Application of artificial intelligence in portal hypertension and esophagogastric varices

Abstract

Esophagogastric variceal bleeding is a common and severe complication of cirrhotic portal hypertension. Hepatic venous pressure gradient measurement and esophagogastroduodenoscopy are the diagnostic gold standards for portal hypertension and esophagogastric variceal bleeding, respectively. With advancements in artificial intelligence in medicine, non-invasive diagnostic methods are increasingly replacing traditional invasive procedures, permitting more rational and personalized patient care. This review summarizes the formation and diagnosis of portal hypertension, as well as the primary prophylaxis, secondary prophylaxis, and management of acute esophagogastric variceal bleeding. This study also highlights the latest progress in artificial intelligence in the diagnosis and treatment of portal hypertension and esophagogastric varices.

Key Words: Cirrhosis; Portal hypertension; Esophagogastric variceal; Artificial intelligence; Diagnosis; Management

Core Tip: This review highlights the latest progress of artificial intelligence (AI) in the diagnosis and treatment of portal hypertension and esophagogastric varices. It emphasizes AI’s potential in early non-invasive diagnosis, risk prediction, and treatment optimization. Key points include the application of machine learning and deep learning in analyzing medical imaging and clinical data to improve diagnostic accuracy and personalized treatment. The review also discusses challenges in AI implementation, such as data quality, model interpretability, and regulatory requirements, and suggests future research directions focusing on enhancing AI’s role in clinical practice.

INTRODUCTION

Cirrhosis, resulting from the progression of various chronic liver diseases, is characterized by hepatocyte loss, diffuse hepatic fibrosis, and vascular hyperplasia[1,2]. The number of deaths caused by cirrhosis has risen significantly to approximately 1 million annually[3]. Clinically, cirrhosis is divided into compensated and decompensated stages. The Baveno VII consensus classifies compensated cirrhosis into two stages based on the presence or absence of clinically significant portal hypertension (CSPH), each with different treatment priorities[4]. Decompensated cirrhosis is defined by significant complications of portal hypertension, such as ascites, variceal bleeding, and overt hepatic encephalopathy (OHE), which severely affect patients’ quality of life and impose a heavy burden on healthcare services[5,6]. Therefore, early diagnosis, monitoring, and precise intervention are crucial for treating hypertension. Hepatic venous pressure gradient (HVPG) measurement is considered the gold standard for assessing portal hypertension[7]. HVPG > 5 mmHg denotes portal hypertension, HVPG ≥ 10 mmHg indicates CSPH, and HVPG > 20 mmHg is an independent risk factor for early rebleeding or death[8,9]. Approximately 50% of patients with cirrhosis have esophagogastric varices (EGVs), which can progress to esophagogastric variceal bleeding (EGVB)[4,10,11]. EGVB is a significant clinical complication of portal hypertension that carries a 6-week mortality rate as high as 20%[10,11]. Thus, identifying high-risk patients and preventing the first episode of EGVB are critical. Esophagogastroduodenoscopy (EGD) is the gold standard for diagnosing EGVs and EGVB[4,10,11]. However, EGD exhibits significant interobserver heterogeneity, and it is an invasive procedure with a risk of variceal bleeding.

In recent years, the emergence of artificial intelligence (AI) has opened new avenues for the early non-invasive diagnosis of cirrhotic portal hypertension and EGVs (Table 1). AI, which centers on algorithms, employs computers to simulate human intelligence and perform a variety of tasks[12-14]. With the widespread use of electronic medical record systems, vast amounts of clinical data are available for extraction, making AI increasingly applicable in various aspects of medical practice. Machine learning is the core technology of AI, and it encompasses shallow learning and deep learning, with the latter incorporating methods such as convolutional neural networks (CNNs)[13]. Deep learning constructs multilayer neural networks to automatically learn complex patterns and features in data, but it faces challenges in data requirements, computational resources, and model interpretability. AI based on machine learning has been applied to the diagnosis of portal hypertension and its complications[15]. This review examines the role of AI in the early non-invasive diagnosis and risk prediction of portal hypertension and EGVs, as well as the current management of EGVB.

Table 1 Summary of key artificial intelligence applications in portal hypertension and esophagogastric varices management.

Applications	Techniques/methods	Research data/performance metrics
Diagnostic tools
Liver and spleen ultrasound elastography	Ultrasound elastography	LSM and SSM values correlated with HVPG
CT/MRI imaging analysis	CT, MRI, deep learning, radiomics	rHVPG model performance (AUC value), virtual HVPG validation, morphological assessment of varices
Deep learning models (DCNN)	Deep learning	Model AUC value (e.g., 0.9), sensitivity, specificity
Prognostic models
HVPG prediction model	Machine learning, CT radiomics	aHVPG model AUC value (e.g., 0.80)
Variceal bleeding risk prediction	Deep learning	Model AUC value (internal: 0.782, external: 0.789), calibration and decision curve analysis
Treatment selection aids
Endoscopic virtual ruler (ENDOAGGEL)	Deep learning	Accuracy for detecting EV and GV (97.00% and 92.00%)
TIPS post-OHE prediction model	Machine learning	Model accuracy in predicting OHE, comparison with traditional models

CT: Computed tomography; MRI: Magnetic resonance imaging; DCNN: Deep convolutional neural networks; HVPG: Hepatic venous pressure gradient; TIPS: Transjugular intrahepatic portosystemic shunt; OHE: Overt hepatic encephalopathy; LSM: Liver stiffness measurement; SSM: Spleen stiffness measurement; rHVPG: Radiomics signature of hepatic venous pressure gradient; AUC: Area under the curve; aHVPG: Automatic machine learning-based quantitative model; EV: Esophageal varices; GV: Gastric varices.

FORMATION AND DIAGNOSIS OF PORTAL HYPERTENSION

Formation of portal hypertension

Portal hypertension is a clinical syndrome caused by increased pressure in the portal vein and its branches, with cirrhosis being the most common cause[16]. In the early stages of portal hypertension, increased portal venous resistance is the primary pathogenic factor, and it usually results from inflammation-induced structural changes in the liver[17]. These permanent structural changes account for 70% of the increased intrahepatic resistance, whereas the remaining 30% is attributable to increased tension in the intrahepatic capillary bed. The increase in vascular tension is a dynamic and reversible change caused by insufficient vasodilator release and increased vasoconstrictor production[18]. Numerous studies have highlighted the importance of HVPG in liver cirrhosis staging, risk stratification, treatment monitoring, and prognosis. Patients with HVPG ≤ 12 mmHg or a ≥ 10% decrease from baseline (defined as “HVPG responders”) have a low risk of variceal rebleeding and reduced risks of ascites, hepatic encephalopathy, and death[19]. However, HVPG measurement remains invasive, and it is not widely used in clinical practice. Therefore, non-invasive diagnostic methods for early identification of CSPH are being explored.

Ultrasound elastography

Ultrasound elastography is a highly sensitive imaging technique for detecting changes in the stiffness of visceral tissues, and it is widely used to assess portal venous pressure[20]. Ultrasound elastography techniques include transient elastography (TE) and two-dimensional shear wave elastography (2D-SWE)[21]. Because of their simplicity and low cost, these techniques represent effective alternatives for patients unable or unwilling to undergo HVPG measurement.

TE

The Baveno VI consensus introduced the concept of compensated advanced chronic liver disease (cACLD), which can be diagnosed using TE-based liver stiffness measurement (LSM) and spleen stiffness measurement (SSM)[22]. LSM has been extensively applied for non-invasive portal pressure assessment[23]. LSM < 10 kPa can exclude cACLD, whereas LSM > 15 kPa strongly suggests cACLD. For patients with chronic viral hepatitis, alcohol-related liver disease, and non-obese (body mass index < 30 kg/m²) non-alcoholic steatohepatitis-induced cACLD, LSM ≥ 25 kPa has a specificity and positive predictive value exceeding 90% for diagnosing CSPH[4]. When LSM < 25 kPa, the platelet count can be measured to assess whether patients with cACLD have concomitant CSPH. For example, when 20 kPa ≤ LSM < 25 kPa and platelet count < 150 × 10⁹/L or when 15 kPa ≤ LSM < 20 kPa and platelet count < 110 × 10⁹/L, more than 60% of patients with cirrhosis have CSPH[4]. Clinically, SSM is less widely used than LSM. For patients with cACLD caused by viral hepatitis, guidelines recommend using a special 100-Hz probe to simultaneously monitor SSM[4,24]. SSM < 21 kPa can exclude CSPH, whereas SSM > 50 kPa indicates CSPH. In patients with viral hepatitis, SSM might be more sensitive than LSM, and it might have greater potential in predicting CSPH[4,24]. However, the cutoffs for LSM and SSM remain controversial and require validation through large-scale studies.

2D-SWE

2D-SWE is a new ultrasound elastography technique that reflects the severity of portal hypertension. Using safer acoustic radiation pulses, 2D-SWE has the potential for widespread clinical application[25]. A meta-analysis of studies on 2D-SWE indicated that LSM < 14 kPa on 2D-SWE can exclude CSPH in patients with cirrhosis[26]. However, because of the heterogeneity of the observed cutoffs, it cannot yet be routinely recommended.

Computed tomography and magnetic resonance imaging

In recent years, imaging techniques have made significant progress in the early non-invasive diagnosis of CSPH. Computed tomography (CT) and magnetic resonance imaging (MRI) provide detailed anatomical information of the liver and spleen, and they can be used to assess hemodynamic changes in the portal venous system, offering strong support for the early diagnosis and risk assessment of CSPH[27,28].

CT features a short imaging time and high spatial resolution. A prospective study involving 385 patients with cirrhosis from five liver disease centers developed a non-invasive radiomic diagnostic model called radiomics signature of HVPG that displayed good performance in detecting CSPH and good inter- and intra-observer consistency[29]. Another study established a computational model, termed virtual HVPG, based on CT angiography to estimate HVPG. Validation illustrated that virtual HVPG had an area under the curve (AUC) of 0.89 [95% confidence interval (CI) = 0.81-0.96] for predicting CSPH and a good correlation with traditional HVPG (r = 0.61, P < 0.001)[30]. MRI offers advantages in soft tissue resolution and accurate lesion detection, but it is more time-consuming and expensive than CT. Shi et al[31] used multiparametric three-dimensional MR elastography to assess the severity of portal hypertension in patients with chronic hepatitis B or C, finding correlations between MR elastography findings and with HVPG in these patients.

Deep learning

Machine learning is a subset of AI that involves training algorithms to make predictions or decisions based on data. It enables computers to learn from and improve with experience without being explicitly programmed. The regulatory landscape for AI in clinical practice is more concerned with the rules, guidelines, and frameworks that govern the development, validation, and use of AI technologies in healthcare settings[32]. In the context of portal hypertension and EGV, machine learning techniques, such as supervised learning, weakly supervised learning, and unsupervised learning, are used to analyze complex datasets and develop predictive models that can assist in diagnosis, risk stratification, and treatment optimization[33].

Deep learning is a subset of machine learning, with CNNs being the most popular type in medical imaging analysis[34]. Deep learning mimics the structure of the human brain’s neural networks to extract various quantitative features from traditional radiomic data, aiding in the creation of non-invasive assessment models for automatic lesion identification and assisting doctors in making more accurate and comprehensive imaging diagnoses[35]. Unlike traditional radiomics approaches that rely on manually extracted features, deep learning models can identify subtle patterns imperceptible to human experts.

A multicenter study used deep convolutional neural networks (DCNNs) to analyze CT or MRIs of the liver and spleen in patients with cirrhosis, developing a model for identifying patients with CSPH with an AUC as high as 0.9, indicating a significant advantage in identifying patients with CSPH. This innovative, non-invasive, rapid method offers a new solution for the clinical diagnosis of CSPH[36]. Another study successfully developed an automatic machine learning-based quantitative model, called aHVPG, for HVPG assessment using CT radiomic features. Compared with other non-invasive models based on imaging and serological data, the aHVPG model had an AUC exceeding 0.80, indicating a significant advantage in assessing the severity of portal hypertension[37].

PRIMARY PROPHYLAXIS OF EGVB

The initial rupture and bleeding of EGV are related to the Child-Turcotte-Pugh grade, variceal diameter, and presence of red signs on the varices[38]. High-risk gastric varices (GVs) are defined by maximum variceal diameter > 10 mm or positive red signs on the varices, although this definition requires further unification[39]. According to the latest Baveno VII consensus, non-selective beta-blockers (NSBBs) are recommended for the primary prophylaxis of decompensation in patients with CSPH[4]. NSBBs act on both β1-adrenergic receptors, reducing heart rate and cardiac output, and visceral vascular β2-adrenergic receptors, causing visceral vasoconstriction and thereby lowering portal venous pressure[40]. Carvedilol also has a unique α1 receptor-blocking effect, which can improve hepatic microcirculation and reduce intrahepatic vascular resistance[41]. Recent studies found that compared with the effects of propranolol, carvedilol more effectively reduces HVPG and prevents EGVB, making it the preferred NSBB for patients with compensated cirrhosis[42]. For patients intolerant to NSBBs or with contraindications and high-risk varices, endoscopic variceal ligation (EVL) aiming to eradicate varices is critical to primary prophylaxis[4,10,11]. Once varices are eradicated, patients should undergo endoscopic examination every 3-6 months for 1 year and then annually thereafter to assess variceal recurrence and retreatment[4]. There is relatively little research on the primary prophylaxis of gastric variceal bleeding. A randomized controlled trial (RCT) found that endoscopic cyanoacrylate injection (ECI) more effectively prevented bleeding of high-risk GVs than NSBBs, but it did not improve survival[43]. Therefore, the application of ECI remains controversial, and it is currently recommended only for high-risk GVs in patients with NSBB contraindications or intolerance.

Because of the lack of large-sample RCTs, the role of interventional treatment methods such as balloon-occluded retrograde transvenous obliteration (BRTO) or transjugular intrahepatic portosystemic shunt (TIPS) in primary prophylaxis is not well defined, and thus, they are currently not recommended for preventing bleeding in patients with EGV[4,10,11].

Ultrasound elastography

TE and SWE are the primary ultrasound elastography techniques used to assess EGV and EGVB risk. Numerous studies illustrated that LSM and SSM as determined by TE are related to the severity of EGV and the risk of EGVB[21,44]. The latest consensus suggests that for patients with compensated cirrhosis, LSM ≥ 20 kPa or platelet count ≤ 150 × 10⁹/L can be used as a standard for further endoscopic screening of varices. For patients requiring EGD, SSM ≤ 40 kPa can exempt them from endoscopy[4]. Therefore, non-invasive testing using the platelet count, LSM, and SSM might be useful in identifying patients requiring EGD for varices. Liu et al[45] used a 100-Hz probe for SSM and found that it had high accuracy and a low miss rate in predicting high-risk varices. Danish et al[46] conducted a prospective study using SWE to test LSM in patients with chronic liver disease and found that liver elastography was a valuable, non-invasive tool for diagnosing varices with a sensitivity and specificity of 44.90% and 51.90%, respectively. Additionally, a meta-analysis evaluating the diagnostic performance of SWE in predicting high-risk esophageal varices (HREVs) revealed that LSM and SSM measured by SWE had high accuracy in predicting HREVs in patients with cirrhosis, highlighting these indices as potential alternatives to EGD[47]. However, there is controversy regarding the cutoffs of LSM and SSM for assessing EGV, and high-quality clinical studies are needed for validation.

CT and MRI

Yang et al[48] conducted a study aiming to non-invasively predict HREVs, establishing a predictive model based on CT-measured liver and spleen volumes. The developed model had better discriminative ability and clinical efficacy in the non-invasive prediction of HREV than previously reported models. Salahshour et al[49] performed enhanced multidetector CT in 124 patients with cirrhosis and found that the presence and diameter of the gastric coronary vein, short gastric vein, and paraesophageal vein were significantly related to the occurrence of EGVB. Based on these findings, they constructed a predictive model for EGVB risk, achieving 75.86% sensitivity and 76.92% specificity. Dual-energy CT, being an emerging imaging technique, is gradually being applied in clinical practice. Hong et al[50] reported that the hepatic extracellular volume fraction based on dual-energy CT combined with the platelet count can be used to predict high-risk varices in patients with cirrhosis, reducing the need for EGD.

MRI has a role similar as CT in the morphological diagnosis of EGV, but it offers the unique advantage of using MR elastography to assess EGVB risk. A study involving 93 patients using MR elastography to measure liver and spleen stiffness and spleen volume found that, compared with LSM and the spleen volume, SSM had a stronger correlation with severe EGV[51].

Deep learning

Deep learning technology has displayed great applicability in the diagnosis and risk assessment of EGV, significantly improving diagnostic accuracy and efficiency. Yan et al[52] developed and validated a radiomics model based on machine learning to diagnose HREV in patients with cirrhosis. Compared with Baveno VI and its extended criteria, this model improved the diagnostic rate by 49.0% and 32.8%, respectively. Chen et al[53] trained and validated a real-time DCNN system called “ENDOAGGEL” for diagnosing EGV and predicting rupture risk. The results illustrated that ENDOAGGEL had higher accuracy than endoscopists in detecting esophageal varices (EVs) (97.00%) and GVs (92.00%). Additionally, ENDOAGGEL accurately identified endoscopic risk factors for EGVB. Therefore, the application of DCNNs will help endoscopists evaluate EGV more objectively and accurately. Lee et al[54] used deep learning to establish a model that effectively predicted the risk of high-risk varices and their rupture bleeding in patients with hepatitis B-related cirrhosis by combining the hepatic and splenic volumes with clinical indicators. The study found that when the ratio of splenic volume to platelets was < 1.63, the possibility of high-risk varices could be dismissed, thus avoiding unnecessary EGD. Specifically, the model demonstrated a sensitivity of 69.4% and specificity of 78.5% for detecting high-risk varices at the cutoff (> 3.78) that balanced sensitivity and specificity.

TREATMENT OF ACUTE EGVB

The general management principles for acute EGVB include airway management, early prophylactic antibiotic and vasoactive drug use, blood volume restoration and maintenance, and restrictive red blood cell transfusion[4,10,11]. All patients should receive early prophylaxis with antibiotics, as studies indicated that prophylactic antibiotics can reduce infection rates and mortality in patients with EGVB. For patients with late-stage cirrhosis who previously received quinolones, cephalosporins (such as intravenous ceftriaxone) are the preferred antibiotics[4,10]. Once acute bleeding is suspected in patients with EGVs, vasoactive drugs should be administered as early as possible and continued until 3-5 days after endoscopic treatment[55]. Once hemodynamic stability is achieved, CT or MRI should be performed as early as possible to clarify the structure of portal venous collaterals and the presence of portal venous thrombosis[4,10,55].

For patients suspicious for acute EGVB, EGD should be performed within 12 hours of admission. For patients with acute EV bleeding, EVL is recommended as the first-line therapy[4]. To date, some small RCTs have reported that ECI more effectively controls acute GV bleeding than EVL, and ECI is the preferred endoscopic treatment for acute GV bleeding internationally, achieving hemostasis success rates of 87%-100%[56]. Because of complications such as glue excretion ulcers and embolization, ECI should be performed by physicians with extensive experience, and the selection of endoscopic treatment should be based on the professional knowledge and technology of local medical institutions[57]. Ultrasound endoscopy-guided ECI is a new endoscopic treatment method that can more clearly display the blood vessels supplying GVs, achieving visualized and precise ECI and thereby improving hemostasis rates and reducing the risk of embolization. However, high-quality studies proving the effectiveness of ECI against acute GV bleeding are lacking[58,59].

In total, 10%-15% of patients with acute EV bleeding might experience rebleeding after initial endoscopic hemostasis. Such high-risk rebleeding is observed in patients with Child-Pugh grade C (< 14) or Child-Pugh grade B (> 7) with endoscopic active bleeding[60,61]. For such patients, early TIPS is recommended as a more effective method to prevent rebleeding. TIPS is also an effective treatment for acute bleeding in gastroesophageal varices type 2 (GOV2) and isolated gastric varices type 1 (IGV1)-type GVs. For patients with high-risk acute GV bleeding, early TIPS should be performed within 72 hours (ideally within 24 hours). Because GVs usually bleed at low portal pressure, TIPS combined with gastric coronary vein embolization can further reduce rebleeding and hepatic encephalopathy rates[62]. For refractory EGVB unresponsive to medical and endoscopic therapy, rescue TIPS can be applied, and hemostasis rates of 90%-100% have been attained[63,64]. For patients with GOV2 and IGV1-type GVs and gastrorenal shunts, BRTO is also a safe and effective measure, and it is recommended as an alternative therapy in the presence of TIPS contraindications (such as hepatic encephalopathy and advanced liver failure)[65,66].

Deep learning

A recent study based on enhanced CT developed a machine learning model to predict the risk and prognosis of death and other endpoint events in 330 patients with acute EGVB enrolled from three clinical centers. Researchers used a trained deep learning model to segment the liver and spleen and then constructed a liver–spleen model based on the extracted original images. The AUC of this model was 0.782 (95%CI = 0.650-0.882) in the internal test group and 0.789 (95%CI = 0.674-0.878) in the external test group. Calibration curve and decision curve analysis illustrated that the newly constructed machine learning model outperformed traditional clinical scoring systems[67].

A retrospective study based on a 5-year dataset developed a machine learning model to predict the occurrence of OHE after TIPS for acute variceal bleeding. The model, which utilized logistic regression, demonstrated a high performance with an AUC of 0.825. Additionally, both the actual OHE status and predicted OHE value were significant predictors in each Cox model, with the model-predicted OHE achieving an AUC of 88.1 in survival prediction. The results demonstrated that this model accurately predicted the risk of post-TIPS OHE and outperformed traditional models, supporting its application in improving the prognosis of patients with EGVB[68]. Other models for AI-guided treatment decisions for cirrhotic variceal bleeding are currently being developed and validated. For example, Qi et al[30] are developing and validating AI-driven model to guide the management of patients with liver cirrhosis and variceal bleeding.

SECONDARY PROPHYLAXIS OF EGVB

After acute EGVB has been controlled, the probability of rebleeding and death remains high. For untreated patients, the rebleeding rate within 1-2 years is 60%[4,10]. Currently, NSBB combined with EVL is recommended as the first-line treatment for preventing EV rebleeding[4]. A long-term follow-up RCT reported that EVL combined with propranolol or carvedilol was more effective in preventing EV rebleeding than EVL or NSBB alone[69]. A retrospective study of 87 patients with cirrhosis and HVPG > 12 mmHg compared the effects of propranolol plus EVL and carvedilol plus EVL in variceal secondary prophylaxis and found that carvedilol exhibited greater efficacy than propranolol in reducing HVPG and carried a lower rebleeding rate[70]. These studies suggest that NSBB therapy combined with EVL is superior to monotherapy and that carvedilol is the preferred NSBB.

TIPS is also an effective method for preventing rebleeding in patients with varices. An RCT involving 72 patients with a history of variceal bleeding revealed that covered stent TIPS achieved a lower rebleeding rate than NSBB therapy combined with EVL. However, this strategy did not improve survival, and it was associated with a higher incidence of hepatic encephalopathy[71]. Currently, the Baveno VII consensus recommends TIPS as a rescue treatment after the failure of NSBBs or carvedilol combined with EVL[4]. The American Association for the Study of Liver Diseases guidelines state that TIPS can be considered as a first-line treatment for secondary prophylaxis in patients with TIPS indications (such as recurrent or refractory ascites)[10]. Concerning patients with GOV2 and IGV1-type GV, a study found that TIPS treatment had a lower rebleeding rate than ECI, whereas no differences in survival and hepatic encephalopathy rates were observed[72]. BRTO is also an effective and safe measure for preventing GV rebleeding. A meta-analysis indicated that BRTO had lower cumulative rebleeding (10.6% vs 18.7%) and hepatic encephalopathy rates (0.00% vs 23.1%) than TIPS, but it was more likely to worsen ascites (22.4% vs 4.3%)[73]. Although BRTO is not as widely used as TIPS, for patients with GV together with significant gastrorenal shunts and TIPS contraindications (such as hepatic encephalopathy or liver failure), BRTO should be considered as the preferred treatment for secondary prophylaxis. However, there are few high-quality studies in this area, and further research is needed.

Deep learning

In a recent multicenter retrospective study involving 727 patients who underwent EVL, Cao et al[74] developed an endoscopic virtual ruler based on AI to measure the EV diameter during endoscopy to identify patients suitable for EVL. Patients were divided into early rebleeding and non-rebleeding groups based on whether rebleeding occurred within 6 weeks after surgery. Multivariate binary logistic regression analysis identified Child-Pugh grade C (P = 0.007), Japanese variceal grade F3 (P = 0.009), and EV diameter (P < 0.001) as potential predictors of early rebleeding after EVL. The AUC for EV diameter was 0.848, vs 0.635 for the Japanese variceal grade (P < 0.001). Therefore, this study concluded that the EV diameter better predicts early rebleeding after EVL than the Japanese variceal grading standard. An EV diameter of ≥ 1.4 cm increases the risk of early rebleeding after EVL, and EVL should be used cautiously[74]. Currently, further research on the application of AI in the secondary prophylaxis of EGVB is needed.

CHALLENGES AND FUTURE PERSPECTIVES

AI demonstrates remarkable precision and reliability, enhancing clinical safety and excelling in diagnosing and continuously monitoring portal hypertension and EGV. It offers significant advantages over conventional non-invasive diagnostic tools. Nevertheless, this technology is in its early stages, and it faces numerous challenges in practical application[15].

The implementation of AI technologies in diverse clinical settings presents distinct challenges. In resource-limited environments, constraints related to computational infrastructure, data availability, and healthcare provider training might hinder the effective deployment of AI tools. Conversely, high-resource settings might face different obstacles such as integration with existing electronic health record systems and navigating complex regulatory requirements[75]. Current AI applications also face technical limitations that can affect their reliability in clinical settings. Variability in imaging quality, motion artifacts, and differences in equipment across healthcare facilities can introduce noise into the data, reducing diagnostic accuracy. Furthermore, the dynamic nature of liver disease and the potential for rapid changes in portal pressure attributable to acute decompensation events pose challenges for longitudinal risk prediction[76]. Therefore, data quality and standardization are fundamental to the successful application of AI in clinical practice. Multicenter collaborations and adherence to standardized data collection protocols are essential to generate diverse, representative datasets. Furthermore, implementing rigorous data cleaning and augmentation techniques can enhance the robustness and generalizability of AI models across different healthcare environments[77]. The implementation of advanced AI solutions, particularly deep learning models, demands significant computational resources. Cloud-based computing platforms offer scalable solutions for healthcare institutions lacking dedicated hardware infrastructure[78].

Cost-effectiveness analyses suggest that AI-driven diagnostic tools can offer significant value, particularly in resource-constrained settings. By reducing the need for invasive procedures such as HVPG measurement and decreasing the frequency of endoscopic surveillance in low-risk patients, these technologies could potentially optimize resource allocation[79]. This requires comprehensive training programs for healthcare providers to effectively implement AI technologies in clinical practice. Interdisciplinary collaboration and continuing medical education initiatives will be critical to ensure that providers can integrate AI into their workflows while maintaining clinical judgment and patient-centered care.

The widespread adoption of AI technologies faces several barriers. Financial constraints related to software licensing, hardware acquisition, and maintenance might limit access in low-resource regions. Resistance to change from healthcare providers accustomed to traditional diagnostic and treatment approaches represents another significant obstacle. Additionally, concerns about data privacy, cybersecurity threats, and potential job displacement perceptions among medical staff might generate skepticism toward AI integration. Therefore, the future development of machine learning algorithms, especially in deep learning, needs to tackle current issues such as high data demands and insufficient model transparency. In addition, combining multimodal data sources (e.g., images, laboratory values, genetic information, clinical history) can boost the accuracy and usefulness of AI tools in predicting complications and guiding portal hypertension treatment. Furthermore, optimizing algorithms to reduce computational intensity without losing diagnostic accuracy represents a key research area, particularly for applications in low-resource settings. Developing robust data preprocessing pipelines and adaptive algorithms is essential to handle the complexity and variability of real-world clinical data.

CONCLUSION

EGVB remains a life-threatening complication of cirrhosis associated with high rebleeding rates and mortality. Although traditional therapies such as pharmacological interventions, endoscopic techniques, and vascular procedures have been mainstays of management, AI is emerging as a transformative force in this field. AI-driven tools offer promising solutions for the early detection of CSPH, risk stratification of variceal bleeding, and implementation of individualized treatment strategies. However, broader clinical adoption requires addressing challenges related to data quality, model interpretability, and regulatory frameworks. Future research should focus on refining non-invasive diagnostic algorithms, expanding AI applications in therapeutic decision-making, and validating these tools through large-scale multicenter trials. The integration of AI into clinical workflows could potentially revolutionize the management of portal hypertension complications, but careful consideration of implementation barriers and equity in healthcare delivery is essential to maximize its benefit for patients globally.