Future directions in noninvasive prediction of cirrhosis decompensation: An opinion review

Abstract

Cirrhosis is the terminal stage of chronic liver disease. The characteristic manifestations of decompensated cirrhosis are portal hypertension (PH) and complications arising from liver dysfunction, at which stage patient survival is markedly reduced. Hepatic venous pressure gradient serves as the gold standard for assessing PH, with values reaching or exceeding 10 mmHg defining clinically significant PH. Nevertheless, hepatic venous pressure gradient measurement is invasive and limited by high cost, restricted accessibility, and operator dependence, which hinder its routine clinical application. Liver stiffness and spleen stiffness measured via transient elastography have emerged as essential tools for noninvasive risk stratification in chronic liver disease. However, existing noninvasive predictive models are constrained by single-center retrospective designs, homogeneous etiologies, and insufficient standardization of spleen stiffness measurement, and their applicability in patients with hepatocellular carcinoma remains uncertain. Furthermore, the neglect of competing risks in traditional survival analyses has led to overestimation of hepatocellular carcinoma risk. Although technologies such as radiomics and machine learning show considerable promise, prospective multicenter validation remains necessary. Further progress will depend on standardized multi-etiology and multicenter cohorts, supported by rigorous internal and external validation, to produce a universal noninvasive tool for clinical decision-making.

Key Words: Decompensated cirrhosis; Noninvasive prediction model; Spleen stiffness measurement; Hepatocellular carcinoma; Radiomics; Baveno VII criteria

Core Tip: The noninvasive model integrating liver stiffness and spleen stiffness has demonstrated high accuracy in predicting clinical decompensation in patients with viral cirrhosis. Future research should prioritize high-risk subgroups, particularly patients with hepatocellular carcinoma. Standardized protocols for spleen stiffness measurement must be established, and multistate models should be adopted to better characterize disease progression. Future models that combine clinical complexity with radiomics and artificial intelligence-enhanced surveillance may sharpen personalized risk assessment in decompensated cirrhosis.

INTRODUCTION

The natural history of cirrhosis is conventionally divided into compensated and decompensated stages, with the onset of ascites, variceal bleeding, or hepatic encephalopathy marking a sharp decline in prognosis[1,2]. Cirrhosis represents the terminal stage of chronic liver disease[3], at which point patient survival is markedly reduced once decompensation occurs[4]. Because most compensated patients eventually progress to decompensation, predicting decompensation offers greater clinical value than estimating mortality alone[5]. Portal hypertension (PH) is the predominant pathophysiological mechanism underlying decompensation, and hepatic venous pressure gradient (HVPG) measurement is the gold standard[6,7]. A value of ≥ 10 mmHg defines clinically significant PH (CSPH) and identifies individuals at heightened risk of decompensation[8]. The invasiveness, cost, limited availability, operator dependency, and technical variability of HVPG measurement restrict its routine application and highlight the need for reliable noninvasive alternatives, as compared in Table 1[9-13]. Liver stiffness measurement (LSM) and spleen stiffness measurement (SSM) via transient elastography (TE) have emerged as essential tools for noninvasive risk stratification in chronic liver disease[14]. However, existing noninvasive predictive models are constrained by single-center retrospective designs, homogeneous etiologies, and insufficient standardization of SSM[15], and their applicability in patients with hepatocellular carcinoma (HCC) remains uncertain[16]. Furthermore, the neglect of competing risks in traditional survival analyses has led to overestimation of HCC risk[17]. This review examines noninvasive tools for predicting decompensation in cirrhosis. We review the strengths and limitations of the Baveno VII consensus, explore the methodological challenges associated with excluding patients with HCC, and discuss the standardization issues in SSM[14]. We also consider how emerging radiomics[18] and artificial intelligence (AI) technologies[19] might help build more equitable and clinically useful decision-support tools.

Table 1 Comparative analysis of hepatic venous pressure gradient and non-invasive models in portal hypertension: Practical considerations for clinical deployment.

Parameter	HVPG measurement	Noninvasive indicators
Invasiveness and safety	As an invasive vascular procedure, HVPG measurement is associated with inherent risks of complications (e.g., access site hematoma, infection, or, rarely, more serious events)	Inherently non-invasive, eliminating procedure-related risks. This profile results in superior patient compliance and excellent suitability for serial monitoring
Cost-effectiveness	The high per-procedure cost imposes a significant economic burden on healthcare systems and patients, particularly when repeated assessments for treatment monitoring are necessary	Substantially more cost-effective. The lower cost per examination alleviates the overall economic burden, making these modalities ideal for large-scale screening and long-term surveillance
Accessibility and infrastructure	Requires a specialized setting, including highly trained interventional radiologists/hepatologists and a dedicated catheterization laboratory. Its implementation in primary care or resource-limited settings is not feasible	High accessibility. Techniques such as vibration-controlled transient elastography (e.g., FibroScan) have a shorter learning curve, and with increasingly portable devices, can be widely deployed in outpatient and primary care settings
Standardization and reproducibility	The accuracy and reproducibility of HVPG measurement are highly operator-dependent. Key sources of variability include: Technique: Measurements obtained using balloon catheters (recommended) may differ from those obtained using straight catheters. Sedation: The use of procedural sedation can alter systemic hemodynamics and pressure readings. Protocol adherence: Poor adherence to the standard protocol – specifically, failing to perform the required three separate measurements – introduces avoidable error	Sources of variability are distinct and often more manageable. Liver stiffness measurement/spleen stiffness measurement: Influenced by operator experience and patient factors (e.g., obesity, ascites, active hepatic inflammation). Standardized operating protocols are widely established and easier to implement uniformly. Serum models: Variability primarily stems from inter-laboratory assay differences, which can be mitigated through standardized calibration. Overall, non-invasive methods are more amenable to standardization and are associated with lower inter-center variability compared to HVPG

HVPG: Hepatic venous pressure gradient.

NONINVASIVE PREDICTION MODELS AND THE BAVENO VII CONSENSUS

The Baveno VII consensus has formally integrated SSM by TE into the clinical decision for patients with cirrhosis. In patients who meet Baveno VI endoscopic screening criteria (liver stiffness ≥ 20 kPa or platelet count ≤ 150 × 10⁹/L) but are not candidates for nonselective beta-blockers, an SSM ≤ 40 kPa identifies a low probability of high-risk varices, allowing endoscopy to be safely avoided. For the assessment of CSPH, SSM < 21 kPa helps rule out the condition, whereas SSM > 50 kPa is highly indicative of its presence[3]. The Baveno VII criteria has been confirmed in multiple independent cohorts[20-22]. In a study of patients with hepatitis B virus-related cirrhosis, Zhang et al[23] demonstrated that the Baveno VII criteria outperformed earlier noninvasive models in ruling out high-risk esophageal varices. Dajti et al[24] showed that incorporating SSM into the Baveno VII criteria improves the noninvasive diagnosis of CSPH. The Baveno VII consensus offers recommendations for patients with SSM < 21 kPa or > 50 kPa but leaves considerable uncertainty for those with values between 21 kPa and 50 kPa. Comparisons with Baveno VII criteria will determine whether prediction models offer meaningful improvements in risk stratification for PH.

HCC EXCLUSION AND COMPETING RISKS

The vast majority of HCC cases originate from the common pathological basis of liver cirrhosis, the two conditions frequently coexist, presenting complex challenges in clinical management[25]. HCC represents an advanced stage of cirrhosis and is a leading cause of mortality in patients with decompensated disease[7,26]. HCC development is influenced by PH; notably, Ripoll et al[27] demonstrated that PH independently predicts HCC development in cirrhotic patients without varices or ascites, regardless of liver dysfunction severity or disease duration. However, the presence of HCC changes the evaluative relationship between noninvasive alternatives and PH[28]. As Allaire et al[29] reported, the correlation between LSM and HVPG is weakened in this population. LSM no longer reliably reflects PH but is substantially confounded by the tumor's own characteristics. Consequently, LSM does not reliably reflect the actual severity of PH in these patients. Whether existing noninvasive prediction models remain valid in this critical subgroup, namely compensated cirrhosis with concurrent HCC, remains an open question[30].

Moreover, the application of the Baveno VII criteria is also restricted in patients with HCC because the key validation studies explicitly excluded this patient population[31]. While this design choice ensures a more homogeneous derivation cohort, it fundamentally limits the generalizability of these criteria to the complex clinical setting of HCC. The nonviral etiologies prevalent in patients with HCC display distinct patterns of decompensation risk compared with those in validated cohorts[32,33]. Whether current noninvasive prediction models retain their accuracy in patients with compensated cirrhosis and concurrent HCC remains unverified. Likewise, it is unclear whether optimal diagnostic thresholds require adjustment for tumor size, number, or location, or whether anti-tumor therapies themselves might confound model performance by altering the hepatic mechanical microenvironment or reducing tumor burden[34-36]. The noninvasive prediction models should not be directly applied to patients with HCC without dedicated prospective validation. Future studies should therefore prospectively enroll this high-risk, heterogeneous subgroup and undergo rigorous external validation to develop more reliable tools for risk stratification.

REPRODUCIBILITY AND STANDARDIZATION OF SSM

The reproducibility and standardization of SSM are essential for the validity of SSM-based predictive models[37]. Conventional imaging modalities (e.g., ultrasonography, computed tomography, magnetic resonance imaging) can evaluate morphological features, including liver size, nodularity, and signs of PH such as ascites, providing indirect evidence of cirrhosis and PH, but they cannot directly quantify LSM and SSM[38]. SSM-based models rely on TE, which estimates tissue stiffness, and international guidelines recommend a quality criterion of interquartile range/median ≤ 30%[39]. However, reproducibility remains unreported, including inter-operator variability, and the dependence of SSM on technique and equipment has not been fully characterized[40,41]. Furthermore, SSM is more susceptible than LSM to factors such as probe type, ascites, obesity, and operator experience. Notably, the use of an SSM^@50 Hz transducer may overestimate stiffness values[42]. Recently, an SSM^@100 Hz probe specifically designed for SSM has been developed, and its measurement accuracy has been significantly enhanced. However, the optimal cutoff values for this probe across TE, point shear wave elastography, and two-dimensional shear wave elastography require further validation[43-46]. Where TE is not accessible, validated serological indices will be required to extend the clinical application of SSM-based models[9,47]. Unified standard operating procedures across institutions will be equally important for multicenter comparability and reproducibility. Future studies should evaluate measurement variability across different probes. Guidance is also needed to identify patients less suited to SSM, particularly those with large-volume ascites or severe obesity.

RADIOMICS AND AI

In recent years, radiomics has been increasingly applied in the diagnosis of liver cirrhosis and the prediction of its complications[48,49]. Radiomics involves extracting high-throughput, quantitative features from medical images; analyzing their associations with clinical outcomes using data mining techniques; and constructing predictive models based on these features[50]. Liu et al[18] developed the model for predicting CSPH by utilizing radiomics features from both the liver and the spleen, which demonstrated strong diagnostic performance. In a large-scale prospective global cohort study, Silvey et al[19] reported that compared with traditional models such as the Child-Pugh and model for end-stage liver disease scores, machine learning algorithms integrating diverse clinical and laboratory data achieved greater accuracy in predicting mortality in hospitalized cirrhotic patients. These findings underscore the potential of radiomics-based approaches in facilitating individualized patient management and enhancing outcomes in patients with liver disease. Further integration of radiomic features with clinical variables through machine learning may yield more robust and accurate noninvasive predictive models[51,52]. Multicenter prospective validation, comparisons with Baveno VII criteria, and practical tools for routine clinical use are needed to define the reliability and generalizability of these approaches across different etiologies and populations.

LIMITATIONS OF CURRENT PREDICTION MODELS

Recently, numerous studies have sought to incorporate noninvasive indicators into predictive models as an alternative to HVPG measurement. Nevertheless, the single-center, retrospective design and exclusive inclusion of viral hepatitis-related cirrhosis could introduce selection and spectrum bias, restricting the generalizability of the model to other etiologies or demographic groups[53]. Although some models have undergone external validation, small sample sizes or close demographic and clinical resemblance between training and validation cohorts limit these efforts. The generalizability of these models in diverse healthcare environments and multi-etiological backgrounds still requires further assessment. Additionally, the risk of model overfitting has not been effectively controlled[54,55].

The clinical application of noninvasive predictive models also raises important equity concerns[56]. Several factors may undermine model performance in real-world settings, including unequal access to elastography, inconsistent imaging protocols and quality standards, regional variation in patient demographics and underlying etiologies, and divergent clinical workflows between primary care and specialized centers[57-59]. Consequently, models developed and validated in high-resource settings may not be directly applicable to under-resourced or community healthcare environments without rigorous external validation. The ICE-I framework proposed by Yu[60] addresses this concern directly, stating that the application of AI in clinical practice demands attention to equity and integration challenges beginning with the initial stages of model development rather than during the validation phase. Such efforts are essential to ensure that the benefits of noninvasive risk stratification are distributed more equitably among patients with cirrhosis across diverse populations.

NONINVASIVE STRATIFICATION AND ENDOSCOPIC SCREENING

Noninvasive predictive models are now recommended as first-line tools for risk stratification in patients with compensated advanced chronic liver disease (cACLD), improving patient comfort and optimizing healthcare resource use[61,62]. However, for patients consistently classified as high risk by these noninvasive models, endoscopic examination remains the gold standard for diagnosing varices and preventing first or recurrent variceal bleeding[63]. Notably, AI-assisted endoscopic technology has introduced new possibilities into this field, enhancing diagnostic accuracy, reducing interoperator variability, and improving both the safety and yield of screening[64,65]. Patients with cACLD should be assessed with LSM and SSM by TE, ideally with a 100 Hz spleen-dedicated probe[66]. In patients with contraindications to or intolerance of nonselective beta-blockers, an SSM ≤ 40 kPa without other warning signs identifies a low probability of high-risk varices according to Baveno VII criteria, allowing endoscopic screening to be safely deferred; those with SSM < 21 kPa may be further classified as having a low probability of CSPH. For patients with SSM in the range of 21-50 kPa, or for those with discordant LSM and SSM findings, validated integrative models or radiomics-based approaches may refine risk assessment. Patients with SSM > 50 kPa, or those persistently identified as high risk by other models, should proceed to endoscopic screening and receive nonselective beta-blockers or endoscopic variceal ligation as clinically indicated[63]. We propose a tiered strategy, derived from the Baveno VII consensus, as a framework for noninvasive risk stratification in cACLD, as illustrated in Figure 1. This approach awaits prospective validation across diverse populations and clinical settings.

Figure 1 Noninvasive stratification and endoscopic screening. EVL: Endoscopic variceal-ligation; LSM: Liver stiffness measurement; NSBB: Non-selective β-blockers; SSM: Spleen stiffness measurement; TE: Transient elastography.

CONCLUSION

The Baveno VII consensus integrates SSM into the clinical decision for cACLD. Noninvasive predictive models reduce reliance on invasive HVPG and show promise for contributing to PH management. Notably, the exclusion of patients with HCC and the lack of competing risk analysis limit its applicability in complex clinical scenarios. Future research should integrate radiomics and machine learning algorithms to develop multi-state models, thereby enabling more precise risk stratification and better informing clinical decision-making.