Development of a deep learning model for guiding treatment decisions of acute variceal bleeding in patients with cirrhosis

Abstract

BACKGROUND

Acute variceal bleeding (AVB) in patients with cirrhosis remains life-threatening; moreover, the current risk stratification methods have certain limitations. Rebleeding and mortality after AVB remain major challenges. Although preemptive transjugular intrahepatic portosystemic shunt (p-TIPS) can improve outcomes, not all patients benefit equally. Accurate risk stratification is needed to guide treatment decisions and identify those most likely to benefit from p-TIPS.

AIM

To develop an artificial intelligence (AI)-driven model to guide AVB treatment decisions, and identify candidates eligible for p-TIPS.

METHODS

Patients with cirrhosis and AVB, from two multicenter retrospective cohorts in China, who received endoscopic variceal ligation plus pharmacotherapy (n = 1227) or p-TIPS (n = 1863) were included. Baseline data within 24 hours of hospital admission were obtained. The AI-AVB model, based on the six-week failure and one-year mortality rates, was developed to predict treatment efficacy and compared with standard risk scores. Outcomes and adverse events of the treatments were compared across the high- and low-risk subgroups stratified using the AI-AVB model.

RESULTS

The AI-AVB model demonstrated superior predictive performance compared to traditional risk stratification methods. In the internal validation cohort, the model achieved an area under the curve (AUC) of 0.842 for predicting six-week treatment failure and 0.954 for one-year mortality. In the external validation cohort, the AUCs were 0.814 and 0.889, respectively. The model effectively identified patients at high risk of first-line treatment failure who may benefit from aggressive interventions such as p-TIPS. In contrast, advancing the treatment strategy for low-risk patients did not notably improve the short-term prognosis.

CONCLUSION

The AI-AVB model can predict treatment outcomes, stratify the failure risk in cirrhotic patients with AVB, aid in clinical decisions, identify p-TIPS beneficiaries, and optimize personalized treatment strategies.

Key Words: Acute variceal bleeding; Liver cirrhosis; Deep learning; Risk stratification; Endoscopic therapy; Preemptive transjugular intrahepatic portosystemic shunt

Core Tip: A novel deep learning model was developed to predict treatment outcomes in patients with acute variceal bleeding, a life-threatening condition that is often observed in patients with cirrhosis. By analyzing clinical data collected within 24 hours of hospital admission, the artificial intelligence model can effectively identify high-risk patients who may benefit from more aggressive treatments, such as preemptive transjugular intrahepatic portosystemic shunt, while also helping avoid unwarranted invasive procedures for low-risk patients.

INTRODUCTION

Acute variceal bleeding (AVB) is a severe and life-threatening complication of portal hypertension frequently observed in patients with cirrhosis[1-4]. Despite significant advances in the management of AVB over recent years, it continues to be associated with substantial mortality[5,6]. The primary treatment modalities for AVB include endoscopy combined with pharmacological treatment and transjugular intrahepatic portosystemic shunt (TIPS)[7]. Endoscopy, often involving band ligation, along with vasoactive drugs, is the first-line treatment for controlling bleeding[3,8]. However, TIPS is considered for patients who are at high risk of treatment failure with endoscopic and pharmacological interventions[9,10]. Owing to the varying severity of bleeding and survival prognosis among patients with AVB, clinicians typically tailor or optimize treatment strategies based on the expected risk[11,12]. High-risk patients, who have a poor prognosis, require more aggressive and invasive interventions, such as preemptive TIPS (p-TIPS), to improve their survival outcomes[13]. Conversely, patients with a favorable prognosis can often be managed with less invasive and safer therapeutic options, thereby avoiding unwarranted procedures[14]. Individualized management based on these predictive factors has demonstrated improved survival outcomes and reduced rates of rebleeding and treatment failure.

The most widely adopted risk stratification criteria for AVB management are those outlined in the Baveno VII consensus[15]. According to the Baveno VII criteria, patients with a hepatic venous pressure gradient (HVPG) of ≥ 20 mmHg, or a Child-Pugh class C (10-13 points) or B (8-9 points) with active bleeding are classified as high risk for treatment failure. While these criteria are intended to guide the use of aggressive interventions[16,17], they are not without limitations. First, the HVPG measurement is an invasive, risky, and expensive procedure requiring specialized equipment and expertise. Hence, obtaining HVPG measurements promptly in emergency settings is challenging, limiting their clinical accessibility[4,18]. Secondly, although the Child-Pugh classification is useful, it includes variables such as ascites and hepatic encephalopathy (HE) that can vary between observers, introducing potential variability and bias[7]. Lastly, the evaluation of active bleeding during endoscopy is highly dependent on the operator’s skill and experience, resulting in substantial subjectivity and heterogeneity in assessments[19,20]. Consequently, these factors contribute to the complexity and reduced accuracy of risk stratification for AVB using the Baveno VII criteria. Yoo et al[21] reported that the Baveno VII criteria achieved an area under the curve (AUC) of 0.70 [95% confidence interval (CI): 0.66-0.75] for predicting varices requiring treatment, indication moderate discriminatory performance in real-world settings.

In recent years, several alternative scoring systems have been proposed to refine the early TIPS criteria for AVB in patients with cirrhosis. These include the model for end-stage liver disease (MELD) score[22] with a threshold of > 19, the recalibrated MELD model or MELD-HE score[23], and the chronic liver failure-consortium acute decompensation score[14]. Among these, the MELD score and its recalibrated versions have demonstrated better discriminative performances, aiding in the identification of patients with AVB who may benefit from p-TIPS[24]. Buckholz et al[25] reported that the MELD score achieved an AUC of 0.76 (95%CI: 0.70-0.82) for predicting six-week mortality, reflecting moderate discriminatory ability[25]. However, the MELD score components, bilirubin levels, international normalized ratio, and creatinine levels mainly reflect liver and renal functions, omitting other factors that may improve prediction. Consequently, the current scoring systems may inadvertently misclassify patients, leading to treatment failure in ostensibly ‘non-high-risk’ patients owing to insufficient intervention[7] and unnecessary p-TIPS in others, increasing the risk of overt HE without survival benefits[9,12,26]. These limitations highlight the need for a more objective, accessible, accurate, and cost-effective tool to guide treatment decisions in complex AVB cases.

With the rapid advancements in artificial intelligence (AI) and medical informatics, deep learning technology has emerged as a powerful data-centric approach, demonstrating remarkable potential in the healthcare sector[27]. Notably, deep learning shows significant promise in predicting treatment outcomes in patients with AVB[28]. The complex nature of AVB, coupled with the myriad factors that influence patient prognosis, often renders traditional predictive models inadequate. The ability of deep learning to analyze complex datasets and identify patterns offers a promising avenue for enhancing predictive accuracy in this field[29]. Specifically, deep learning algorithms can be trained on a variety of variables, including clinical parameters, laboratory results, and endoscopic findings[30], to generate models that predict the likelihood of various treatment outcomes, such as rebleeding, complications, or mortality[31].

This study aimed to develop and validate a deep learning model using baseline data from patients with cirrhosis-related AVB at initial admission to stratify the risks of rebleeding and mortality, thereby guiding treatment decisions and identifying candidates likely to benefit from p-TIPS.

MATERIALS AND METHODS

Study population

This study included two distinct cohorts of patients with cirrhosis-associated AVB. The first cohort was part of a multicenter retrospective study involving 30 tertiary hospitals across 18 provinces in China (Supplementary Table 1). This study included patients who received endoscopic variceal ligation (EVL) combined with pharmacotherapy (PT) between February 2013 and May 2020. Comprehensive demographic and clinical data were collected, and one-year treatment and survival outcomes were recorded. The second cohort included a multicenter retrospective study of patients with cirrhosis-associated AVB who underwent p-TIPS treatment. This study was conducted from January 2010 to June 2020 in six tertiary hospitals in China, with patient follow-up extending until death, liver transplantation, or up to 2 years, whichever came first.

This study adhered to the ethical principles of the Declaration of Helsinki, and the study protocols were approved by the Ethics Committee of Zhongda Hospital, Southeast University. Given the retrospective nature of this study, the requirement for informed consent was waived.

Inclusion and exclusion criteria

The inclusion criteria for both the training and validation cohorts were as follows: (1) Age between 18 and 80 years; (2) A confirmed diagnosis of cirrhosis based on liver biopsy or a combination of clinical, biochemical, and imaging findings; and (3) Endoscopic confirmation of AVB[8]. Exclusion criteria were: (1) Severe heart failure, chronic obstructive pulmonary disease; (2) Complete portal vein thrombosis, recurrent HE, heart failure, previous liver transplantation, and previous TIPS; (3) Hepatocellular carcinoma exceeding the Milan criteria or classified as advanced/terminal stage; (4) Previous endoscopic injection sclerotherapy or EVL within 3 months prior to the study; (5) Possibility of a bleeding source beyond the esophageal or gastro-esophageal varices; (6) Therapies other than the EVL + PT combination (e.g., EVL/PT monotherapy) for management; and (7) Incomplete or missing data. All enrolled patients adhered to a restrictive transfusion protocol and received prophylactic antibiotics, as clinically indicated[32]. Variables with > 20% missing data were excluded. For those with ≤ 20% missing data, the missing values were imputed using the median for continuous variables and the mode for categorical variables.

As this study aimed to identify subgroups benefiting from p-TIPS, extremely high-risk patients who may not meet the criteria for p-TIPS treatment were excluded, i.e., those with uncontrollable initial bleeding, for whom TIPS may be ineffective, or even those for whom salvage TIPS becomes necessary (according to the Baveno criteria, such as patients with a Child-Pugh score ≥ 14 or a MELD score > 30)[33]. P-TIPS is defined as the placement of covered TIPS within 72 hours of admission as a preventive treatment before rebleeding occurs following EVL + PT therapy. Patients who had undergone p-TIPS were excluded during the model development and validation phases but were included in the evaluation of the model’s p-TIPS prediction efficiency. A flow diagram of the study population is illustrated in Figure 1.

Figure 1 Flowchart of study population. AVB: Acute variceal bleeding; AI: Artificial intelligence; EIS: Endoscopic injection sclerotherapy; EVL: Endoscopic variceal ligation; TIPS: Transjugular intrahepatic portosystemic shunt; p-TIPS: Preemptive transjugular intrahepatic portosystemic shunt; PT: Pharmacotherapy.

Endpoints

The primary endpoint of the study was the rate of treatment failure within 6 weeks following either EVL + PT or p-TIPS treatment. Treatment failure was defined as rebleeding, including evidence of gastrointestinal hemorrhage such as hematemesis, melena, or hematochezia, or endoscopic evidence of recurrent bleeding. Secondary endpoints included one-year mortality (from any cause, in the absence of liver transplantation), requirement for intensive care unit (ICU) admission, and treatment-related adverse events, such as the onset or worsening of ascites and HE.

Cohorts for training, internal validation, and external validation

A total of 1227 patients with AVB receiving EVL + PT therapy were allocated to the training, internal validation, and external validation cohorts. Of these, 835 patients from the Fifth Medical Center of the Chinese PLA General Hospital were randomly assigned to either a training cohort (668 patients) or an internal validation cohort (167 patients) in a 4:1 ratio. The remaining 392 patients from 29 tertiary medical centers constituted the external validation cohort. Random assignment was performed using the scikit-learn package in Python with the ‘train test split’ function, to ensure balanced representation between the training and internal validation cohorts. The external validation cohort remained independent when assessing the generalizability of the model across different institutions.

AI-AVB model development

The proposed machine learning model, AI-AVB, is primarily based on the residual structure, pyramid architecture, and multilevel feature fusion strategy. It extracts the features of 22 input variables to generate prediction results for each endpoint; the overall architecture is shown in Figure 2A and B.

Figure 2 Workflow of the development and testing of artificial intelligence-acute variceal bleeding model. A: Overview of the prediction pipeline; B: Architecture of artificial intelligence-acute variceal bleeding; C: Architecture of atrous spatial pyramid pooling; D: Architecture of blocks. SBP: Systolic blood pressure; HR: Heart rate; AVB: Acute variceal bleeding; AI: Artificial intelligence; P: Positive; N: Negative; ICU: Intensive care unit; AUC: Area under the curve; RB: Residual block; CB: Convolution block; LB: Linear block; ASPP: Atrous spatial pyramid pooling; ReLU: Rectified linear unit; Conv: Convolution; Norm: Normalization; Drop: Dropout.

The initial input into the AI-AVB included 22 variables that were processed in three residual stages. Each residual stage was primarily comprised a complex one-dimensional residual block and a one-dimensional pooling layer. The three stages had two, three, and six residual blocks, respectively. The structure of the residual blocks is shown in Figure 2C and D. A shortcut was used to superimpose input features onto features generated by three one-dimensional convolution sets with kernel sizes of 1, 3, and 1. This cross-layered connection structure can limit gradient disappearance and network degradation in deep networks. At the end of each residual module, a dropout layer was appended to limit model overfitting.

After the residual-stage process, the third-stage features were input into the modified atrous spatial pyramid pooling (ASPP). The ASPP comprised four sets of one-dimensional atrous convolutions with dilation rates of 1, 3, 5, and 7, respectively. These atrous convolutions improve the multiscale context feature extraction ability while expanding the field of vision of AI-AVB. The features were first input into four sets of atrous convolutions and then concatenated. After four linear blocks and two pooling layers, features fused with multiscale context information were generated.

The multiscale features generated by the three residual stages and the ASPP features were fused using three sets of feature fusion blocks. Each set of feature fusion blocks included the corresponding residual stage and the last fusion features. Each fusion block adopted several convolution and pooling layers to down sample the residual-stage features. The down sampled features were superimposed with the last fusion feature and then integrated using a linear block to generate fusion features.

At the end of AI-AVB, the ASPP features and three groups of fusion features were concatenated and integrated using linear layers. A softmax layer was then used to generate the prediction probabilities for all classes of each endpoint.

Training strategy

To improve the anti-overfitting ability of the AI-AVB, we applied online data augmentation to the training cohort. The augmentation probability was initially 0.5 and was adjusted based on the decline curves of the training and validation cohorts. Augmentation methods primarily included random missing data, random variable vibration, and random label vibration.

To prevent AI-AVB from falling into a local optimum, cosine annealing was used to adjust the learning rate. The initial learning rate was 0.1, the minimum value was 0.0002, and the warm restart period was 8 epochs. We adopted an early stop strategy for training. The default number of epochs was 80. In terms of hyperparameters, the dropout rate was 0.4. The loss function was the crossing entropy loss, which is described as where is the number of classes, is the actual value of class, and is the predicted value. After calculating the loss value, chain derivation was performed according to the loss value, and the AdamW optimizer was used to update the model parameters.

Comparison with machine learning models

To demonstrate the effectiveness and advancement of the proposed AI-AVB method, we compared and analyzed six frequently used machine learning methods based on our dataset with AI-AVB. These methods included support vector machine[34], logistic regression[35], decision tree[36], random forest[37], extreme gradient boosting[38], and light gradient boosting machine[39]. To ensure fairness of comparisons, the same training, validation, and test cohorts were used to train and test the AI-AVB and the six existing methods. All six methods used Optuna to automatically adjust hyperparameters according to the performance of the validation cohort, with the AUC as the primary evaluation metric to guide tuning. Accuracy was used as a secondary reference when AUC values were identical[40]. Each model was optimized using Optuna’s Tree-structured Parzen Estimator sampler over 100 trials. Search spaces included model-specific parameters such as learning rate (0.01-0.3), maximum tree depth (3-10), and number of estimators (50-500) for tree-based algorithms, and regularization strength (0.01-10) for linear models. All experiments were conducted using a fixed random seed (123) to ensure stability and reproducibility.

Efficacy assessment of different treatments in patients with AVB

Using the previously developed and validated AI-AVB model, all patients with cirrhosis and AVB who underwent EVL + PT and p-TIPS were stratified into high- and low-risk groups for EVL + PT failure. Within these high- and low-risk subgroups, we compared the outcomes and adverse effects associated with the two treatment strategies. The outcomes assessed were as follows: (1) The rate of treatment failure within 6 weeks; (2) One-year mortality; (3) ICU admission; and (4) The development or worsening of ascites and HE.

Further analysis involved both univariate and multivariate regression models to investigate the independent predictive value of baseline characteristics, Child-Pugh scores, MELD scores, and treatment modalities for six-week treatment failure and survival at discharge in both the high- and low-risk groups for standard therapy failure.

Statistical analysis

Normally distributed continuous variables are expressed as mean ± SD and compared using Student’s t-test, while the non-normally distributed continuous variables are presented as medians (interquartile ranges) and analyzed using the Mann-Whitney test. Categorical variables are represented as counts and percentages, and compared using the χ² test or Fisher’s exact test. The performance metrics of the proposed and comparative methods were evaluated using a receiver operating characteristic curve, confusion matrix, and the following indices: AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The Delong test was used to compare the AUCs of the other models with that of AI-AVB.

RESULTS

Baseline characteristics of patients and outcomes

A total of 1227 patients with cirrhosis and AVB were included in the study and underwent EVL + PT. Of these, 835 patients from the Fifth Medical Center of the Chinese PLA General Hospital were randomly allocated into a training cohort (668 patients) and an internal validation cohort (167 patients) in a 4:1 ratio. The remaining 392 patients from 29 tertiary medical centers were designated as the external validation cohort. The demographic characteristics, initial clinical conditions, laboratory results, liver-related risk scores, treatment modalities, and outcomes for the three cohorts are presented in Table 1.

Table 1 Baseline characteristics and outcomes of acute variceal bleeding patients in the training, internal validation and external validation cohort, mean ± SD/n (%).

Characteristics	Training cohort (n = 668)	Internal validation cohort (n = 167)	P value	External validation cohort (n = 392)
Demographic characteristics
Age (years)	52.5 ± 11.8	52.6 ± 10.2	0.927	53.1 ± 11.8
Sex			0.273
Male	473 (70.8)	126 (75.4)		278 (70.9)
Female	195 (29.2)	41 (24.6)		114 (29.1)
Etiology of cirrhosis			0.394
Chronic HBV infection	401 (60.0)	117 (70.1)		246 (62.8)
Chronic HCV infection	41 (6.1)	10 (6.0)		28 (7.1)
Alcohol	73 (10.9)	16 (9.6)		56 (14.3)
Others	92 (13.8)	14 (8.3)		35 (8.9)
Cryptogenic	61 (9.1)	10 (6.0)		27 (6.9)
Medical history
Previous variceal bleeding	240 (35.9)	65 (38.9)	0.407	163 (41.6)
Location of varices			0.299
Esophageal varices only	388 (58.0)	105 (62.9)		214 (54.6)
Esophageal and gastric varices	280 (42.0)	62 (37.1)		178 (45.4)
Hepatic encephalopathy	65 (9.7)	17 (10.2)	0.437	33 (8.4)
Ascites			0.502
Mild	237 (35.4)	76 (45.5)		142 (36.2)
Moderate	96 (14.4)	31 (18.6)		68 (17.4)
Massive	45 (6.7)	13 (7.8)		30 (7.7)
Heart rate at admission (beats/minute)	85.0 ± 15.6	85.7 ± 16.6	0.632	87.5 ± 15.5
Systolic blood pressure at admission (mmHg)	115.8 ± 17.5	116.5 ± 16.2	0.640	113.6 ± 19.8
Diastolic blood pressure at admission (mmHg), median IQR	71.0 (63-79)	71.0 (62.5-79.5)	0.746	64.0 (55.5-72.5)
Laboratory examination
White blood cell (× 109 cell/L), median IQR	7.3 (5.4-10.0)	7.6 (6.2-9.0)	0.052	7.6 (5.3-9.7)
Red blood cell (× 109 cell/L)	2.7 ± 0.8	2.8 ± 0.7	0.483	2.3 ± 0.6
Hemoglobin (g/L)	77.3 ± 22.3	78.5 ± 23.4	0.521	70.8 ± 24.8
Platelet count (× 109/L), median IQR	84.0 (61.0-117.0)	85.0 (62.5-128.0)	0.867	94.5 (69.0-129.0)
NEC (× 109/L), median IQR	5.7 (4.3-8.0)	5.7 (4.3-8.2)	0.674	6.3 (3.8-12.5)
AST (U/L), median IQR	94.0 (83.7-118.2)	104.0 (93.0-125.5)	0.561	113.0 (103.5-133.0)
ALT (U/L), median IQR	54.0 (46.0-67.0)	53.0 (46.0-62.5)	0.681	56.0 (49.0-70.0)
TBIL (μmol/L)	38.0 ± 63.2	34.4 ± 38.0	0.577	33.0 ± 33.2
Albumin (g/L)	28.5 ± 5.4	29.0 ± 5.0	0.267	31.7 ± 8.8
INR	1.4 ± 2.7	1.3 ± 0.3	0.606	1.6 ± 1.5
APTT (second)	38.4 ± 16.3	36.6 ± 9.8	0.176	37.2 ± 13.4
TT (second)	19.9 ± 3.6	19.4 ± 3.2	0.124	19.8 ± 5.8
PT (second), median IQR	14.2 (12.9-15.8)	13.9 (12.9-15.6)	0.269	15.4 (13.9-17.9)
Creatinine (μmol/L), median IQR	92.0 (81.0-107.0)	90.0 (80.0-102.5)	0.814	86.0 (75.0-105.0)
Risk stratification index
MELD score (points)	11.3 ± 3.2	12.1 ± 5.0	0.049	12.2 ± 3.7
Child-Pugh score (points), median IQR	8.0 (7.0-9.0)	8.0 (7.0-9.1)	0.797	7.0 (5.5-8.5)
Child-Pugh class			0.395
A (5-6)	142 (21.3)	29 (17.4)		125 (31.9)
B (7-9)	407 (60.9)	111 (66.5)		180 (45.9)
C (10-13)	119 (17.8)	27 (16.2)		87 (22.2)
Early TIPS criteria			0.745
Low risk	534 (79.9)	136 (81.4)		280 (71.4)
High risk	134 (20.1)	31 (18.6)		112 (28.6)
Treatment procedure
Initial pharmacological therapy			0.275
Octreotide	177 (26.5)	47 (28.1)		90 (22.9)
Somatostatin	403 (60.3)	91 (54.5)		255 (65.1)
Terlipressin	88 (13.2)	29 (17.4)		47 (12.0)
Outcome measurements
6-week treatment failure to control bleeding	78 (11.8)	22 (13.2)	0.113	55 (14.0)
ICU requirement	155 (23.2)	38 (22.8)	0.785	81 (20.7)
1-year mortality	126 (18.8)	30 (18.0)	0.261	67 (17.1)
Treatment-related adverse events
Hepatic encephalopathy	86 (12.9)	26 (15.6)	0.093	74 (18.9)
New or worsening ascites	18 (2.7)	3 (1.8)	0.194	10 (2.6)

HBV: Hepatitis B virus; HCV: Hepatitis C virus; IQR: Interquartile range; NEC: Neutrophil count; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; TBIL: Total bilirubin; INR: International normalized ratio; APTT: Activated partial thromboplastin time; TT: Thrombin time; PT: Prothrombin time; MELD: Model for end-stage liver disease; TIPS: Transjugular intrahepatic portosystemic shunt; ICU: Intensive care unit.

The mean age of patients in the training, internal validation, and external validation cohorts was 52.5 ± 11.8 years, 52.6 ± 10.2 years, and 53.1 ± 11.8 years, respectively. Most patients in the three cohorts were males (70.8%, 75.4%, and 70.9%, respectively). Chronic hepatitis B virus (HBV) infection was the most common etiology of cirrhosis in all cohorts.

The six-week treatment failure rates were 11.8%, 13.2%, and 14.0% in the training, internal validation, and external validation cohorts, respectively. The one-year mortality rates for the three cohorts were 18.8%, 18.0%, and 17.1%, respectively.

Evaluation and the internal and external validation of model performance

An AI-AVB model was developed to predict the likelihood of treatment failure within 6 weeks and mortality in 1 year following EVL + PT treatment in patients with cirrhosis and AVB. For the six-week treatment failure prediction in the internal validation cohort, the AI-AVB model achieved an AUC of 0.842, an accuracy of 0.940, a PPV of 0.455, and an NPV of 0.974. In the external validation cohort, the AI-AVB model demonstrated an AUC of 0.814, accuracy of 0.939, PPV of 0.278, and NPV of 0.971 (Figure 3 and Supplementary Table 2).

Figure 3 Receiver operating characteristic curves and confusion matrix of different models for clinical outcomes in acute variceal bleeding patients. A-C: Receiver operating characteristic curves of seven models for 6-week treatment failure, 1-year mortality, intensive care unit (ICU) requirement, respectively, in the internal and external validation cohort; D-F: Confusion matrix for 6-week treatment failure, 1-year mortality, ICU requirement, respectively, in the internal and external validation cohort using artificial intelligence-acute variceal bleeding algorithms. True labels on the vertical axis and predicted labels on the horizontal axis. ICU: Intensive care unit; AUC: Area under the curve; SVM: Support vector machine; LR: Logistic regression; DT: Decision tree; RF: Random forest; XGB: Extreme gradient boosting; LGBM: Light gradient boosting machine; AVB: Acute variceal bleeding; AI: Artificial intelligence.

To predict one-year mortality in the internal validation cohort, the AI-AVB model achieved an AUC of 0.954, an accuracy of 0.964, a PPV of 0.625, and an NPV of 0.981. In the external validation cohort, the model demonstrated an AUC of 0.889, accuracy of 0.959, PPV of 0.455, and NPV of 0.984 (Supplementary Table 3). Regarding ICU admission after treatment, the AI-AVB model had an AUC of 0.866 and 0.812 for the internal and external validation cohorts, respectively (Supplementary Table 4 and Supplementary Figures 1-3).

Improved short-term outcomes in patients with high-risk AVB receiving p-TIPS

As shown in Table 2, for high-risk patients, the six-week treatment failure rate was significantly lower in the p-TIPS group (3.8%) than in the EVL + PT group (39.6%) (Table 2). Similarly, the one-year mortality rate was markedly lower in the p-TIPS group (14.6%) than in the EVL + PT group (37.3%) (Figure 4). These findings indicate that p-TIPS significantly improved short-term outcomes in high-risk patients.

Figure 4 Cumulative incidence of treatment outcomes in high or low-risk acute variceal bleeding patients. A: 6-week treatment failure in high or low-risk patients; B and C: 1-year survival rates in high and low risk patients. EVL: Endoscopic variceal ligation; p-TIPS: Preemptive transjugular intrahepatic portosystemic shunt; PT: Pharmacotherapy.

Table 2 Baseline characteristics and outcomes of acute variceal bleeding patients treated with endoscopic variceal ligation plus pharmacotherapy or preemptive transjugular intrahepatic portosystemic shunt, mean ± SD/n (%).

Characteristics	High-risk for EVL + PT failure			Low-risk for EVL + PT failure
	EVL + PT (n = 220)	p-TIPS (n = 704)	P value	EVL + PT (n = 1007)	p-TIPS (n = 1159)	P value
Demographic characteristics
Age (years)	52.4 ± 10.9	52.6 ± 11.7	0.167	53.3 ± 11.8	51.5 ± 12.5	0.042
Sex			0.011			0.020
Male	162 (73.6)	457 (64.9)		715 (71.0)	729 (62.9)
Female	58 (26.4)	247 (35.1)		292 (29.0)	430 (37.1)
Etiology of cirrhosis			0.072			0.005
Chronic HBV infection	154 (70.0)	476 (67.6)		610 (60.6)	761 (65.7)
Chronic HCV infection	6 (2.7)	35 (5.0)		73 (7.2)	100 (8.6)
Alcohol	23 (10.5)	63 (8.9)		122 (12.1)	92 (7.9)
Others	18 (8.1)	92 (13.1)		123 (12.2)	133 (11.5)
Cryptogenic	19 (8.6)	38 (5.4)		79 (7.8)	73 (6.3)
Medical history
Previous variceal bleeding	93 (42.3)	319 (45.3)	0.537	375 (37.2)	447 (38.6)	0.618
Location of varices			0.692			0.544
Esophageal varices only	131 (59.5)	432 (61.4)		576 (57.2)	679 (58.6)
Esophageal and gastric varices	89 (40.5)	272 (38.6)		431 (42.8)	480 (41.4)
Hepatic encephalopathy	32 (14.5)	116 (16.4)	0.236	83 (8.2)	105 (9.1)	0.280
Ascites			0.021			0.066
Mild	89 (40.5)	263 (37.3)		366 (36.3)	395 (34.1)
Moderate	42 (19.1)	144 (20.4)		153 (15.2)	220 (19.0)
Massive	9 (4.1)	72 (10.2)		79 (7.8)	104 (9.0)
Heart rate at admission (beats/minute)	84.4 ± 14.4	83.1 ± 15.3	0.384	87.8 ± 16.4	86.8 ± 15.7	0.033
Systolic blood pressure at admission (mmHg)	112.0 ± 12.9	113.8 ± 17.9	0.067	117.4 ± 14.5	115.8 ± 16.0	0.003
Diastolic blood pressure at admission (mmHg)	67.5 ± 10.8	70.6 ± 12.4	0.002	69.5 ± 9.9	68.1 ± 11.5	0.042
Laboratory examination
White blood cell (× 109 cell/L)	8.4 ± 5.9	8.0 ± 3.9	0.246	6.0 ± 3.9	6.3 ± 4.4	0.095
Red blood cell (× 109 cell/L)	2.5 ± 0.7	2.4 ± 0.6	0.387	2.8 ± 0.7	3.5 ± 7.3	0.003
Hemoglobin (g/L)	75.2 ± 22.8	72.6 ± 22.6	0.038	77.7 ± 22.3	82.5 ± 24.9	< 0.001
Platelet count (× 109/L)	87.0 ± 57.8	86.4 ± 54.3	0.888	95.0 ± 89.0	97.6 ± 86.6	0.492
NEC (× 109/L)	7.6 ± 8.8	6.8 ± 5.7	0.115	5.1 ± 8.4	4.7 ± 5.9	0.196
AST (U/L)	124.9 ± 233.5	106.1 ± 167.5	0.189	76.1 ± 493.4	50.3 ± 99.0	0.082
ALT (U/L)	62.0 ± 101.9	55.6 ± 40.8	0.176	49.6 ± 214.2	34.9 ± 40.8	0.022
TBIL (μmol/L)	104.1 ± 120.8	60.2 ± 49.5	< 0.001	23.7 ± 23.8	27.3 ± 25.4	< 0.001
Albumin (g/L)	25.2 ± 4.6	25.6 ± 6.9	0.421	29.2 ± 5.3	32.9 ± 8.7	< 0.001
INR	1.6 ± 0.5	2.1 ± 1.8	< 0.001	1.4 ± 2.9	1.4 ± 1.4	0.898
APTT (second)	49.2 ± 19.1	46.2 ± 19.7	0.047	36.0 ± 14.6	35.3 ± 10.9	0.203
TT (second)	21.4 ± 3.8	22.1 ± 7.0	0.156	19.5 ± 3.5	19.3 ± 5.4	0.315
PT (second)	18.5 ± 6.7	23.9 ± 20.3	0.001	14.2 ± 3.5	15.6 ± 3.4	< 0.001
Creatinine (μmol/L)	100 ± 82.9	86.8 ± 58.4	0.009	82.5 ± 70.1	73.6 ± 44.1	< 0.001
Risk stratification index
MELD score (points)	13.5 ± 4.0	15.9 ± 6.5	0.002	11.2 ± 3.6	12.5 ± 3.8	0.034
Child-Pugh score (points), median IQR	7.9 (5.1-10.7)	8.2 (6.1-10.3)	0.068	7.4 (5.5-9.3)	7.2 (5.6-8.8)	< 0.001
Child-Pugh class			0.167			< 0.001
A (5-6)	42 (19.2)	101 (14.4)		254 (25.2)	336 (29.0)
B (7-9)	131 (59.5)	424 (60.3)		567 (56.3)	719 (62.1)
C (10-13)	47 (21.4)	179 (25.4)		186 (18.5)	104 (8.9)
Early TIPS criteria			< 0.001			< 0.001
Low risk	161 (73.2)	390 (55.4)		789 (78.4)	707 (61.0)
High risk	59 (26.8)	314 (44.6)		218 (21.6)	452 (39.0)
Treatment procedure
Initial pharmacological therapy			0.087			< 0.001
Octreotide	283 (28.1)	314 (27.1)		31 (14.1)	140 (19.9)
Somatostatin	597 (59.3)	730 (63.0)		152 (69.1)	513 (72.9)
Terlipressin	127 (12.6)	115 (9.9)		37 (16.8)	51 (7.2)
p-TIPS stent diameter
< 8 mm		137 (19.5)			61 (5.3)
8 mm		554 (78.2)			1025 (88.4)
10 mm		16 (2.3)			73 (6.3)
Outcome measurements
6-week treatment failure to control bleeding	87 (39.6)	27 (3.8)	< 0.001	68 (6.8)	22 (1.9)	0.013
ICU requirement	102 (46.4)	144 (20.5)	0.005	172 (17.1)	151 (13.0)	0.049
1-year mortality	82 (37.3)	103 (14.6)	< 0.001	141 (14.0)	121 (10.4)	0.036
Treatment-related adverse events
Hepatic encephalopathy	36 (16.4)	285 (30.5)	< 0.001	150 (14.9)	314 (27.1)	< 0.001
New or worsening ascites	12 (5.5)	10 (1.4)	< 0.001	19 (1.9)	10 (0.9)	< 0.001

EVL: Endoscopic variceal ligation; PT: Pharmacotherapy; p-TIPS: Preemptive transjugular intrahepatic portosystemic shunt; HBV: Hepatitis B virus; HCV: Hepatitis C virus; IQR: Interquartile range; NEC: Neutrophil count; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; TBIL: Total bilirubin; INR: International normalized ratio; APTT: Activated partial thromboplastin time; TT: Thrombin time; PT: Prothrombin time; MELD: Model for end-stage liver disease; TIPS: Transjugular intrahepatic portosystemic shunt; ICU: Intensive care unit.

Regarding treatment-related adverse events, HE was more prevalent in the p-TIPS group, with an incidence of 30.5%, compared to 16.4% in the EVL + PT group. Conversely, the incidence of new or worsening ascites was lower in the p-TIPS group (1.4%) than that in the EVL + PT group (5.5%).

The univariate and multivariate regression analyses (Table 3) further supported these findings. In the high-risk subgroup, different treatment strategies (EVL + PT vs p-TIPS) were significantly associated with mortality, with p-TIPS revealing a hazard ratio (HR) of 2.377 (95%CI: 1.855-2.842, P < 0.001) in the univariate analysis and an adjusted HR of 1.671 (95%CI: 1.096-2.259, P = 0.029) in the multivariate analysis.

Table 3 Univariate and multivariate regression analysis for composite outcomes in acute variceal bleeding patients treated with endoscopic variceal ligation plus pharmacotherapy or preemptive transjugular intrahepatic portosystemic shunt, mean ± SD/n (%).

Variables	High-risk patients for EVL + PT failure				Low-risk patients for EVL + PT failure
	Univariate analysis		Multivariate analysis		Univariate analysis		Multivariate analysis
	HR (95%CI)	P value	HR (95%CI)	P value	HR (95%CI)	P value	HR (95%CI)	P value
1-year mortality
Age (years, ≥ 60 vs < 60)	1.125 (0.924-1.371)	0.241			1.312 (1.009-1.706)	0.043	1.189 (0.841-1.680)	0.328
Sex (male vs female)	1.095 (0.909-1.320)	0.337			0.981 (0.765-1.257)	0.879
Etiology of cirrhosis	1.123 (0.892-1.414)	0.325			0.792 (0.587-1.068)	0.127
Location of varices	0.997 (0.808-1.229)	0.974			0.802 (0.609-1.055)	0.115
Hepatic encephalopathy (yes vs no)	1.464 (1.204-1.781)	< 0.001	1.132 (0.875-1.465)	0.345	1.369 (1.066-1.757)	0.014	1.027 (0.738-1.429)	0.876
Ascites (yes vs no)	1.784 (1.421-2.241)	< 0.001	1.409 (1.134-1.957)	< 0.001	2.282 (1.648-3.160)	< 0.001	2.088 (1.538-2.834)	< 0.001
Systolic blood pressure at admission (≥ 120 mmHg vs < 120 mmHg)	1.019 (0.847-1.227)	0.84			1.054 (0.829-1.340)	0.668
Diastolic blood pressure at admission (≥ 90 mmHg vs < 90 mmHg)	1.353 (0.779-2.348)	0.283			0.777 (0.458-1.316)	0.348
MELD score (> 13 vs ≤ 13)	4.968 (3.543-6.967)	< 0.001	1.320 (1.046-1.665)	0.006	5.284 (3.331-8.381)	< 0.001	2.208 (1.155-3.504)	< 0.001
Child-Pugh class (C vs A + B)	4.137 (1.977-7.926)	< 0.001	3.309 (2.535-4.554)	< 0.001	-	0.996
Different treatment strategies (EVL + PT vs p-TIPS)	2.377 (1.855-2.842)	< 0.001	1.671 (1.096-2.259)	0.029	0.954 (0.671-1.355)	0.791
6-week treatment failure
Age (years, ≥ 60 vs < 60)	1.251 (1.170-1.363)	0.009	1.090 (0.898-1.322)	0.376	1.225 (1.075-1.526)	0.028	1.128 (0.870-1.389)	0.437
Sex (male vs female)	1.087 (0.917-1.272)	0.346			0.997 (0.792-1.238)	0.947
Etiology of cirrhosis	1.134 (0.917-1.392)	0.248			0.821 (0.631-1.068)	0.144
Location of varices	1.025 (0.839-1.240)	0.839			0.852 (0.665-1.094)	0.205
Hepatic encephalopathy (yes vs no)	1.430 (1.211-1.704)	< 0.001	1.251 (1.010-1.548)	0.040	1.378 (1.094-1.725)	0.007	1.180 (0.984-1.548)	0.230
Ascites (yes vs no)	1.750 (1.415-2.187)	< 0.001	1.256 (1.093-1.667)	< 0.001	2.001 (1.515-2.666)	< 0.001	1.875 (1.375-2.571)	< 0.001
Systolic blood pressure at admission (≥ 120 mmHg vs < 120 mmHg)	1.051 (0.875-1.260)	0.591			1.079 (0.855-1.339)	0.563
Diastolic blood pressure at admission (≥ 90 mmHg vs < 90 mmHg)	1.303 (0.846-2.115)	0.288			0.826 (0.540-1.280)	0.353
MELD score (> 13 vs ≤ 13)	5.020 (3.750-6.667)	< 0.001	1.516 (1.167-1.923)	0.002	4.520 (3.697-6.750)	< 0.001	3.427 (2.167-4.154)	< 0.001
Child-Pugh class (C vs A + B)	6.125 (2.567-9.450)	< 0.001	4.231 (2.540-6.221)	< 0.001		0.995
Different treatment strategies (EVL + PT vs p-TIPS)	2.350 (1.903-2.999)	< 0.001	1.705 (1.300-2.250)	< 0.001	1.955 (1.700-2.300)	0.018	1.051 (0.746-1.320)	0.082

EVL: Endoscopic variceal ligation; PT: Pharmacotherapy; HR: Hazard ratio; CI: Confidence interval; MELD: Model for end-stage liver disease; p-TIPS: Preemptive transjugular intrahepatic portosystemic shunt.

Comparable efficacy of EVL + PT and p-TIPS in patients with low-risk AVB

The six-week treatment failure rate was 6.8% in the EVL + PT group (68 patients) and 1.9% in the p-TIPS group (22 patients; P = 0.013). The one-year mortality rates were comparable in the EVL + PT (14.0%) and p-TIPS (10.4%) groups (P = 0.036). HE occurred in 14.9% and 27.1% of the EVL + PT and p-TIPS groups, respectively (P < 0.001), whereas new or worsening ascites was observed in 1.9% and 0.9% of the EVL + PT and p-TIPS groups, respectively (P < 0.001; Table 2).

Further multivariate regression analysis (Table 3) revealed that for patients with low-risk AVB, the presence of ascites and a MELD score > 13 were independent predictors of poor short-term prognosis. Specifically, ascites had an HR of 22.825 (95%CI: 16.485-31.604, P < 0.001), and MELD > 13 had an HR of 7.208 (95%CI: 4.155-12.504, P < 0.001). Conversely, the adjusted HR between the two treatment strategies (EVL + PT vs p-TIPS) was 0.954 (95%CI: 0.671-1.355, P = 0.791), indicating that the aggressive treatment strategy was not an independent factor directly contributing to the survival benefit of patients with low-risk AVB (Supplementary Tables 5-7).

DISCUSSION

This study developed and validated a novel AI-driven model using deep learning technologies based on clinical examination data obtained within 24 hours of the initial hospital admission for patients with cirrhosis presenting with AVB. The AI-AVB model demonstrated exceptional performance in distinguishing between high-risk and low-risk patients for first-line treatment failure, accurately predicting six-week treatment failure and one-year mortality rates compared with other risk stratification methods. The results confirmed that the AI-AVB model in clinical practice has the potential to significantly enhance personalized treatment strategies, ensuring that high-risk patients receive the most appropriate and aggressive interventions while excluding low-risk patients from unwarranted invasive procedures.

Our findings corroborate those of previous studies, highlighting the key predictors of mortality and treatment failure in AVB. Advanced age, HE, and ascites are significant predictors of adverse outcomes[3,41,42]. Advanced age (≥ 60 years) was particularly influential in the low-risk subgroup, showing an HR of 1.312 in univariate analysis. HE (HR = 1.464) and ascites (HR = 17.848) were critical in the high-risk group. Moreover, elevated MELD scores (> 13) and higher Child-Pugh classifications were robustly associated with increased mortality, emphasizing the importance of liver and renal functions in determining prognosis[43]. Specifically, the MELD scores showed an HR of 4.968 (P < 0.001) and 7.208 (P < 0.001) in in high-and low-risk patients, respectively. These findings suggest these variables were more indicative of patient outcomes than the severity of the bleeding episodes. They can be attributed to advances in effective hemostatic treatments, such as endoscopy, PT, and the use of TIPS, along with improved general medical management practices, including restrictive transfusion strategies and prophylactic antibiotic use[44-46]. These insights underscore the need for early and aggressive management of the underlying liver dysfunction and related complications to improve outcomes in patients with AVB.

The superior performance of our decision-support model likely stems from its use of deep learning and AI technologies, outperforming traditional clinical risk scores (such as the Baveno VII criteria, MELD, Child-Pugh, and HVPG)[15,23] (Supplementary Tables 8-10) and other ML algorithms. First, the model architecture, incorporating a residual structure, pyramid architecture, and multilevel feature fusion strategy[47], enables it to effectively capture and process complex representations within the data. The inclusion of a wide range of variables, such as clinical parameters, laboratory results, and endoscopic findings, allows for comprehensive analyses and robust prediction capabilities[48,49]. Additionally, the use of advanced techniques such as ASPP and online data augmentation enhanced the ability of model characterization and generalization. The integration of these sophisticated methods ensured that the AI-AVB model could accurately stratify patients according to risk and predict critical outcomes.

By leveraging comprehensive data collected within the first 24 hours of admission, clinicians can utilize the model’s predictions to identify high-risk patients who would benefit from more aggressive and invasive interventions, such as p-TIPS or liver transplantation, thereby reducing the likelihood of treatment failure[13,50]. Although these patients may incur additional treatment risks or adverse effects, the significant improvement in the success rates following six weeks of treatment and one-year survival justifies the approach[9]. Conversely, for patients identified as low-risk, although p-TIPS offers superior hemostatic efficacy, the increased incidence of HE in our study may ultimately negate the benefits, failing to improve prognosis independently. In such cases, the model can guide clinicians toward less-invasive treatments, minimizing unnecessary surgical interventions and their associated risks (such as puncture failure, HE, liver failure, secondary infections, and stent occlusion or displacement)[51-53]. This risk-stratified approach to post-treatment failure and mortality ensures more efficient resource allocation and tailored care according to each patient’s specific risk profile, ultimately improving the overall treatment outcomes and patient prognosis.

Beyond its clinical utility, the AI-AVB model may yield meaningful socioeconomic gains by aligning therapeutic intensity with individualized risk. Accurate stratification directs high-risk patients to timely, resource-intensive care (e.g., p-TIPS and close monitoring), areas where we observed evident outcome advantages, while guiding low-risk patients toward EVL + PT, which achieved comparable outcomes without routine p-TIPS. This risk-based triage reduces unwarranted procedures and procedure-related complications (notably HE), lowers direct procedural costs, and avoids downstream expenditures from treatment failure. In parallel, prioritizing candidates that are most likely to benefit optimizes the ICU capacity and endoscopy/interventional radiology scheduling, thereby facilitating patient flow in emergency settings.

This study has some limitations. First, in our multicenter Chinese cohort, HBV-related cirrhosis was predominant and other etiologies were underrepresented, rendering underpowered etiology-specific analyses with limited generalizability, warranting further validation in etiology-balanced cohorts. Second, although AI-AVB demonstrated strong overall performance, its specificity and PPV were lower than its sensitivity and NPV in certain validation settings, possibly owing to disease-class imbalance and deliberate prioritization of sensitivity to avoid missing high-risk cases. Clinically, this supports the use of the model to exclude low-risk patients; however, high-risk predictions should be interpreted in conjunction with additional clinical assessments. Third, the retrospective design may have introduced biases, although we minimized this risk using a large multicenter cohort. Additionally, data collected within the first 24 hours of admission, while relevant, may not fully capture patient variability and disease progression. Patients with severe renal impairment, portal vein thrombosis, heart failure, or hepatocellular carcinoma were excluded, limiting the generalizability of the results. The model relies on comprehensive data, which necessitates robust data collection systems that may not be universally available. Moreover, the small external validation dataset requires further validation in diverse patient populations. Future improvements should include prospective validations in various clinical settings and the incorporation of additional variables, such as continuous monitoring data, to enhance predictive accuracy.

CONCLUSION

The AI-AVB model, developed using advanced deep learning techniques, excels in stratifying high- and low-risk patients for first-line treatment failure among patients with cirrhosis-associated AVB, accurately predicting the six-week treatment failure and one-year mortality rates. This model offers a promising tool for ensuring that patients with AVB receive personalized interventions to improve their overall prognosis and resource allocation.

ACKNOWLEDGEMENTS

The authors acknowledge all the clinical and research staff from the research centers.