intrahepatic portosystemic shunt; Acute variceal bleeding; Overt hepatic encephalopathy; Machine learning; Logistic regression

Core Tip: This study developed a machine learning (ML) model to predict overt hepatic encephalopathy (OHE) after transjugular intrahepatic portosystemic shunt (TIPS) in patients with acute variceal bleeding (AVB). Utilizing a 5-year retrospective dataset of 218 patients, key features such as Child-Pugh score, age, and portal vein thrombosis were identified. The ML model demonstrated a strong performance, with an area under the curve of 0.825. This ML model effectively predicts post-TIPS OHE, providing a valuable tool for tailoring personalized treatment plans. Its superior performance over traditional models supports its integration into clinical practice to enhance outcomes for patients with AVB undergoing TIPS.

INTRODUCTION

Transjugular intrahepatic portosystemic shunt (TIPS) is a widely recognized intervention for managing complications of portal hypertension, particularly in patients suffering from acute variceal bleeding (AVB)[1-3]. Despite TIPS effectively reducing portal pressure and controlling bleeding, it carries a significant risk of overt hepatic encephalopathy (OHE) with an incidence of 10% to 30%[4]. Multiple studies and meta-analyses have demonstrated that post-TIPS hepatic encephalopathy, especially OHE, significantly increases the risk of early mortality in patients. Moreover, post-TIPS OHE and Child-Pugh grade have been recognized as key independent predictors of early mortality[5]. This risk constrains TIPS from being considered first-line treatment for AVB. Accurate prediction of OHE post-TIPS is essential for optimizing patient management and improving clinical outcomes.

Previous studies have developed various predictive models for assessing and stratifying the risk factors for OHE, yet none have been widely adopted in clinical practice[6-8]. Recent advancements in machine learning (ML) have significantly improved the accuracy of clinical outcome predictions by leveraging large datasets and utilizing sophisticated algorithms that can uncover intricate patterns and relationships within the data, enhancing predictive capabilities in medical practice[9]. ML models incorporate diverse variables from clinical, biochemical, and procedural data, thereby enhancing predictive accuracy[10,11]. Various ML algorithms, including logistic regression (LR) and more sophisticated approaches such as extreme gradient boosting (XGBoost), have demonstrated strong potential for improving clinical outcome predictions by capturing essential predictive factors across multiple medical conditions. By utilizing advanced techniques such as recursive feature elimination (RFE) and employing robust evaluation metrics such as the area under the curve (AUC), these models can enhance prediction accuracy. In addition, the integration of model results with preoperative indicators offers valuable insights into patient prognosis following TIPS, contributing to more informed decision-making and personalized treatment strategies.

This study developed an ML model for predicting post-TIPS OHE in patients with AVB using a 5-year longitudinal dataset and assessed the impact of model metrics on patient survival. The model’s predictions are expected to serve as crucial indicators of long-term survival following TIPS.

MATERIALS AND METHODS

Patient criteria

This retrospective, single-center study was conducted at our institution with approval of the institutional review board, following the principles of the Declaration of Helsinki. Patients who underwent TIPS for AVB between April 2017 and December 2020 were screened. The inclusion criteria were: (1) Liver cirrhosis diagnosis confirmed by clinical evaluation, imaging, or biopsy; and (2) Initial AVB due to portal hypertension. The exclusion criteria were: (1) Previous TIPS or surgical shunt; (2) Repeated occurrences of AVB and related treatments; (3) Liver cancer or other extrahepatic malignancies; (4) TIPS unrelated to AVB; (5) Severe cardiac, pulmonary, or renal insufficiency and expected survival time of less than 3 months; and (6) Loss to follow-up within 1 month post-TIPS (Figure 1). All patients with AVB received treatment with vasoactive drugs, received endoscopic hemostasis as appropriate, and underwent TIPS within 7 days of first onset. After TIPS, all patients were administered lactulose and rifaximin to prevent OHE onset. Within 7 days before the TIPS procedure, demographic and baseline clinical characteristics were collected.

Figure 1 Study design. A: Development and validation of the overt hepatic encephalopathy prediction model; B: Flowchart of patient enrollment and grouping. LR: Logistic regression; RFE: Recursive feature elimination; K-M: Kaplan-Meier; SHAP: SHapley Additive exPlanations; DCA: Decision curve analysis; AVB: Acute variceal bleeding; TIPS: Transjugular intrahepatic portosystemic shunt; ML: Machine learning.

Follow-up and outcomes

Patients were followed up regularly at specific intervals-1 week, 1 month, 3 months, 6 months, and 1 year after the TIPS procedure-and annually thereafter until death or the end of the follow-up period. Key outcome measures included the occurrence of OHE, defined as the development of clinically apparent hepatic encephalopathy requiring medical intervention, and overall survival (OS), defined as the time from TIPS placement to death from any cause. Follow-up data were collected through clinical visits, imaging studies, and biochemical tests. The occurrence of OHE and other complications was recorded based on standardized diagnostic criteria, and survival status was confirmed through hospital records or follow-up communications.

TIPS procedure

All procedures were successfully performed by the same interventional radiology team with 5 years to 20 years of experience. Using the Seldinger technique, the right internal jugular vein was punctured successfully, followed by pre-dilation with a 10F dilator. Under guidewire guidance, the RUPS-100 sheath (COOK Medical, Bloomington, IN, United States) was advanced to the inferior vena cava to measure inferior vena cava pressure. The catheter was positioned in the hepatic vein and confirmed by a “smoke” sign. Under fluoroscopic guidance, a puncture needle was introduced to puncture the left or right branch of the portal vein via the hepatic vein through the liver parenchyma. Then, a catheter was advanced into the main portal vein for pressure measurement and contrast imaging to assess portal venous flow and varices. Embolization was performed using coils and/or tissue glue. An exchange for a super-stiff guidewire was done, and the liver parenchymal tract was dilated with a 6 mm balloon. The Viatorr covered stent (Gore Medical, Flagstaff, AZ, United States) was deployed to create a hepatic-portal vein shunt, followed by dilation of the stent with an 8 mm balloon. Follow-up imaging confirmed stent position and shunt function, and pressures in the inferior vena cava and portal vein were remeasured. The right internal jugular vein sheath was removed, and the puncture site was compressed and bandaged securely. The goal was a portosystemic gradient of less than 12 mmHg or a reduction of 50% from the baseline portal pressure gradient for AVB.

Data pre-processing

A dataset was built using baseline demographic characteristics (sex, age, etiology), along with clinical, biochemical, and procedural data. We employed embedded methods (EMs) with LR as the classifier, along with RFE for feature selection. The selected features were used to train models and predict OHE after TIPS. The data were normalized to standardize the variables to a common scale. Subsequently, the dataset was randomly divided into training and test sets in a ratio of 7:3. To address the issue of the imbalanced sample sizes between patients with and without OHE, we used the synthetic minority oversampling technique (SMOTE) to balance the training dataset. The SMOTE algorithm was implemented using the imblearn package in Python 3.7.13.

Model development

In the present study, three common ML algorithms, including random forest classifier, XGBoost, and LR classifier, were developed and validated using the scikit-learn and XGBoost packages in Python 3.7.13 to predict OHE after TIPS. Ten-fold cross-validation was used for model derivation and internal validation. The grid search algorithm was used during the training process for each model to optimize the model’s hyperparameters on the training set as the standard of the AUC of the receiver operating characteristic curve. We integrated SHapley Additive exPlanations (SHAP), a locally interpretable method, to elucidate the optimal model. SHAP applies game theory to explain ML model outputs, linking optimal credit assignment with localized explanations using classical SHAP values. This enabled identification of pivotal features influencing model outputs and provided insights into the decision-making process of the models.

Statistical analyses

Statistical analyses were performed using RStudio (version 4.3.1). Categorical variables are expressed as numbers and percentages, while continuous variables are summarized as the median with 95% confidence interval (CI) or mean with SD. Baseline characteristics of patients were analyzed using the Student’s t-test, Mann-Whitney U test, and χ² test. OS in the test sets was estimated using the Kaplan-Meier analysis, and the log-rank test was used to compare OS differences between patients with and without post-TIPS OHE. In Kaplan-Meier analysis, we stratified the patients into OHE and non-OHE groups based on the actual OHE status or predicted OHE status. Covariates were adjusted to meet the proportional hazards, linearity, and independence assumptions of Cox regression. Stepwise Cox regression analysis was conducted, and variables with P < 0.05 were included in further Cox modeling. The final model was validated using the concordance C statistic (known as the C-index) and Brier score. All data were analyzed using two-sided tests, and P < 0.05 was considered statistically significant.

RESULTS

Patient characteristics

A total of 218 patients were included in the study (male/female: 163/55), with a median age of 52.14 ± 11.96 years at the time of the TIPS procedure. The predominant cause of AVB was portal hypertension resulting from hepatitis B virus infection (70.18%). The median follow-up time was 5 years. During this period, the 5-year survival rate was 78.4% (47/218), while the incidence of OHE was 15.1% (33/218). Detailed demographic data and characteristics of the patients are presented (Table 1).

Table 1 Baseline characteristics, mean ± SD/n (%).

Variables	Total, n = 218	Training set, n = 152	Test set, n = 66	P value
Age in years	52.14 ± 11.96	51.86 ± 11.93	52.80 ± 12.10	0.592
OS	47.83 ± 22.95	48.56 ± 22.66	46.14 ± 23.69	0.475
Etiology				0.044
Hepatitis B virus	153 (70.18)	110 (72.37)	43 (65.15)
Hepatitis C virus	11 (5.05)	9 (5.92)	2 (3.03)
Alcoholism	24 (11.01)	15 (9.87)	9 (13.64)
Autoimmune disease	7 (3.21)	7 (4.61)	0 (0.00)
Other etiologies	23 (10.55)	11 (7.24)	12 (18.18)
HE				0.856
No	208 (95.41)	145 (95.39)	63 (95.45)
Mild	8 (3.67)	5 (3.29)	3 (4.55)
Moderate and severe	2 (0.92)	2 (1.32)	0 (0.00)
Ascites				0.686
No	73 (33.49)	54 (35.53)	19 (28.79)
Mild	85 (38.99)	56 (36.84)	29 (43.94)
Moderate	59 (27.06)	41 (26.97)	18 (27.27)
Severe	1 (0.46)	1 (0.66)	0 (0.00)
PVT				0.594
No	177 (81.19)	122 (80.26)	55 (83.33)
Yes	41 (18.81)	30 (19.74)	11 (16.67)
SMVT				0.412
No	197 (90.37)	139 (91.45)	58 (87.88)
Yes	21 (9.63)	13 (8.55)	8 (12.12)
SVT				1
No	213 (97.71)	149 (98.03)	64 (96.97)
Yes	5 (2.29)	3 (1.97)	2 (3.03)
BCS				0.318
No	213 (97.71)	147 (96.71)	66 (100.00)
Yes	5 (2.29)	5 (3.29)	0 (0.00)
CTPV				0.085
No	203 (93.12)	145 (95.39)	58 (87.88)
Yes	15 (6.88)	7 (4.61)	8 (12.12)
Pressure-related indicators
Pre-IVCP	5.66 ± 4.07	5.53 ± 4.09	5.97 ± 4.03	0.461
Pre-PVP	28.07 ± 6.62	28.05 ± 6.50	28.14 ± 6.96	0.927
Pre-PPG	22.41 ± 5.51	22.52 ± 5.43	22.17 ± 5.73	0.665
Post-IVCP	8.28 ± 4.62	8.26 ± 4.65	8.33 ± 4.59	0.911
Post-PVP	18.00 ± 5.87	18.02 ± 5.78	17.94 ± 6.12	0.926
Post-PPG	9.72 ± 3.71	9.76 ± 3.73	9.61 ± 3.69	0.775
Blood and biochemical indicators
TBIL	22.79 ± 15.02	22.97 ± 14.75	22.38 ± 15.75	0.789
ALB	33.94 ± 5.19	33.88 ± 5.32	34.10 ± 4.92	0.775
CR	90.51 ± 78.67	90.90 ± 73.21	89.61 ± 90.59	0.911
ALT	24.93 ± 18.65	24.18 ± 17.19	26.65 ± 21.71	0.37
Prolonged PT	1.86 ± 2.43	1.93 ± 2.30	1.69 ± 2.72	0.504
INR	1.30 ± 0.23	1.31 ± 0.22	1.27 ± 0.24	0.187
Sodium	140.68 ± 12.47	140.06 ± 3.41	142.11 ± 22.12	0.266
PLT	94.59 ± 77.50	89.96 ± 71.21	105.24 ± 90.05	0.182
WBC	5.17 ± 3.51	5.05 ± 3.29	5.46 ± 3.97	0.426
Clinical scores and classifications
Child-Pugh score	7.17 ± 1.71	7.18 ± 1.75	7.15 ± 1.61	0.897
Child-Pugh classification				0.776
A	85 (38.99)	61 (40.13)	24 (36.36)
B	111 (50.92)	75 (49.34)	36 (54.55)
C	22 (10.09)	16 (10.53)	6 (9.09)
MELD score	10.86 ± 3.70	11.05 ± 3.62	10.42 ± 3.87	0.243
MELD classification				0.54
Mild	103 (47.25)	67 (44.08)	36 (54.55)
Moderate	78 (35.78)	57 (37.50)	21 (31.82)
Severe	37 (16.97)	27 (18.42)	9 (13.64)
MELD-Na score	11.50 ± 4.64	11.66 ± 4.33	11.14 ± 5.31	0.447
ALBI score	-0.95 ± 0.62	-0.94 ± 0.62	-0.99 ± 0.61	0.591
ALBI classification				0.343
Low risk	52 (23.85)	39 (25.66)	13 (19.70)
High risk	166 (76.15)	113 (74.34)	53 (80.30)
FIPS score	2.49 ± 1.09	2.56 ± 0.98	2.34 ± 1.30	0.184
CLIF-C AD score	41.57 ± 10.60	42.01 ± 8.10	40.55 ± 14.86	0.351
CLIF-C AD classification				0.84
Low risk	147 (67.43)	101 (66.45)	46 (69.70)
Intermediate risk	65 (29.82)	47 (30.92)	18 (27.27)
High risk	6 (2.75)	4 (2.63)	2 (3.03)
Previous surgeries and procedures
EBL				0.986
No	195 (89.45)	136 (89.47)	59 (89.39)
Yes	23 (10.55)	16 (10.53)	7 (10.61)
Partial splenectomy				0.629
No	198 (90.83)	139 (91.45)	59 (89.39)
Yes	20 (9.17)	13 (8.55)	7 (10.61)
EIS				0.605
No	211 (96.79)	146 (96.05)	65 (98.48)
Yes	7 (3.21)	6 (3.95)	1 (1.52)
GCAE				1
No	217 (99.54)	151 (99.34)	66 (100.00)
Yes	1 (0.46)	1 (0.66)	0 (0.00)
PSE				0.989
No	213 (97.71)	148 (97.37)	65 (98.48)
Yes	5 (2.29)	4 (2.63)	1 (1.52)

ALB: Albumin; ALBI: Albumin-bilirubin; ALT: Alanine aminotransferase; BCS: Budd-Chiari syndrome; CLIF-C AD: Chronic liver failure consortium acute decompensation; CR: Creatinine; CTPV: Cavernous transformation of the portal vein; EBL: Endoscopic band ligation; EIS: Endoscopic injection sclerotherapy; FIPS: Fibrosis-4 index for liver fibrosis; GCAE: Gastric antral vascular ectasia; IVCP: Inferior vena cava pressure; HE: Hepatic encephalopathy; INR: International normalized ratio; MELD: Model for end-stage liver disease; MELD-Na: Model for end-stage liver disease sodium; OS: Overall survival; PLT: Platelet count; PPG: Portal pressure gradient; PSE: Partial splenectomy; PT: Prothrombin time; PVP: Portal venous pressure; PVT: Portal vein thrombosis; SMVT: Superior mesenteric vein thrombosis; SVT: Splenic vein thrombosis; TBIL: Total bilirubin; WBC: White blood cell.

ML models for OHE in patients after TIPS procedure

The dataset was randomly divided into a training set (n = 152, 70%) and a test set (n = 66, 30%). In the training set, there were 23 (15.1%) patients with OHE. In the test set, there were 10 (15.1%) patients with OHE. SMOTE algorithm generated 106 cases with OHE in the training set. We employed EMs for feature selection, reducing the initial set of features to 16. Subsequently, we applied RFE to further refine the features down to 9. In the test set, the AUCs for the RF, XGBoost, and LR model were 0.696, 0.796, and 0.825, respectively (Figure 2). The decision curve analysis showed that the LR model provided more net benefits for predicting OHE after TIPS. SHAP algorithm showed that the most important factors were Child-Pugh score, age, and portal vein thrombosis (PVT) (Figure 3).

Figure 2 Comparison of various machine learning models in a classification task and decision curve analysis for the logistic regression model. A: Receiver operating characteristic (ROC) curves for random forest, extreme gradient boosting, and logistic regression (LR); B: ROC curve for the LR model, presented individually with an area under the curve value of 0.825; C: Decision curve analysis for the LR model, illustrating the net benefit at various threshold probabilities in comparison to the “treat all” and “treat none” strategies. AUC: Area under the curve; XGB: Extreme gradient boosting; LR: Logistic regression.

Figure 3 Analysis of feature importance using SHapley Additive exPlanations values in the logistic regression model. A: Bar chart displaying the mean SHapley Additive exPlanations (SHAP) values, which indicate the average impact of each feature on the model output; B: Beeswarm plot showing the distribution of SHAP values for individual predictions; C: Summary plot combining SHAP values and feature importance, with feature values represented by color (blue for low and red for high). PVT: Portal vein thrombosis; CLIF-C: Chronic liver failure consortium; ALBI: Albumin-bilirubin; TBIL: Total bilirubin; PPG: Portal pressure gradient; EIS: Endoscopic injection sclerotherapy; CTPV: Cavernous transformation of the portal vein; PVP: Portal venous pressure; ALB: Albumin; SMV: Superior mesenteric vein thrombosis; CLIF-C AD: Chronic liver failure consortium acute decompensation; FIPS: Fibrosis-4 index for liver fibrosis; SHAP: SHapley Additive exPlanations.

Predictive performance and prognostic value of OHE in survival

In Kaplan-Meier analysis, the prediction of OHE was based on the LR model, with a threshold set such that patients with a predicted probability greater than 0.5 were classified as predicted OHE status. Patients without post-TIPS OHE had a longer OS than patients with post-TIPS OHE (log-rank test, P = 0.005) (Figure 4A). Similarly, a longer OS was observed for patients with predicted OHE status by the ML model (log-rank test, P = 0.020) (Figure 4B). In multivariate Cox analysis with stepwise selection, the OHE value predicted by the ML model (P = 0.012) is the prognostic factor in the Cox proportional hazards model. As the same time, the actual OHE status was a prognostic factor in the analysis of OS (P = 0.019) (Figure 5 and Supplementary Tables 1-3), the AUCs of the Cox models with the actual OHE status and predicted OHE value were 90.6 (95%CI: 80.2-100.0) and 88.1 (95%CI: 76.5-99.7), respectively, which were very close to each other and higher than the Cox model based on traditional preoperative data (AUC: 83.0, 95%CI: 67.0-99.1) (Figure 6).

Figure 4 Kaplan-Meier survival curves comparing different overt hepatic encephalopathy status groups. A: Survival analysis based on actual overt hepatic encephalopathy (OHE) status; B: Survival analysis based on predicted OHE status. CI: Confidence interval; HR Hazard ratio; OHE: Overt hepatic encephalopathy.

Figure 5 Multivariate Cox regression analyses of overall survival. A: Cox regression model established with preoperative data; B: Cox regression model established with preoperative data and actual overt hepatic encephalopathy (OHE) status; C: Cox regression model established with preoperative data and predicted OHE value. Na⁺: Sodion; AIC: Akaike information criterion; TBIL: Total bilirubin; OHE: Overt hepatic encephalopathy.

Figure 6 Calibration plot comparing the predicted risk to the actual risk for three different models: actual model (black line), traditional model (yellow line), and machine learning model (blue line). The performance metrics include the area under the curve and Brier score for each model. AUC: Area under the curve; ML: Machine learning.

DISCUSSION

We utilized 5-year long-term follow-up data and applied ML methods to successfully establish a predictive model for post-TIPS OHE. Through survival analysis, we observed that the predicted value of OHE was associated with differences in post-TIPS OS among patients, further validating the model’s value. In this study, we employed EMs for feature selection, reducing the initial set of features to 16. Subsequently, we applied RFE to further refine the features down to 9. Our findings indicate that the Child-Pugh score and age emerged as the most critical predictors, followed by PVT, significantly influencing the model’s output (Figure 3).

The Child-Pugh score is a widely used indicator for assessing patients who underwent TIPS preoperatively, serving as a crucial measure of liver function reserve[12,13]. Numerous studies have demonstrated its correlation with both the occurrence of post-TIPS OHE and patient survival[5,6,8]. Patients with poorer liver function have more neurotoxins bypassing the liver and entering the nervous system post-TIPS, increasing the risk of OHE[14]. However, due to the combination of subjective and objective factors in the Child-Pugh score, using it alone to predict post-TIPS OHE is unreliable. In our ML model, although the Child-Pugh score significantly impacts the model, other scoring systems such as the albumin-bilirubin score and chronic liver failure consortium acute decompensation score also play a role. By incorporating different scores and clinical variables, the ML model can better handle the nonlinear relationships between variables compared to traditional statistical models, resulting in more reliable predictions.

With advancing age, gastrointestinal motility decreases, leading to delayed intestinal emptying and higher incidences of constipation and intestinal flora imbalance. Additionally, the aging brain is more sensitive to toxic metabolites such as plasma ammonia[15]. Moreover, aging affects the overall physiological reserve and liver functional capacity, potentially reducing surgical tolerance and increasing risk[16,17]. These findings highlight the need to consider age during preoperative assessments and postoperative care for TIPS patients.

The formation of PVT results from disrupted hepatic blood flow, often accompanied by further liver function impairment[18]. In patients with PVT, the potential influence on post-TIPS OHE can be attributed to the presence of additional portosystemic shunts beyond the TIPS stent, which reduces portal vein blood flow velocity and subsequently affects the central nervous system. Although some studies have reported an association between PVT and OHE, the specific mechanisms underlying this relationship require further investigation[17,19]. Since our enrolled patients experienced AVB, they did not receive anticoagulant therapy either before or after the TIPS procedure, in principle. There is also controversy regarding anticoagulant therapy following TIPS in different guidelines[4,20-22]. Therefore, in AVB patients with PVT, more detailed clinical examination data and multidisciplinary team discussions may be necessary.

In addition to constructing an ML model to predict the status of OHE, we further investigated the prognostic differences between the predicted OHE value from the ML model and the actual OHE status. This analysis was conducted on the test dataset that was split using random seeds. The predicted OHE value was applied to a classification model to evaluate prognostic differences using Kaplan-Meier curves, revealing that both predicted and actual OHE values significantly influenced OS (Figure 4). In further exploring the prognostic factors, the final Cox models consistently identified age and sodium levels as significant factors affecting post-TIPS outcomes. Older patients face higher mortality risks due to reduced physiological reserves and greater susceptibility to OHE complications. Additionally, hyponatremia, which is common in cirrhotic patients, reflects fluid imbalances that can worsen postoperative outcomes. These findings highlight the importance of managing age and sodium levels to improve long-term survival after TIPS. Both the predicted and actual OHE values were equally effective in predicting patient survival, as demonstrated by comparable model performance (AUC: 0.846/0.798, P = 0.368) (Supplementary Figure 1), indicating no significant difference in the performance of the two models in predicting post-TIPS survival. The predicted OHE value was similar to the actual OHE status in predicting patient survival and performed better than traditional preoperative models, demonstrating the reliability and validity of the ML model. In multivariate Cox analysis, both the predicted OHE value and actual OHE status were important prognostic factors in their respective Cox models (Figure 5 and Supplementary Tables 4-6). Although a recent Italian study found that post-TIPS episodic OHE did not increase mortality in patients with cirrhosis[23], our research indicates that post-TIPS OHE is a significant factor affecting patient survival post-TIPS (Figure 5). Long-term follow-up studies from other centers have also shown that post-TIPS OHE is significantly associated with a higher risk of mortality[24-26]. Given the strong prognostic implications of OHE in post-TIPS survival, it is essential for the medical team to thoroughly communicate the risks, diagnosis, and management of OHE with patients and their families.

For the clinical application of our model, predicting the occurrence of OHE after TIPS allows for the individualized treatment of patients. This enables the provision of optimal treatment plans for individual patients, making the treatment of AVB more personalized. Additionally, the predictive model offers a tool for long-term follow-up, guiding decision-making by identifying patients who will benefit the most from the TIPS procedure and allowing for better prevention and treatment measures for potential OHE. This is where our model has an advantage over previous studies that focused on short-term follow-up after TIPS. Furthermore, the predictive performance of our model is superior compared to other models.

This study had several limitations. First, as a retrospective single-center study, there may have been selection bias. Second, although we used survival analysis to further validate our ML model, the relatively small sample size resulted in a small test set, which may have compromised the stability of the results. Additionally, our study focused on predicting OHE and did not consider other complications occurring over a longer period. Lastly, some procedural features were missing in this study, such as specific physical parameters of the stent and the stent implantation angle. Certain potential risk factors, such as sarcopenia[27], which are believed to influence the prognosis of patients undergoing TIPS for AVB, were not included in this study. Future research could incorporate multicenter data and include more clinical course information to further improve the model.

CONCLUSION

In conclusion, our study demonstrated that our model exhibited high and stable performance in predicting the occurrence of OHE following TIPS in patients with AVB. The predicted OHE value was identified as an independently significant predictor of OS in validation cohorts. The ML model demonstrated promising potential as a computer-assisted diagnostic tool, enabling clinicians to more accurately detect post-TIPS OHE and thus inform and optimize patient treatment plans.

ACKNOWLEDGEMENTS

We thank Professor Wang JY for his support and supervision of this study.