Graphical Abstract
INTRODUCTION
Portal hypertension is a major complication of cirrhosis, leading to significant morbidity and mortality due to its consequences, including ascites, esophageal varices (EV), variceal hemorrhage, and overt clinical decompensation [1-4]. Notably, EV occurs in approximately 50% of patients with cirrhosis [5]. Variceal hemorrhage remains a life-threatening condition, with a six-week mortality rate of 15–25% following an acute EV hemorrhage and a one-year rebleeding risk of up to 60% without prophylaxis [6-8]. Recently, the American Association for the Study of Liver Diseases practice guidelines recommended that patients with high-risk varices (HRV) undergo endoscopic surveillance and receive endoscopic variceal ligation to prevent variceal hemorrhage [2]. This highlights the importance of timely and accurate identification of HRV, which is crucial for optimizing clinical management and improving patient outcomes.
Esophagogastroduodenoscopy (EGD) is considered the gold standard for diagnosing EV [9]. However, a large percentage of patients undergoing screening, particularly those with compensated cirrhosis, do not have EV or have only small EV without high-risk features, and therefore, do not require prophylactic therapy [10]. These patients would not benefit from EGD but may suffer complications from the procedure. Therefore, developing a novel non-invasive tool for better detecting EV and predicting the risk of bleeding is urgently needed, ultimately reducing the burden of unnecessary EGDs.
Over the last decade, laboratory marker panels and elastography methods have received significant attention in routine clinical practice. The fibrosis-4 (FIB-4) score and liver stiffness measurement (LSM) are commonly used non-invasive tests for liver fibrosis in patients with chronic liver disease [11,12]. As noted in the Baveno VI consensus guideline, utilizing LSM (<20 kPa) in combination with platelet (PLT) count (>150,000/μl) can identify patients with a very low likelihood of having HRV, making EGD screening unnecessary for these individuals [13]. A recent study has demonstrated the value of FIB-4 in predicting the development of liver cirrhosis-related outcome events [14]. Additionally, studies have confirmed that spleen stiffness measurement (SSM) correlates with portal hypertension and is considered as a promising non-invasive alternative to EGD [15,16]. Indeed, the most recent Baveno VII consensus guideline has recommended that SSM be considered an additional tool for assessing portal hypertension and HRV [17].
Despite these strong correlations, the ability of FIB-4, LSM, and SSM to predict EV is only moderate when considering each score separately [18,19]. We proposed, for the first time, the hypothesis that combining these three non-invasive markers can improve the overall performance, as FIB-4, LSM, and SSM are associated with different and potentially complementary aspects of liver characteristics.
To test this hypothesis, we conducted a multicenter study incorporating the FIB-4 score (or its individual components as separate predictors), LSM, and SSM into a single model, termed FIB-4plus, to detect EV and HRV in patients with compensated cirrhosis whose EV severity was assessed using EGD. We used a novel machine learning method to develop and validate the FIB-4plus score for predicting EV and HRV in patients with compensated cirrhosis, aiming to provide a non-invasive and costeffective alternative to EGD that enables the accurate identification of EV and HRV in this population.
MATERIALS AND METHODS
Study population and design
This international multicenter study included patients with compensated cirrhosis who underwent EGDs at 17 hospitals in China and one university hospital center in Croatia. Patients with compensated cirrhosis were prospectively recruited from 14 Chinese hospitals between October 2020 and December 2022 to form a training cohort for the development of the FIB-4plus score. It is retrospective analysis of prospectively collected data. Two additional independent cohorts were established for external validation: validation cohort 1, which comprised patients with compensated cirrhosis prospectively enrolled in two Chinese hospitals between March 2016 and December 2022, and validation cohort 2, which comprised patients with compensated cirrhosis retrospectively enrolled in a Croatian University Hospital Center between January 2016 and December 2022. These three cohorts included patients who underwent LSM and SSM using twodimensional shear wave elastography (2D-SWE).
In addition, a cohort of eligible patients with compensated cirrhosis who underwent transient elastography (TE) to measure LSM and SSM was prospectively recruited from four Chinese hospitals between August 2021 and March 2022 to further validate the utility of our model in detecting EV and HRV in these patients.
This study was approved by the ethics committee of the principal investigator’s hospital (approval number LDYYLL2020-246) and was registered with ClinicalTrials. gov (NCT04546360). All patients signed an informed consent form.
Inclusion and exclusion criteria
Adult participants (≥18 years) with compensated cirrhosis, confirmed by compatible clinical, liver histological biopsy, or radiologic findings, who underwent both EGD and 2D-SWE (training cohort, validation cohorts 1 and 2) or TE (TE cohort) for measuring LSM and SSM, were included. All patients were required to have time between 2D-SWE or TE measurements, laboratory tests, and EGD within 1 month. The exclusion criteria were as follows: (1) patients with previous or current decompensation (defined as cirrhotic patients with ascites, variceal hemorrhage, or hepatic encephalopathy, etc.); (2) individuals with noncirrhotic etiologies for por tal hyper tension; (3) unqualified acquisitions for LSM or SSM, as described in our previous studies [20,21]; (4) patients who experienced cirrhosis regression following antiviral therapy (histology showing fibrosis stage <F4); (5) those who received nonselective beta-blockers or underwent endoscopic variceal ligation; and (6) those with concurrent hepatocellular carcinoma.
LSM and SSM measurements
In training and validation cohorts 1 and 2, both LSM and SSM were performed using the Aixplorer 2D-SWE system (SuperSonic Imagine, Aix-en-Provence, France) with an abdominal 3.5 MHz curved array probe (SuperSonic Imagine). In the TE cohort, LSM and SSM were performed using Fibroscan® (Echosens, Paris, France) equipped with an M or XL probe (Echosens) for LSM and a 100 Hz specific TE-probe (Echosens) for SSM. The patients fasted for at least 6 hours before scanning. Both 2D-SWE and TE examinations were conducted by experienced operators following standardized procedures; further details are provided in Supplementary Materials and Methods, Parts I and II.
Endoscopic evaluation of EV and HRV
Standard endoscopies were conducted by expert operators who were blinded to the clinical and laboratory data of the study population [21]. Details are described in the Supplementary Materials and Methods, Part III. According to the Baveno VII consensus guidelines, HRV is defined as large varices (varix size, ≥5 mm), or red signs, or Child– Pugh C [17].
Data collection
Clinical and laboratory data, including demographic characteristics, 2D-SWE or TE examination, routine blood tests, liver and renal biochemistry, coagulation, and etiology of cirrhosis, were recorded. Liver disease severity was evaluated by calculating the Child–Pugh score and the model for end-stage liver disease score [22]. The FIB-4 score was calculated as follows: FIB-4=age (years)×aspartate aminotransferase (AST, U/L)/(PLT [109/L]×alanine aminotransferase [ALT, U/L]1/2) [23].
Overview of machine learning models
Several approaches and strategies have been employed to construct a new model by combining LSM, SSM, and the FIB-4 score to develop an optimal FIB-4plus score. LSM, SSM, and FIB-4 values were combined as continuous variables to develop the model. In addition, we developed models that included the individual components of the FIB-4 score (i.e., age, AST, ALT, and PLT) as separate predictors rather than using the composite FIB-4 score as a single predictor. Additionally, univariable and multivariable logistic regression (LR) analyses were conducted to evaluate the significant associations between each variable of the FIB-4plus score (i.e., age, AST, ALT, PLT, LSM, and SSM) and the outcome, with all results showing statistical significance (all P<0.01) (Supplementary Materials and Methods, Part IV).
In this study, two machine learning algorithms, LR and eXtreme Gradient Boosting (XGBoost), were used to develop the LR-FIB-4plus and XGBoost-FIB-4plus models, respectively. Here, we briefly describe the two algorithms. LR is a widely used machine learning algorithm, which is used to solve classification problems, particularly binary classification problems. It predicts the probability that a single sample belongs to a certain class by utilizing a logistic function (usually the Sigmoid function) to map the output value of a linear regression between 0 and 1. The differences between statistical LR and machine learning LR are detailed in the Supplementary Materials and Methods, Part V. We developed an alternative model that integrated LSM, SSM, and the individual components of the FIB-4 score (i.e., age, AST, ALT, and PLT) as independent predictors, using statistical LR to predict EV and HRV. The performance metrics of this statistical LR model are detailed in the Supplementary Materials and Methods, Part VI. XGBoost, introduced by Chen and Guestrin [24] in 2016, is a prevalent tree-based machine learning ensemble algorithm. It integrates many tree models to form a strong classifier, resulting in better prediction results in terms of efficiency and accuracy [25].
Ten-fold cross-validation was used in the machine learning model-building process. We used the Shapley Additive exPlanations (SHAP) method to interpret the model predictions and determine the effects of features on the predictions. SHAP, a model-agnostic explanation technique, is useful for interpreting machine learning models at the cohort and patient levels. For each feature across all patients, the SHAP values were separately aggregated and averaged to measure the importance of the features for prediction. The importance plot of the SHAP features illustrates the overall importance of each feature. The larger the mean absolute SHAP value, the higher the importance of the feature for model prediction. The SHAP summary plot demonstrates the effect of each feature on the predictions. A positive SHAP value suggests that the outcome (i.e., HRV) is more likely to be due to the feature, whereas a negative SHAP value implies that the outcome is less likely to be due to the feature.
Outcomes
The primary outcome of this study was the development of a single score (i.e., FIB-4plus) that combines FIB-4, LSM, and SSM to predict HRV in patients with compensated cirrhosis using a machine learning algorithm. The secondary outcome was the prediction of EV, identified using this score.
Statistical analysis
Data are presented as continuous or categorical variables. We first used the Shapiro–Wilk test to determine whether the data followed a normal distribution. Data were given as mean±standard deviation for normally distributed data, while median and interquartile ranges were used for non-normally distributed data. Between-group comparisons were conducted using the t-test or the Mann–Whitney U-test, as appropriate. A two-tailed P-value <0.05 was considered significant. The IBM SPSS Statistics 26 software (IBM Co., Armonk, NY, USA) was used for this study.
The diagnostic performance of the FIB-4plus score was evaluated using several metrics, including accuracy, area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score, in both the training and validation cohorts. Among these metrics, the AUROC and F1-score served as primary indicators for model performance comparison. The DeLong test was used to assess the differences in the AUROCs.
We selected a rule-out cut-off that achieves sensitivity of ≥95% and a rule-in cut-off that achieved specificity of ≥95% for the outcome of HRV. For the outcome of EV, we selected cut-off values that achieve sensitivity or specificity >90%. A subgroup analysis was conducted to evaluate the model’s performance in subgroups stratified by cirrhosis etiology. Additionally, LSM, SSM, PLT, and FIB-4 were used to predict EV and HRV in both the training and validation cohorts. The diagnostic performance of the ruleout and rule-in cut-offs for LSM and SSM in predicting EV and HRV was also assessed. Furthermore, we investigated the performance of the Baveno VI criteria [13] and the Baveno VII algorithm [17] for ruling out HRV using 2D-SWE. The performance of the FIB-4plus score was validated in a TE cohort. To further assess and validate the models’ performance, decision curve analysis (DCA) and calibration curve were employed. The Python software (version 3.7, Python Software Foundation) was used in this study.
RESULTS
Patient characteristics
In total, 502 patients with compensated cirrhosis who underwent EGD in China and Croatia were included in the analysis (Fig. 1). The number of patients included at each center is detailed in the Supplementary Materials and Methods, Part VII. The baseline characteristics of the patients are summarized in Table 1. In the training cohort, the median age was 49.0 years. Among them, 67.2% were male, and 80.2% were diagnosed with hepatitis B virus (HBV) infection. Additionally, 85.8% of the patients were undergoing treatment at the time of the elastography examination. Similarly, in validation cohort 1, the median age was 51.5 years, with 68.6% of the patients being male and 79.7% having an HBV infection. In validation cohort 2, the median age was 62.0 years, and 79.3% of the participants were male, with alcohol-related liver disease being the most common etiology (36.6%). In the TE cohort, both the median age (52.0 years) and the proportion of male participants (67.6%) were similar to those of the training cohort.
Based on EGD, 84 (31.3%), 22 (18.6%), 19 (23.2%), and 9 (26.5%) patients had HRV in the training, the validation cohort 1, the validation cohort 2, and the TE cohorts, respectively. A comparison of the population characteristics of patients with HRV and without HRV is presented in Supplementary Table 1. The comparison of characteristics between patients who underwent 2D-SWE and those who underwent TE is provided in Supplementary Table 2.
Performance of the FIB-4plus score in the prediction of EV and HRV
We evaluated the diagnostic performance of the FIB-4plus score in predicting EV and HRV, as summarized in Table 2 and illustrated in Figure 2. Across all three cohorts, the XGBoost-FIB-4plus score demonstrated superior performance in predicting HRV, achieving an AUROC of 0.927 (95% confidence interval [CI] 0.897–0.957) in the training cohort, and 0.919 (95% CI 0.843–0.995) and 0.902 (95% CI 0.820–0.984) in the first and second external validation cohorts, respectively (Fig. 2). The XGBoost-FIB-4plus score achieved an accuracy, sensitivity, specificity, PPV, NPV, and F1-score of 0.825, 0.726, 0.924, 0.813, 0.880, and 0.767, respectively, in the training cohort (Table 2).
For detecting EV, the XGBoost-FIB-4plus score presented an AUROC of 0.900 (95% CI 0.864–0.935). The XGBoost-FIB-4plus score achieved an accuracy, sensitivity, specificity, PPV, NPV, and F1-score of 0.798, 0.848, 0.748, 0.799, 0.807, and 0.823, respectively, in the training cohort (Table 2, Supplementary Fig. 1). In the validation cohorts, the XGBoost-FIB-4plus score consistently achieved AUROCs exceeding 0.80 for predicting EV.
Both the LR-FIB-4plus and XGBoost-FIB-4plus models showed improved diagnostic performance compared with LSM, SSM, PLT, or FIB-4 alone. Their diagnostic performance for detecting EV and HRV across all cohorts is shown in Supplementary Table 3 and Supplementary Figure 2. For HRV, the AUROCs for LSM, SSM, PLT, and FIB-4 were 0.610 (95% CI 0.539–0.680), 0.792 (95% CI 0.735–0.849), 0.753 (95% CI 0.694–0.813), and 0.686 (95% CI 0.619–0.752) in the training cohort, 0.775 (95% CI 0.662–0.889), 0.881 (95% CI 0.777–0.985), 0.764 (95% CI 0.655–0.874), and 0.773 (95% CI 0.676–0.869) in the validation cohort 1, and 0.766 (95% CI 0.643–0.889), 0.798 (95% CI 0.686–0.910), 0.614 (95% CI 0.462–0.766), and 0.584 (95% CI 0.432–0.736) in the validation cohort 2, respectively. Details of LSM and SSM performance in predicting EV and HRV are presented in Supplementary Table 4. Overall, the SSM demonstrated higher accuracy and AUROC compared with the LSM, PLT, and FIB-4. Furthermore, we investigated the performance of the Baveno VI criteria and the Baveno VII algorithm for ruling out HRV using 2D-SWE (Supplementary Table 5). However, the results were not satisfactory.
We also included the composite FIB-4 score as a single predictor for model development. Supplementary Table 6 shows the diagnostic performance of the FIB-4plus score, which combines the LSM, SSM, and the FIB-4 score as continuous variables for predicting HRV. The XGBoost-FIB-4plus score outperformed the LR-FIB-4plus score, achieving AUROCs of 0.869 (95% CI 0.822–0.916), 0.905 (95% CI 0.830–0.980), and 0.842 (95% CI 0.745–0.938) in the training cohort, validation cohort 1, and validation cohort 2, respectively (Supplementary Fig. 3).
Finally, the XGBoost-FIB-4plus score, which combines LSM, SSM, and the individual components of the FIB-4 score (i.e., age, AST, ALT, and PLT) as separate predictors, demonstrated the highest discrimination ability (Table 2; Fig. 2). An open web page for HRV detection is freely accessible at https://FIB-4plus.shinyapps.io/dynnomapp/.
Performance of the XGBoost-FIB-4plus score for ruling out and ruling in EV and HRV
By applying a rule-out cut-off value of 0.420, the XGBoost-FIB-4plus score achieved sensitivity and NPV greater than 0.95 in both the training cohort and validation cohort 1, outperforming the Baveno VI criteria and the Baveno VII algorithm. Similarly, using a rule-in cut-off value of 0.640, the XGBoost-FIB-4plus score demonstrated specificity and PPV greater than 0.95. These results highlight the superior performance of the XGBoost-FIB-4plus score in both ruling out and ruling in HRV, supporting its potential utility in clinical practice. Detailed results are presented in Supplementary Table 7.
Model interpretation
Figure 3 and Supplementary Figure 4 show the SHAP feature importance and SHAP summary plots for the LRFIB-4plus and XGBoost-FIB-4plus models. This explainable method offers two types of insights: (1) a global explanation, which highlights feature-level contributions to the overall model predictions, and (2) a local explanation, which provides case-specific insights into individual predictions. Features were ranked based on the mean absolute SHAP values, reflecting their relative importance in influencing predictions. A higher SHAP value for a feature indicates a stronger association with the likelihood of EV or HRV. To predict HRV using the XGBoost-FIB-4plus model, the feature importance was ranked as follows: SSM, PLT, LSM, ALT, AST, and age. SSM and PLT consistently emerged as the two most critical predictors across both the LR-FIB-4plus and XGBoost-FIB-4plus models (Fig. 3). This systematic feature ranking reinforces the dominant role of SSM and PLT in prediction models and highlights the interpretability of these models in clinical decision making.
Performance of the FIB-4plus score for predicting EV and HRV in patients with HBV infection
Subgroup analysis was conducted to evaluate the diagnostic performance of the XGBoost-FIB-4plus score specifically in patients with HBV infection. Given the limited number of HBV-infected patients in the validation cohort 2, the analysis focused on the training cohort and validation cohort 1. The XGBoost-FIB-4plus score achieved AUROCs of 0.923 (95% CI 0.887–0.960) and 0.905 (95% CI 0.798– 1.000) in the training cohort and validation cohort 1, respectively (Supplementary Fig. 5; Supplementary Table 8).
Validation of TE-based models in the prospective external validation cohort
The performance of the FIB-4plus score was further assessed in the TE cohort, which utilized TE-derived LSM and SSM values along with the individual components of the FIB-4 score as separate predictors. The XGBoost-FIB-4plus score demonstrated consistently high diagnostic accuracy, achieving an AUROC of 0.904 (95% CI 0.799– 1.000) for predicting HRV, whereas the LR-FIB-4plus score achieved a slightly lower AUROC of 0.831 (95% CI 0.686– 0.977) (Supplementary Fig. 6). In the detection of EV, the XGBoost-FIB-4plus score achieved an AUROC exceeding 0.9 (Supplementary Table 9).
Performance of the XGBoost model combining LSM, SSM, and PLT for predicting EV and HRV
We also developed an alternative model that combined LSM, SSM, and PLT using the XGBoost algorithm to predict EV and HRV. The ROC curves of this model are shown in Supplementary Figure 7, and the importance of these three clinical features, as assessed using the SHAP method, is displayed in Supplementary Figure 8. To predict the HRV, the XGBoost model achieved AUROCs of 0.855 (95% CI 0.808–0.901), 0.891 (95% CI 0.812–0.971), and 0.828 (95% CI 0.726–0.930) for the training, validation cohort 1, and validation cohort 2, respectively (Supplementary Fig. 7). For EV detection, the AUROCs ranged from 0.780 to 0.831. Therefore, the performance of the XGBoost model combining the LSM, SSM, and PLT was lower than that of the FIB-4plus model that we developed earlier.
Evaluation of the FIB-4plus score performance for EV and HRV diagnosis by DCA and calibration curve
The DCA for the XGBoost-FIB-4plus score consistently showed a net benefit in the training cohort and all three validation cohorts across various threshold probabilities. The model outperformed the ‘treat none’ strategy, highlighting its practical utility in decision making (Fig. 4, Supplementary Fig. 9). Furthermore, calibration curves demonstrated good diagnostic performance of the XGBoost-FIB-4plus model across the three validation cohorts (Supplementary Figs. 9, 10).
DISCUSSION
Similar to previous studies [10,26], our current study highlights that a significant proportion (45.8%) of patients who underwent EGD did not have EV. Notably, the presence of EV on EGD serves as one of the surrogate markers of clinically significant portal hypertension [2]. This finding underscores that, for well-compensated cirrhotic patients undergoing screening, EGD may lead to unnecessary costs and potential complications without any added benefits [10,21]. Consequently, there is a growing interest in developing an accurate and non-invasive tool for the detection of EV in compensated cirrhosis, given the invasive and costly nature of the gold standard method.
We first validated the suboptimal performance of the Baveno VI criteria [13] and the Baveno VII algorithm [17] (Baveno VI criteria combined with SSM ≤40 kPa) in ruling out HRV using 2D-SWE. Similarly, our recent study showed that the Baveno VII algorithm was less effective for ruling out clinically significant portal hypertension, with an NPV <90% [20]. To address this, we developed and validated a novel single model, termed FIB-4plus, designed to detect EV and HRV in patients with compensated cirrhosis. The FIB-4plus score is not only highly accurate but is also entirely based on routinely collected data, including elastography examination and laboratory test parameters.
Using a large international multicenter cohort of patients with compensated cirrhosis, we combined LSM, SSM, and the FIB-4 score into the FIB-4plus score for the first time. The FIB-4plus score demonstrated excellent diagnostic performance in detecting HRV in both the training and validation cohorts. It consistently performed well in the subgroup analysis of patients with HBV infection. Impor tantly, the FIB-4plus score showed superior predictive performance for EV and HRV in compensated cirrhosis compared with LSM, SSM, PLT, or FIB-4 alone. Furthermore, the XGBoost-FIB-4plus score, trained using the XGBoost algorithm, achieved an AUROC of 0.927 in the training cohort and 0.919 and 0.902 in the first and second external validation cohorts, respectively. The accuracy values ranged from 0.80 to 0.90, with sensitivity and NPV exceeding 0.95 for ruling out HRV, suggesting that the FIB-4plus score is an accurate and non-invasive tool for HRV in patients with compensated cirrhosis. Further analysis using DCA and calibration curves confirmed the practical applicability of the FIB-4plus score in predicting HRV. Thus, our findings indicate that the FIB-4plus score has the potential to predict EV and HRV in patients with compensated cirrhosis, thereby offering clinicians a valuable tool to guide clinical management and improve patient outcomes.
In this study, we chose to combine LSM, SSM, and FIB-4 values because each parameter measures a different yet complementary aspect of advanced liver disease [27]. In recent years, numerous studies have demonstrated that LSM is a promising non-invasive tool that assesses the degree of liver fibrosis, which is crucial in diagnosing conditions such as cirrhosis and portal hypertension [28,29]. The correlation between LSM and the extent of liver fibrosis makes it valuable in predicting the progression and potential complications of liver diseases [30,31]. At the same time, performing SSM alongside LSM during the same examination is simple and fast. The SSM can reflect changes related to portal hypertension and its severity, irrespective of the etiology[28]. High SSM values are associated with a higher risk of variceal bleeding and other portal hypertension-related complications. A recent metaanalysis has shown that, compared with LSM, the diagnostic performance of SSM was significantly better for detecting the presence of EV [32]. The FIB-4 score, based on blood laboratory parameters, is one of the most widely adopted serum markers, which can be increased in patients with advanced liver disease [31]. However, there is only little data on the combination of two or more noninvasive tests to predict portal hypertension and/or its related complications [15,21,26,27].
Interestingly, a recent study by Vutien et al. [27] introduced the FIB-5 score, a system that combines the FIB-4 score and TE-derived LSM. This research demonstrated that the FIB-5 score offers superior predictive accuracy for assessing the risk of complications (ascites, hepatic encephalopathy, or variceal bleeding) related to portal hypertension in individuals with compensated liver disease (AUROC=0.868), outperforming the predictive capabilities of the FIB-4 score (AUROC=0.672) and LSM (AUROC=0.688) when used independently.
With the advancement of machine learning, clinicians can now transform large volumes of data into practical models, significantly enhancing their ability to diagnose diseases more effectively [25,33] A machine learning model was previously described for the non-invasive detection of EV and HRV in patients with cirrhosis [10]. This approach utilized a machine learning algorithm with routine laboratory parameters, including international normalized ratio, AST, PLT, urea nitrogen, hemoglobin, and the presence of ascites. In the validation cohort, the AUROC ranged from 0.75 to 0.82. This is an example of a noninvasive assessment of HRV using a machine learning algorithm, but its predictive capability was not comparable to that we observed across all three datasets in our study (AUROC ranged from 0.902 to 0.927). Our previous study built a diagnostic model that combined TE-derived LSM, PLT, and total bilirubin based on the light gradient boosting machine algorithm [21], which achieved an AUROC of 0.74 to 0.86 in detecting HRV, and spared more screening EGDs in all cohorts as compared with the Baveno VI criteria.
In this study, we further designed and combined the LSM, SSM, and the FIB-4 score into a single, more accurate score (FIB-4plus), offering clinicians a valuable tool for predicting HRV in patients with compensated cirrhosis. Both LR-FIB-4plus and XGBoost-FIB-4plus scores demonstrated higher specificity and overall classification accuracy for predicting HRV than LSM, SSM, PLT, or FIB-4 alone. One particularly noteworthy finding was that SSM emerged as the most important clinical feature in the prediction, representing a more useful parameter for HRV detection. As our results showed, SSM demonstrated superior accuracy and AUROC in predicting HRV in patients with compensated cirrhosis compared with LSM, PLT, and FIB-4. Researchers have confirmed that SSM is significantly associated with portal hypertension and may be a superior marker for predicting clinical outcomes compared wi th LSM in pat ients wi th cirrhosis [15,16,28,32,34,35]. In contrast, LSM exhibited a poor correlation with the prediction of clinically significant portal hypertension when used for detecting hepatic venous pressure gradient >12 mmHg, likely due to the increasing influence of extra-hepatic factors on the progression of por tal hyper tension [36]. In fact, both the European Association for the Study of the Liver guidelines [37] and the Baveno VII consensus [17] emphasized that SSM plays an important role in the diagnosis of portal hypertension and HRV in patients with compensated advanced chronic liver disease.
Subgroup analyses can provide a more in-depth understanding of the diagnostic performance of the FIB-4plus score in specific patient populations. This study assessed the performance of the FIB-4plus score in a subgroup of patients with HBV infections. Notably, for the XGBoost-FIB-4plus score, an AUROC of over 0.9 was achieved in both the training and validation cohorts. In the training cohort, the FIB-4plus score demonstrated an NPV of 0.865 for ruling out HRV, whereas in the validation cohort, the NPV was 0.946. In addition to the two external validations based on 2D-SWE measurements, we prospectively recruited a cohort of eligible patients with compensated cirrhosis who underwent TE examinations at four Chinese hospitals. We further validated our model for the detection of EV and HRV in these patients. The FIB-4plus score, developed by including the TE-derived LSM and SSM, maintained a good performance in predicting HRV. In the TE cohort, the XGBoost-FIB-4plus score achieved an AUROC of over 0.9. The consistent results across different analyses proved the generalizability and stability of the FIB-4plus score, regardless of the elastographic modality. However, further validation in other ethnic populations is required to determine its broader applicability.
The strengths of this study include its large sample size of patients from two countries, with 502 patients from different institutions included for analysis. This study enrol led a large number of wel l -character ized, compensated patients. Another important strength is that we used robust modeling techniques and different strategies to develop the optimal LR and XGBoost models. Specifically, the XGBoost-FIB-4plus score, which included the individual components of the FIB-4 score (i.e., age, AST, ALT, and PLT) as separate predictors rather than the composite FIB-4 score as a single predictor, demonstrated the highest discrimination. Ultimately, we developed a new and readily available scoring system, termed FIB-4plus, based on routinely collected data, including elastography examinations and laboratory test parameters.
There are also limitations to this study that should be acknowledged. First, the FIB-4plus score was developed using data collected prospectively from Chinese patients and externally validated in the Chinese and Croatian populations. Nevertheless, validation studies involving diverse and representative ethnic groups are necessary to ensure generalizability. Second, the etiology of cirrhosis in our study was unevenly distributed, with virus-infected cirrhosis accounting for the majority of cases. Therefore, we performed a subgroup analysis of patients with HBV infection. Future prospective studies on patients with cirrhosis with other etiologies are warranted to validate our model. Additionally, we did not further investigate the potential factors influencing LSM based on different etiologies. Third, we prospectively recruited a cohort of patients who underwent TE examinations at four Chinese hospitals and validated our model for these patients. However, our cohort did not undergo simultaneous 2D-SWE or TE examinations. Thus, we were unable to compare the performance of the FIB-4plus score (which utilizes LSM and SSM from 2D-SWE) and the Baveno criteria (which uses LSM and SSM from TE) in avoiding endoscopic screening. Nevertheless, we did evaluate the performance of the Baveno VI criteria [13] and the Baveno VII algorithm [17] for ruling out HRV using 2D-SWE. Further studies are required to validate the FIB-4plus score for avoiding endoscopic screening in comparison with the Baveno criteria. Fourth, we did not explore the impact of antiviral therapy and viral suppression status on the efficacy of this score. Further research is needed in the future to investigate the effect of antiviral treatment and viral suppression status on the outcomes studied. Finally, we did not evaluate the prognostic performance of the FIB-4plus score in predicting long-term clinical outcomes. such as decompensation (e.g., ascites, HRV bleeding, and hepatic encephalopathy) or mortality. Nevertheless, we planned to validate the prognostic relevance of the FIB-4plus score in our prospective observational cohort of patients with compensated cirrhosis in terms of predicting complications related to portal hypertension.
In conclusion, we combined the LSM, SSM, and the FIB-4 score into a single scoring system (FIB-4plus) that could predict HRV in patients with compensated cirrhosis. This novel scoring system is an effective tool for clinicians and holds great promise for guiding clinical management and improving patient outcomes.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Supplement1
Print
