Application of artificial intelligence model in precise risk stratification and treatment decision-making for acute variceal bleeding in cirrhosis

Abstract

This editorial focuses on a recent study by Xiang et al published on the World Journal of Gastroenterology that developed and validated an artificial intelligence-driven model to refine risk stratification and guide personalized treatment plans for acute variceal bleeding (AVB) in cirrhotic patients. The predictive model was constructed using clinical data collected within the first 24 hours following patient admission. Its superior predictive accuracy translates into a critical clinical application: The model effectively identifies high-risk patients who would derive the greatest benefit from preemptive transjugular intrahepatic portosystemic shunt, and simultaneously identifies low-risk patients for whom such an invasive procedure can be safely avoided. The artificial intelligence-AVB model represents a practical application of precision medicine by translating complex data into actionable, personalized strategies for the treatment of AVB.

Key Words: Non-invasive assessment; Precision medicine; Acute variceal bleeding; Artificial intelligence; Preemptive transjugular intrahepatic portosystemic shunt

Core Tip: This editorial reviews the non-invasive diagnostic techniques for portal hypertension and the application progress of artificial intelligence in the screening, diagnosis, and treatment of cirrhosis-related gastroesophageal varices. It focuses on exploring the application value of artificial intelligence in the risk stratification assessment of high-risk gastroesophageal variceal bleeding, post-bleeding condition monitoring, and the formulation of individualized diagnosis and treatment decisions.

This editorial refers to "Development of a deep learning model for guiding treatment decisions of acute variceal bleeding in patients with cirrhosis" by Xiang et al, 2025; https://doi.org/10.3748/wjg.v31.i41.111361.

---

INTRODUCTION

Acute variceal bleeding (AVB) is the most life-threatening fatal complication of cirrhosis-related portal hypertension[1,2]. The short-term mortality rate of AVB is 10%-20% within 6 weeks[3], and its high mortality rate constitutes a major public health challenge that urgently needs to be addressed worldwide. The core pathophysiological basis of AVB is intrahepatic vascular remodeling and sinusoidal capillarization induced by the persistent activation of hepatic stellate cells and excessive deposition of extracellular matrix during the progression of cirrhosis. These irreversible changes lead to a progressive elevation in intrahepatic vascular resistance, accompanied by splanchnic vasodilation and a systemic hyperdynamic circulatory state, ultimately resulting in tortuous dilatation and thinning of the esophagogastric variceal wall, followed by rupture and bleeding[4,5]. Despite the widespread application of standard therapeutic regimens such as vasoactive agents, endoscopic therapy, and transjugular intrahepatic portosystemic shunt (TIPS), the prognosis of patients with AVB remains poor[6,7]. Accurate risk stratification and early intervention become crucial, as they can significantly improve patient outcomes[8]. Traditional methods for stratifying risk in AVB have relied on established clinical scoring systems. Traditional scoring systems, such as model for end stage liver disease[9] and Child-Pugh[6], have limitations. These scoring tools cannot fully capture the dynamics and complex pathophysiology of AVB, which may lead to inaccurate risk classification and affect personalized treatment for patients[10].

The development of artificial intelligence (AI) models is enhancing the precision of risk stratification and individualized management in cirrhotic AVB[11]. The main strength of these models is to integrate and process large-scale, multi-dimensional datasets, including encompassing clinical profiles, laboratory results, imaging features, and electronic health records[12]. We focus on a recent study by Xiang et al[13] published on the World Journal of Gastroenterology, which developed and validated a non-invasive model for AVB in patients with hepatitis B virus-related cirrhosis. The model aims to identify individuals at high risk of standard treatment failure and those most likely to benefit from preemptive TIPS (p-TIPS) therapy. This novel approach not only optimizes patient selection criteria but also significantly improves clinical efficacy. In this editorial, we aim to discuss the application of AI models in precision risk stratification and individualized management of AVB in patients with cirrhosis. We also discuss the research limitations and clinical significance of the AI-AVB model.

APPLICATION OF NON-INVASIVE DETECTION OF PORTAL HYPERTENSION

Portal hypertension, a complication of hepatic disorders such as cirrhosis, is defined by compromised portal venous blood flow and pathologically elevated pressure. The most critical life-threatening complication is esophageal-gastric variceal rupture hemorrhage, a major contributor to death among individuals with portal hypertension. This risk is attributed to its abrupt onset and massive blood loss[14]. Assessing portal hypertension has required invasive approaches, including hepatic venous pressure gradient (HVPG) measurement as a typical diagnostic tool. HVPG levels of 10 mmHg or higher can be diagnosed as clinically significant portal hypertension (CSPH). This technique requires transvenous catheterization and carries the risk of bleeding or infection, making it difficult for primary healthcare institutions to routinely perform[15,16].

Non-invasive assessment has fundamentally altered this paradigm. Two core systems prevail: Imaging evaluation and multi-marker combined models[17]. In terms of imaging, ultrasound elastography indirectly assesses portal venous pressure by quantifying liver or spleen stiffness[18], while computed tomography (CT) and magnetic resonance imaging can reveal the morphological characteristics of blood vessels[19].

The Baveno VII consensus endorses the ANTICIPATE model as a non-invasive predictor for portal hypertension[20]. The model directly classifies the risk level of CSPH through the combined threshold of liver stiffness measurement (LSM) and platelet count (PLT). If the conditions of “LSM 15-20 kPa and PLT < 110 × 10⁹/L” or “LSM 20-25 kPa and PLT < 150 × 10⁹/L” are met, the patient is considered at high risk for CSPH. It is recommended to further perform endoscopic examination to screen for esophageal and gastric fundus varices, or consider HVPG testing for confirmation. If LSM is ≥ 25 kPa, CSPH diagnosis can be made directly with a specificity > 90%, and clinical intervention should be initiated immediately.

Although non-invasive diagnosis integrating transient elastography (TE) and PLT has formed a standardized protocol within the Baveno consensus framework, it still has limitations in complex clinical scenarios, including measurement errors of TE in obese patients, shifts in diagnostic thresholds in the setting of comorbid liver diseases, and diagnostic uncertainties caused by fluctuating PLT in some patients[17].

The development of machine learning (ML) technology has provided new ideas for breaking through the above bottlenecks: By integrating multi-dimensional clinical data, such as liver function indices, inflammatory markers, imaging features, etc., researchers can construct ML-based prediction models that further improve the accuracy of diagnosing CSPH[21,22].

A study published in Cell Reports Medicine in 2022[23] addressed the limitations of invasive HVPG testing, a diagnostic gold standard for cirrhotic portal hypertension, by developing a non-invasive AI model based on CT images. The deep learning network with V-Net architecture automatically segmented three-dimensional liver and spleen volumes (Dice coefficient ≥ 0.97). Then, an AutoML tool extracted 2436 radiomics characteristics from the liver and spleen to the build model. Among 372 patients (split 6:4 into training/internal validation cohorts) and 25 external validation cases, the model achieved areas under the curve (AUCs) > 0.80 for all HVPG stratifications (≥ 10, 12, 16, 20 mmHg), outperforming traditional imaging tools and serum markers. Another investigation[20] conducted a retrospective review of clinical information collected from 200 individuals diagnosed with viral cirrhosis. TE was applied to obtain LSM and spleen stiffness measurements (SSMs), whereas CT was used to measure spleen length diameter. A noninvasive prediction model was constructed, which achieved an AUC of 0.944 when predicting clinical decompensation. This was significantly greater than existing models such as liver stiffness-platelet score (0.834) and venous reticular index (0.824), and the model performed well in an external validation cohort. It can effectively reduce the need for HVPG measurement. This study confirms that non-invasive tests can substitute for HVPG in risk assessment and treatment decision-making for compensated advanced chronic liver disease patients. These results from non-invasive methods for diagnosing portal hypertension show significant clinical value. However, future studies must prioritize improving the accuracy and reliability of these predictive models to ensure their effective integration into routine practice.

APPLICATION OF AI IN CIRRHOTIC AVB

The application of AI in gastroesophageal varices of liver cirrhosis can be divided into the pre-bleeding screening stage and the post-bleeding decision-making stage[15]. Endoscopy provides an accurate diagnosis for varices, but at the same time, due to its invasive nature, it places a burden on patients. The development of non-invasive prediction models offers another approach[24]. AI-driven research now aims to fuse multimodal inputs-spanning ultrasound, CT, laboratory markers, and clinical data through deep learning algorithms into a unified prognostic framework[25]. This system automatically detects vascular morphological characteristics and establishes meaningful connections between serum biomarkers and clinical parameters to enhance the prediction of high-risk varices.

Yang et al[26] derived and internally tested a non-invasive logistic regression model incorporating LSM and SSM to identify high-risk esophageal varices in patients with viral cirrhosis, aiming to reduce unnecessary endoscopic screening. In a retrospective design, 200 patients with hepatitis B or C-related cirrhosis were enrolled, with 140 assigned to the derivation cohort and the remaining 60 to an external validation set. The final model was expressed as Ln(P/1 - P) = -8.184 - 0.228 × SSM + 0.642 × LSM. The AUC reached 0.965 in the derivation cohort and a perfect 1.000 in the external validation cohort. At a cut-off value of 0.27, both sensitivity and negative predictive value were 100%, while specificity was 82.43% and positive predictive value 83.52%, substantially outperforming existing non-invasive tools including liver stiffness-spleen diameter to platelet ratio score, variceal risk index, aspartate aminotransferase to alanine aminotransferase ratio, and Baveno VI criteria. The Hosmer-Lemeshow test yielded P values above 0.05, indicating satisfactory calibration, and decision curve analysis confirmed a consistent net clinical benefit. The perfect negative predictive value enables safe exclusion of high-risk varices, thereby sparing a large proportion of patients from invasive endoscopic procedures.

Dong et al[27] pioneered and validated fibrosis-4 plus (FIB-4plus), a ML-derived scoring system targeting high-risk varices in compensated cirrhosis to curtail over-reliance on esophagogastroduodenoscopy. Their multicenter cohort enrolled 502 patients across 17 Chinese and one Croatian center. Integrating FIB-4 (age, aspartate aminotransferase, alanine aminotransferase, PLT), LSM, and SSM, they built logistic regression-FIB-4plus and XGBoost-FIB-4plus models. XGBoost-FIB-4plus outperformed others, achieving AUCs of 0.927 (training), 0.919 and 0.902 in the two validation cohorts. For esophageal varices, AUC exceeded 0.8 across all cohorts, maintaining stability in hepatitis B virus and TE subgroups. SHapley Additive exPlanations analysis singled out SSM and PLT as top predictors. FIB-4plus enables accurate non-invasive screening to reduce unnecessary endoscopies, though etiological biases and absent long-term data remain constraints.

Another study[28] focuses on building and validating a radiomic model based on ML for non-invasive diagnosis of esophageal varices with high bleeding risk. A cohort of 796 patients was recruited, with regions of interest in the liver, spleen, and esophagus (5 cm above the cardia) marked on portal venous-phase enhanced CT scans. Altogether, 2358 radiomic features, including texture and statistical features, were extracted. These features were then reduced using principal component analysis and modeled with support vector machines. The results showed the esophageal varices with high bleeding risk model exhibited an AUC of 0.983 (training cohort), 0.834 (internal validation cohort), and 0.736 (external validation cohort), with good sensitivity and specificity. Both the diagnostic accuracy and net reclassification improvement were notably greater than those of Baveno VI and its extended criteria.

Although the above-mentioned research is helpful for non-invasive prediction of the risk of esophageal and gastric variceal bleeding in patients with liver cirrhosis, it still has obvious limitations. Most models focus on predicting the risk of bleeding, but they cannot provide precise guidance for key treatment decision points of AVB.

In the clinical treatment of liver cirrhosis complicated by AVB, variceal ligation and p-TIPS treatment are two commonly used surgical methods[8]. The effect of variceal ligation is limited to controlling bleeding symptoms and cannot significantly reduce portal vein pressure. Rebleeding risk remains relatively high. P-TIPS is the core method for second-line treatment of AVB in liver cirrhosis, and its indication selection directly affects patient prognosis. Excessive intervention may increase complication risk, including hepatic encephalopathy and infection, while delayed intervention may lead to increased rebleeding rates[29]. Most traditional models rely on static clinical indicators and ignore the clinical value of dynamic data, such as hemodynamic fluctuations and organ function changes within 24 hours of acute bleeding. Xiang et al[13] innovatively adopted a deep learning model to integrate clinical data within 24 hours of AVB and constructed a predictive model for the identification of p-TIPS indications. This bridges a pivotal research gap while equipping clinicians with tools for individualized treatment planning.

CLINICAL SIGNIFICANCE AND LIMITATIONS

The main strength of the article is the integration of deep learning with clinical data, an approach utilizing information obtained within the first 24 hours of patient admission. Among the high-risk patients identified by the AI-AVB model, those who only received endoscopic variceal ligation plus pharmacotherapy (EVL+PT) and did not subsequently implement p-TIPS had a 6-week treatment failure rate as high as 39.6%. In the high-risk group, prophylactic p-TIPS was performed within 72 hours of admission following EVL+PT, reducing the 6-week treatment failure rate to 3.8% and the 1-year mortality rate from 37.3% to 14.6%. For low-risk patients, EVL+PT alone achieved a 1-year survival rate comparable to that of p-TIPS (14.0% vs 10.4%), and avoided the increased risk of hepatic encephalopathy associated with p-TIPS (14.9% vs 27.1%). Overall, this model supports aggressive intervention in high-risk patients and avoidance of overtreatment in low-risk patients. The model demonstrated an AUC of 0.842 for predicting 6-week treatment failure and an impressive AUC of 0.954 for predicting 1-year mortality in the internal validation cohort. In the external validation cohort, it maintained excellent performance with AUCs of 0.814 and 0.889, surpassing traditional methods. For patients with AVB, the examinations required by the model can be completed after admission, with endoscopy performed concurrently to control bleeding. High-risk patients identified by the model then underwent early prophylactic p-TIPS. Table 1 compares the non-invasive diagnostic tools for high-risk esophageal varices in cirrhotic patients[26-28,30,31].

Table 1 Comparison of non-invasive diagnostic tools for high-risk esophageal varices in cirrhotic patients[26-28,30,31].

Model	Parameters	Detection modalities	Target etiology	Diagnostic performance (AUC)	Main limitations
FIB-4plus (XGBoost)	FIB-4 (age, AST, ALT, PLT) + LSM + SSM	Serological testing + TE	Compensated cirrhosis (predominantly HBV-related)	Training: 0.927; validation 1: 0.919; validation 2: 0.902	Etiological selection bias; lack of long-term prognostic data; relatively low interpretability of XGBoost black-box model
LSM + SSM logistic regression model	LSM, SSM	TE	Viral cirrhosis (HBV/HCV)	Derivation: 0.965; external validation: 1.000	Restricted to viral cirrhosis; small total sample size (n = 200); no long-term follow-up data
Multi-organ CT radiomics SVM model	2358 radiomic features extracted from portal venous-phase CT (liver, spleen, and esophagus 5 cm above the cardia)	Multislice spiral CT (portal venous phase enhancement)	Mixed-etiology cirrhosis	Training: 0.983; internal validation: 0.834; external validation: 0.736	Marked performance degradation in external validation; complex radiomic feature extraction; requires dedicated software and trained personnel
Meta-analysis of SSM alone	SSM	Ultrasound elastography (TE/SWE) + magnetic resonance elastography		0.83	Significant between-study heterogeneity; low positive predictive value (54%); positive results still require confirmatory endoscopy
Meta-analysis comparing CT, LSM, and MRI	Head-to-head comparison of diagnostic performance between CT, LSM, and MRI	Multislice spiral CT + TE + MRI	All cirrhosis subtypes	CT: 0.94; LSM: 0.85; MRI: 0.83	Indirect comparison between studies; moderate heterogeneity; no standardized universal diagnostic threshold provided

AUC: Area under the curve; FIB-4plus: Fibrosis-4 plus; LSM: Liver stiffness measurement; SSM: Spleen stiffness measurement; CT: Computed tomography; SVM: Support vector machine; MRI: Magnetic resonance imaging; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; PLT: Platelet count; TE: Transient elastography; SWE: Shear wave elastography; HBV: Hepatitis B virus; HCV: Hepatitis C virus.

Existing research has a common limitation in etiological representativeness. These studies were dominated by individuals with hepatitis B-associated cirrhosis, while patients with other etiologies accounted for a smaller proportion, therefore requiring further validation in cohorts with more balanced etiologies. Alcohol-related liver disease is surging, whereas the incidence of hepatitis B and C is receding under active intervention[32]. In a multicenter cohort[33], non-alcoholic cirrhosis conferred a significantly elevated 6-week mortality risk vs alcoholic cirrhosis among AVB patients with equivalent standard liver function scores. Consequently, liver disease etiology warrants incorporation as a pivotal factor into AVB risk stratification frameworks to augment predictive accuracy[34]. In alcoholic cirrhosis patients, non-invasive assessment modalities exhibit vulnerability to both alcohol consumption fluctuations and obesity; consequently, their predictive efficacy is compromised[20,35]. Therefore, further subgroup analyses stratified by cirrhosis etiology are warranted to clarify the predictive performance of the model across distinct etiological subtypes (e.g., differences in AUC, sensitivity, and specificity) and to examine how the relative contributions of key predictors vary by etiology. This will help clinicians make more accurate p-TIPS decisions for patients with AVB due to cirrhosis of different etiologies, thereby enhancing the clinical adaptability and individualized applicability of the model. In some validation scenarios, integrating emerging biomarkers (e.g., serum markers reflecting the degree of liver fibrosis and molecular indicators related to portal vein pressure) or radiological models (e.g., imaging-based liver stiffness assessment) with the existing AI-AVB model may enhance its predictive accuracy.

CONCLUSION

The AI-AVB model marks a new stage of “data-driven” precision medicine for the diagnosis and management of cirrhotic AVB. Future research or clinical practice should focus on overcoming existing limitations through prospective multicenter studies, the inclusion of patients with diverse etiologies and complex conditions, and optimization of algorithms to enhance the comprehensiveness of predictions.

ACKNOWLEDGEMENTS

The author expresses gratitude to the clinical team for their continuous support, which has laid a solid foundation for the exploration of acute variceal bleeding in liver cirrhosis this time.