Noninvasive prediction of esophagogastric varices in hepatitis B: An extreme gradient boosting model based on ultrasound and serology

Abstract

BACKGROUND

Severe esophagogastric varices (EGVs) significantly affect prognosis of patients with hepatitis B because of the risk of life-threatening hemorrhage. Endoscopy is the gold standard for EGV detection but it is invasive, costly and carries risks. Noninvasive predictive models using ultrasound and serological markers are essential for identifying high-risk patients and optimizing endoscopy utilization. Machine learning (ML) offers a powerful approach to analyze complex clinical data and improve predictive accuracy. This study hypothesized that ML models, utilizing noninvasive ultrasound and serological markers, can accurately predict the risk of EGVs in hepatitis B patients, thereby improving clinical decision-making.

AIM

To construct and validate a noninvasive predictive model using ML for EGVs in hepatitis B patients.

METHODS

We retrospectively collected ultrasound and serological data from 310 eligible cases, randomly dividing them into training (80%) and validation (20%) groups. Eleven ML algorithms were used to build predictive models. The performance of the models was evaluated using the area under the curve and decision curve analysis. The best-performing model was further analyzed using SHapley Additive exPlanation to interpret feature importance.

RESULTS

Among the 310 patients, 124 were identified as high-risk for EGVs. The extreme gradient boosting model demonstrated the best performance, achieving an area under the curve of 0.96 in the validation set. The model also exhibited high sensitivity (78%), specificity (94%), positive predictive value (84%), negative predictive value (88%), F1 score (83%), and overall accuracy (86%). The top four predictive variables were albumin, prothrombin time, portal vein flow velocity and spleen stiffness. A web-based version of the model was developed for clinical use, providing real-time predictions for high-risk patients.

CONCLUSION

We identified an efficient noninvasive predictive model using extreme gradient boosting for EGVs among hepatitis B patients. The model, presented as a web application, has potential for screening high-risk EGV patients and can aid clinicians in optimizing the use of endoscopy.

Key Words: Esophagogastric varices; Machine learning; Extreme gradient boosting; Ultrasound; Serological markers

Core Tip: We constructed a noninvasive predictive model using machine learning for esophagogastric varices in hepatitis B patients. An extreme gradient boosting model, based on ultrasound and serological markers, achieved high accuracy (area under the curve = 0.96) in predicting high-risk esophagogastric varices. Key predictive variables included albumin, prothrombin time, portal vein flow velocity and spleen stiffness. A web-based application was developed to facilitate clinical use, offering real-time risk assessment. This model provides a promising tool for targeted screening, potentially reducing the need for costly and risky endoscopic procedures in low-risk individuals.

INTRODUCTION

Hepatitis B virus infection affects around 296 million individuals globally[1]. Hsu et al[2] projected a 39% increase in the global annual mortality from hepatitis B between 2015 and 2030. One of the severe complications associated with chronic hepatitis B is the development of esophagogastric varices (EGVs)[3]. These dilated submucosal veins in the esophagus and stomach are a major cause of morbidity and mortality due to the risk of life-threatening hemorrhage[3]. The gold standard for detecting EGVs is endoscopy, which, despite its high sensitivity and specificity, comes with several drawbacks. Endoscopy is not only costly and uncomfortable for patients but also carries inherent risks such as bleeding and infection[4]. Studies indicate that significant bleeding can occur in 1%-2% of endoscopic procedures[5]. These limitations highlight the urgent need for a noninvasive method to identify patients at high risk for EGVs, thus optimizing the use of endoscopy and minimizing unnecessary procedures in low-risk individuals. Recent advances in noninvasive diagnostic techniques suggest that combining ultrasound and serological markers can provide a reliable alternative for predicting severe EGVs[6,7]. Parameters such as spleen stiffness, portal vein (pv) flow velocity and serological markers such as prothrombin time and platelet count have shown potential in previous studies[8]. These noninvasive markers could be integrated into a predictive model to accurately identify high-risk patients.

Machine learning (ML) has emerged as a particularly powerful tool for analyzing complex multidimensional data such as the combination of ultrasound imaging and serological markers. Traditional statistical methods, such as logistic regression, often struggle when dealing with large, complex datasets where interactions between variables are not linear or straightforward. ML algorithms, however, excel at capturing intricate patterns and relationships within data, making them more suitable for identifying subtle, nonlinear associations between clinical variables[9,10]. Unlike traditional methods, which typically rely on predefined relationships, ML models can autonomously uncover new insights from data and are better at handling multicollinearity and other forms of variable interdependence[11,12]. For example, in the context of predicting EGVs, ML can evaluate multiple input factors simultaneously, such as spleen stiffness, pv flow velocity and serological markers, without requiring the simplification or assumptions that traditional methods impose. ML algorithms such as extreme gradient boosting (XGBoost) and random forest (RF) can provide feature importance rankings, aiding in the interpretability of the model by identifying the most critical predictors of outcomes. This makes it possible to build more accurate and clinically applicable models. The use of ML in this study was therefore critical, as it enabled the processing of large amounts of data while improving the precision of predictions over traditional statistical methods.

This study retrospectively analyzed data from 310 hepatitis B patients who underwent endoscopy, including their ultrasound and serological parameters. The aim was to construct a high-accuracy noninvasive predictive model for EGVs by comparing the performance of 11 ML algorithms. By leveraging the strengths of ML, this study aimed to contribute to clinical practice by providing an effective tool for the targeted screening of high-risk patients.

MATERIALS AND METHODS

Study design and protocol

This was a retrospective, observational analysis aimed at developing and validating a noninvasive predictive model for EGVs in patients with hepatitis B. By utilizing historical patient data, we aimed to leverage ML techniques to identify high-risk individuals who would benefit from endoscopic screening. The study was conducted at Mengchao Hepatobiliary Hospital. The data were collected from January 2016 to December 2023.

This study was performed in accordance with the ethical standards of the institutional and national research committees and with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Prior to data collection, approval was obtained from the Institutional Review Board of Mengchao Hepatobiliary Hospital, approval No. 2022_028_01, ensuring all patient information was handled with the utmost confidentiality and in compliance with regulatory standards. all data were anonymized to protect privacy. All patients provided informed consent.

Patient population

Inclusion criteria were as follows (Figure 1): (1) Diagnosis of chronic hepatitis B, confirmed by serological markers (hepatitis B surface antigen positive for > 6 months); (2) Endoscopic examination to assess the presence of EGVs; (3) Ultrasound and serological data within 3 months of the endoscopic examination; and (4) Age ≥ 18 years. Exclusion criteria were: (1) Other causes of liver disease (e.g., hepatitis C, alcoholic liver disease, or autoimmune hepatitis); (2) Prior history of treatment for EGVs (e.g., banding or sclerotherapy); (3) Incomplete medical records or missing key data points required for the analysis; and (4) Co-infection with human immunodeficiency virus or other significant comorbidities that could affect liver function and varices formation.

Figure 1 Patient inclusion flow. EGVs: Esophagogastric varices; HIV: Human immunodeficiency virus; HBsAg: Hepatitis B virus surface antigen.

Data collection

The data were obtained from the electronic medical records of patients treated at Mengchao Hepatobiliary Hospital, such as patient demographics (age and gender); medical history (duration of hepatitis B or previous liver-related complications); and clinical signs and symptoms (jaundice, ascites or hepatic encephalopathy).

Ultrasound examination

Ultrasound examinations were performed using high-resolution ultrasound machines (Siemens Sequia, 5C-1). The key parameters measured included liver and spleen stiffness (using 2D shear wave elastography to assess tissue stiffness) (Figure 2). Pv flow velocity was measured using Doppler ultrasound. Spleen long diameter and thickness were measured in the coronal plane, and the presence of collateral branches and ascites was assessed using B-mode ultrasound. To ensure optimal imaging quality, all patients fasted for at least 8 hours prior to the examination. Patients were positioned either supine or in the left lateral decubitus position, depending on the parameter being measured.

Figure 2 Esophagogastric varices of varying severity and the corresponding spleen stiffness measurements. A: A 50-year-old male patient with mild esophagogastric varices (EGVs) as determined by endoscopy, with a median spleen stiffness value of 11.6 kPa; B: A 33-year-old male patient with moderate EGVs as determined by endoscopy, with a median spleen stiffness value of 19.5 kPa; C: A 59-year-old female patient with severe EGVs as determined by endoscopy, with a median spleen stiffness value of 25.4 kPa.

Liver and spleen stiffness were measured with minimal probe pressure to avoid artifacts. Specifically, measurements were taken from all eight segments of the liver, with five regions of interest in each segment. The stiffness values from these five regions of interest in each segment were averaged to obtain the final measurement for that segment. This method ensured a comprehensive and consistent assessment of liver stiffness across all regions. Pv flow velocity was measured: The sample box was adjusted to 3 mm and placed at the pv 1 cm from the hepatic hilum. By adjusting the patient’s position, the angle was maintained at no more than 45; measurements were repeated three times; and the maximum pv flow velocity was recorded, with the average flow velocity being used as the subject of the study. A similar approach was applied for measuring the spleen vein flow velocity. Ascites was assessed primarily in the hepatorenal recess and pelvis. To minimize measurement errors, all ultrasound examinations were performed by experienced sonographers, who were blinded to clinical data. Quality control was enhanced by reviewing all measurements by a second operator. Discrepancies in measurements, such as extreme outliers or inconsistent values, were addressed by consulting with the multidisciplinary team to determine whether the measurement should be excluded or re-evaluated. In cases of suspected inappropriate measurements or unreliable data, patients were excluded from the final analysis. All findings were promptly recorded in the electronic medical records system, and any discrepancies were resolved through consensus during multidisciplinary meetings.

Serological assessment

Blood samples were collected within 1 week of the endoscopic examination, typically after an 8-hour fast, drawn in the morning to reduce diurnal variability, and processed within 2 hour to ensure accuracy. The following parameters were recorded: Prothrombin time, alanine aminotransferase, aspartate aminotransferase, albumin, total bilirubin, direct bilirubin, indirect bilirubin, creatinine, alkaline phosphatase, red blood cell count, white blood cell count, and international normalized ratio for blood coagulation.

Endoscopic evaluation

Endoscopic evaluation was performed using standard esophagogastroduodenoscopy procedures, which allowed direct visualization of the esophagus, stomach and upper small intestine using a flexible endoscope with a camera and light source. Patients fasted for > 8 hours before the procedure for better visualization, and pre-procedural assessments were conducted to identify any risks. Moderate sedation with drugs such as midazolam and fentanyl were administered, and vital signs were continuously monitored. The endoscope was gently inserted through the mouth to inspect the mucosal lining of the esophagus, stomach and duodenum for varices, erosions, ulcers and other abnormalities, focusing on grading the varices by size and location. Severe EGVs, named high-risk EGVs, were diagnosed based on size (grade 2: 5-10 mm, grade 3: > 10 mm) and the presence of red signs or bleeding indicators. The procedure was performed by experienced gastroenterologists, assisted by endoscopy technicians and specialized nursing staff, who provided patient care, handled equipment, and ensured all instruments were available and properly disinfected. The findings were documented by the gastroenterologists, who also made clinical decisions based on the observations. The different grades of EGVs are shown in Figure 2.

ML model development

We used a comprehensive suite of ML algorithms to construct and validate the predictive model for EGVs in hepatitis B patients. The following algorithms were utilized: RF, adaptive boosting, artificial neural network, decision tree, extra trees, gradient boosting machine, k-nearest neighbors, light gradient boosting machine, logistic regression, support vector machine, and XGBoost. Prior to model training, essential feature engineering techniques were applied to the dataset. Continuous variables were normalized to ensure uniformity, and Min-Max scaling was performed to adjust the range of features between 0 and 1. Missing data were handled through multiple imputation by chained equations for continuous variables. For categorical variables, imputation was performed using the most frequent category (mode) or logistic regression imputation, depending on the nature of the variable. This ensured a complete and consistent dataset for model training and validation.

Feature selection and model explanation

Interpreting ML models can be challenging because of their inherent complexity. In our study, we utilized SHapley Additive exPlanation (SHAP) to address this “black box” issue by ranking the importance of input features and explaining the results of the predictive model. Initially, univariate analysis was conducted within the modeling group to identify statistically significant factors and those potentially predictive of EGVs. A P threshold of 0.05 was used to determine statistical significance, and variables meeting this criterion were considered for inclusion in the subsequent multivariate analysis. These factors were incorporated into the ML modeling strategy. We evaluated the performance of 11 different ML models by plotting their receiver operating characteristic (ROC) and decision curve analysis (DCA) curves. The area under the ROC curve quantified the ability of the model to distinguish between high-risk and low-risk patients across all classification thresholds, which is particularly suited for clinical predictive tasks. While area under the curve (AUC) evaluated diagnostic accuracy, DCA assessed clinical utility by quantifying the net benefit of using the model to guide interventions across threshold probabilities. It explicitly weighed the harms of unnecessary treatments (false positives) against the benefits of timely interventions (true positives), thereby aligning statistical performance with real-world clinical decision-making. The integration of both metrics ensured that our models were statistically sound and clinically actionable.

The top five models based on AUC were further analyzed using bee swarm plots. Considering both ROC and DCA results, XGBoost was selected as the final model for our study. Dependency plots, overall force plots and decision plots were generated to illustrate the influence of various factors on the predictions of the model.

SHAP values assisted in feature selection by identifying the most critical predictors. This process reduced the number of features from the initial set to the top four most important, which were used to construct a web-based calculator (https://pectgew2rqefrdqyjgqcrh.streamlit.app/). By inputting specific values, clinicians can determine a patient's risk of severe EGVs. This tool enhanced the practical application of our model in clinical settings, providing a user-friendly interface for healthcare providers. The SHAP method provided both global and local explanations for the model. Global explanations offered consistent attribution values for each feature, showing their associations with the risk of severe EGVs, while local explanations demonstrated specific predictions for individual patients based on their data. This approach ensured that our predictive model was accurate and interpretable and reliable for clinical use.

Statistical analysis

The statistical analysis and ML model development were conducted using several robust analytical tools, including Python with libraries such as Scikit-learn, Pandas, NumPy and Matplotlib, as well as R for DCA. Descriptive statistics, including mean ± SD, were calculated for continuous variables to summarize central tendencies and variability, while frequency distributions and percentages described categorical variables.

Inferential statistical methods were applied to compare the two groups (patients with and without severe EGVs). t-tests were used to compare the means of continuous variables, χ² tests assessed associations between categorical variables, and Mann-Whitney U tests were used for nonparametric comparisons when the data did not follow a normal distribution. Pearson and Spearman correlation coefficients were calculated to examine relationships between continuous variables and identify potential collinearities. The performance of the ML models was evaluated using several key metrics to assess their accuracy and clinical utility. AUC measured the ability of the model to distinguish between patients with and without severe EGVs, sensitivity (recall) measured the proportion of actual positive cases correctly identified, and specificity measured the proportion of actual negative cases correctly identified. Positive predictive value and negative predictive value indicated the proportions of true positives and true negatives, respectively. The F1 score, the harmonic mean of precision and recall, balanced false positives and false negatives, while accuracy measured the proportion of correctly classified instances.

To ensure robustness, five-fold cross-validation was used. The training set was divided into five subsets, with the model trained on four subsets and validated on the remaining one. This process was repeated five times, with each subset used once as the validation data. Stratified cross-validation maintained the proportion of severe EGV cases in each fold, preserving the original class distribution of the dataset. This combination of statistical methods and model evaluation techniques provided a comprehensive analysis, ensuring the development of a robust and accurate predictive model for EGVs in hepatitis B patients.

RESULTS

Basic characteristics of the patients

The general characteristics of all patients are shown in Table 1. A total of 310 patients were included in the study, with a mean age of 53.4 ± 12.3 years. Among them, 124 patients were identified as high risk for EGV. The training cohort included 248 patients, with 99 identified as high risk. The validation cohort included 62 patients, with 25 identified as high risk. The baseline characteristics of the training cohort and the validation cohort were matched and comparable.

Table 1 Basic characteristics, mean ± SD.

Index	Total	Train set	Test set	P value
Liver stiffness	13.2 ± 5.0	13.2 ± 5.3	12.9 ± 4.0	0.91
Spv speed	35.3 ± 9.1	35.5 ± 9.3	34.5 ± 8.6	0.4
Platelet count	132.9 ± 63.8	131.3 ± 64.7	139.4 ± 60.2	0.23
Model for end-stage liver disease	30.2 ± 1.9	30.3 ± 1.8	29.9 ± 2.1	0.044
Creatinine	75.5 ± 27.1	76.2 ± 29.6	72.5 ± 12.9	0.74
Alanine aminotransferase	38.0 ± 39.0	36.4 ± 34.6	44.5 ± 52.8	0.046
Aspartate aminotransferase	47.8 ± 62.1	44.7 ± 53.3	60.2 ± 88.4	0.057
Total bilirubin	30.5 ± 34.5	30.7 ± 34.2	29.6 ± 36.0	0.43
Direct bilirubin	16.2 ± 27.8	16.4 ± 27.6	15.5 ± 29.0	0.48
Indirect bilirubin	14.2 ± 9.8	14.3 ± 8.9	14.0 ± 13.0	0.27
Alkaline phosphatase	108.8 ± 55.9	105.2 ± 53.1	123.0 ± 64.4	0.027
Red blood cell count	4.3 ± 0.8	4.3 ± 0.8	4.4 ± 0.7	0.35
White blood cell count	5.3 ± 1.9	5.3 ± 1.9	5.3 ± 1.6	0.64
Age	53.4 ± 12.3	53.6 ± 12.4	52.5 ± 11.8	0.57
Spleen stiffness	15.1 ± 4.8	15.2 ± 4.9	14.8 ± 4.7	0.83
pv	1.2 ± 0.2	1.2 ± 0.2	1.2 ± 0.2	0.59
pvspeed	27.5 ± 6.7	27.2 ± 6.6	28.5 ± 7.2	0.14
splong	13.3 ± 2.6	13.3 ± 2.5	13.4 ± 2.7	0.88
spwide	4.6 ± 1.0	4.7 ± 0.9	4.6 ± 1.0	0.34
spv	0.8 ± 0.2	0.8 ± 0.2	0.9 ± 0.2	0.41
Prothrombin time	15.5 ± 3.2	15.7 ± 3.4	14.8 ± 2.3	0.045
Albumin	39.7 ± 8.9	40.0 ± 8.9	38.6 ± 8.8	0.4
International normalized ratio	1.2 ± 0.3	1.3 ± 0.3	1.2 ± 0.3	0.22
Sex, n (%)
Female	84 (27.1)	67 (27.0)	17 (27.4)	1
Male	226 (72.9)	181 (73.0)	45 (72.6)	-
Child-Pugh class, n (%)
1	168 (54.2)	131 (52.8)	37 (59.7)	0.29
2	119 (38.4)	100 (40.3)	19 (30.6)	-
3	23 (7.4)	17 (6.9)	6 (9.7)	-
Collateral, n (%)
No	244 (78.7)	196 (79.0)	48 (77.4)	0.86
Yes	66 (21.3)	52 (21.0)	14 (22.6)	-
Severe EGV, n (%)
No	186 (60.0)	149 (60.1)	37 (59.7)	1
Yes	124 (40.0)	99 (39.9)	25 (40.3)	-

Spv speed: Spleen vein speed; pv: Portal vein; pvspeed: Portal vein speed; splong: Spleen length; spwide: Spleen width; spv: Splenic vein; EGV: Esophagogastric varices.

Basic characteristics of the training set

The results of the modeling cohort are shown in Table 2. There were 99 patients in the high-risk group. The following factors showed significant differences between the high- and low-risk groups: Sex, portal vein speed (pvspeed), platelet count, creatinine, total bilirubin, direct bilirubin, indirect bilirubin, spleen stiffness, pvspeed, spleen width (spwide), splenic vein (spv), albumin, prothrombin time and Child-Pugh score. Through collinearity analysis, and included potentially significant predictors for high-risk prediction in subsequent analyses. These included: Age, spleen stiffness, pv, pvspeed, spleen length, spv, prothrombin time, albumin, Child-Pugh score, sex, presence of collateral vessels (collateral), and international normalized ratio.

Table 2 Data in training cohort, mean ± SD.

Index	Total	Low risk	High risk	P value
Liver stiffness	13.2 ± 5.3	13.2 ± 5.9	13.3 ± 4.2	0.33
Spv speed	35.5 ± 9.3	36.4 ± 9.2	34.1 ± 9.3	0.026
Platelet count	131.3 ± 64.7	128.0 ± 73.2	136.1 ± 49.2	0.048
Model for end-stage liver disease	30.3 ± 1.8	30.5 ± 2.0	30.0 ± 1.5	0.25
Creatinine	76.2 ± 29.6	76.9 ± 36.8	75.2 ± 12.8	0.018
Alanine aminotransferase	36.4 ± 34.6	40.0 ± 42.8	30.8 ± 14.1	0.88
Aspartate aminotransferase	44.7 ± 53.3	43.8 ± 49.7	46.1 ± 58.5	0.25
Total bilirubin	30.7 ± 34.2	28.3 ± 37.1	34.3 ± 29.2	< 0.001
Direct bilirubin	16.4 ± 27.6	14.5 ± 29.4	19.1 ± 24.5	< 0.001
Indirect bilirubin	14.3 ± 8.9	13.8 ± 9.8	14.9 ± 7.4	0.042
Alkaline phosphatase	105.2 ± 53.1	101.3 ± 50.0	111.1 ± 57.3	0.37
Red blood cell count	4.3 ± 0.8	4.3 ± 0.9	4.2 ± 0.7	0.82
White blood cell count	5.3 ± 1.9	5.4 ± 2.1	5.3 ± 1.7	0.67
Age	53.6 ± 12.4	52.3 ± 13.7	55.6 ± 10.0	0.025
Spleen stiffness	15.2 ± 4.9	13.7 ± 4.5	17.3 ± 4.7	< 0.001
pv	1.2 ± 0.2	1.3 ± 0.2	1.2 ± 0.1	0.16
pvspeed	27.2 ± 6.6	28.0 ± 6.8	26.0 ± 6.1	0.03
splong	13.3 ± 2.5	13.6 ± 2.8	12.9 ± 2.0	0.11
spwide	4.7 ± 0.9	4.5 ± 0.9	4.8 ± 0.9	0.012
spv	0.8 ± 0.2	0.9 ± 0.2	0.8 ± 0.2	0.016
Prothrombin time	15.7 ± 3.4	15.5 ± 3.2	17.8 ± 3.6	0.032
Albumin	40.0 ± 8.9	40.7 ± 8.9	25.13 ± 6.5	0.015
International normalized ratio	1.3 ± 0.3	1.2 ± 0.3	1.3 ± 0.4	0.25
Sex, n (%)
Female	67 (27.0)	26 (17.4)	41 (41.4)	< 0.001
Male	181 (73.0)	123 (82.6)	58 (58.6)	-
Child–Pugh class, n (%)
1	131 (52.8)	71 (47.7)	60 (60.6)	0.025
2	100 (40.3)	70 (47.0)	30 (30.3)	-
3	17 (6.9)	8 (5.4)	9 (9.1)	-
Collateral, n (%)
No	196 (79.0)	120 (80.5)	76 (76.8)	0.53
Yes	52 (21.0)	29 (19.5)	23 (23.2)	-

Spv speed: Spleen vein speed; pv: Portal vein; pvspeed: Portal vein speed; splong: Spleen length; spwide: Spleen width; spv: Splenic vein; EGV: Esophagogastric varices.

Model construction and validation

We used 11 ML methods for model construction and validation. Initially, we included all 13 selected factors and plotted the ROC curve DCA curves for all models in the validation cohort (Figure 3A and B). These models had the largest AUC: Extra tree (0.97), RF (0.97), XGBoost (0.96), light gradient boosting machine (0.96) and adaptive boosting (0.94). Sensitivity, specificity and other metrics for all models were calculated and presented in Table 3. To explore the impact of different numbers of variables on the overall AUC, we used recursive feature elimination to evaluate and rank feature importance. We analyzed how the AUC varied with different numbers of features for each model and plotted the corresponding AUCs (Figure 3C). Through comparison, we found that incorporating the top four ranked variables was sufficient for the model to achieve a high AUC. Based on a comprehensive consideration of AUC and DCA, XGBoost was selected as the primary model for further study.

Figure 3 Curves for various prediction metrics across all models. A: Receiver operating characteristic curves; B: Decision curve analysis curves; C: Area under the curve values corresponding to different numbers of features. AUC: Area under the curve; RF: Random forest; AdaBoost: Adaptive boosting; ANN: Artificial neural network; DT: Decision tree; ET: Extra trees; GBM: Gradient boosting machine; KNN: K-nearest neighbors; LightGBM: Light gradient boosting machine; LR: Logistic regression; SVM: Support vector machine; XGB: Extreme gradient boosting.

Table 3 Statistical measures for all models.

Model name	Area under curve	Accuracy	Precision	Recall	Specificity	F1 score	Positive predict value	Positive predict value
Random forest	0.97 (0.94-1.00)	0.93 (0.87-0.98)	0.92 (0.85-1.00)	0.87 (0.72-1.00)	0.95 (0.92-1.00)	0.88 (0.80-0.98)	0.92 (0.85-1.00)	0.91 (0.84-1.00)
AdaBoost	0.94 (0.88-0.99)	0.84 (0.76-0.94)	0.85 (0.69-1.00)	0.71 (0.53-0.90)	0.91 (0.84-1.00)	0.76 (0.62-0.91)	0.88 (0.69-1.00)	0.85 (0.74-0.95)
Artificial neural network	0.77 (0.61-0.85)	0.63 (0.52-0.74)	0.51 (0.30-0.70)	0.49 (0.31-0.73)	0.69 (0.55-0.83)	0.47 (0.32-0.67)	0.46 (0.30-0.70)	0.73 (0.56-0.85)
Decision tree	0.79 (0.67-0.89)	0.81 (0.69-0.90)	0.74 (0.56-0.94)	0.69 (0.50-0.86)	0.85 (0.75-0.97)	0.71 (0.55-0.86)	0.74 (0.56-0.94)	0.84 (0.71-0.93)
Extra tree	0.97 (0.95-1.00)	0.94 (0.87-0.98)	0.91 (0.83-1.00)	0.85 (0.71-1.00)	0.91 (0.91-1.00)	0.88 (0.80-0.98)	0.93 (0.83-1.00)	0.92 (0.84-1.00)
Gradient boosting machine	0.92 (0.84-0.98)	0.86 (0.75-0.94)	0.83 (0.67-1.00)	0.72 (0.55-0.92)	0.91 (0.83-1.00)	0.76 (0.64-0.91)	0.84 (0.67-1.00)	0.84 (0.75-0.96)
K-nearest neighbors	0.89 (0.80-0.96)	0.82 (0.71-0.90)	0.73 (0.54-0.90)	0.72 (0.55-0.91)	0.83 (0.72-0.95)	0.68 (0.57-0.86)	0.73 (0.54-0.90)	0.85 (0.72-0.95)
Lightgbm	0.96 (0.91-0.99)	0.86 (0.79-0.95)	0.81 (0.73-1.00)	0.74 (0.54-0.92)	0.94 (0.87-1.00)	0.79 (0.65-0.92)	0.85 (0.74-1.00)	0.86 (0.76-0.95)
Logistic regression	0.73 (0.60-0.85)	0.67 (0.55-0.77)	0.54 (0.33-0.74)	0.61 (0.41-0.82)	0.68 (0.53-0.83)	0.58 (0.38-0.73)	0.58 (0.33-0.74)	0.76 (0.61-0.89)
Support vector machine	0.74 (0.62-0.86)	0.61 (0.50-0.73)	0.47 (0.23-0.72)	0.34 (0.15-0.56)	0.77 (0.64-0.90)	0.39 (0.19-0.58)	0.47 (0.23-0.72)	0.63 (0.52-0.80)
Extreme gradient boosting	0.96 (0.92-0.99)	0.86 (0.81-0.97)	0.87 (0.74-1.00)	0.78 (0.62-0.95)	0.94 (0.86-1.00)	0.83 (0.70-0.94)	0.84 (0.74-1.00)	0.88 (0.78-0.98)

AdaBoost: Adaptive boosting.

We generated bee swarm plots for the top five models based on AUCs (Figure 4). Beeswarm plots provided a visual representation of feature importance and the direction of the effect of each feature on the model output. In these plots, each point represented a patient, and it was positioned along the x-axis according to the SHAP value of the corresponding feature for that patient. Features were ranked in descending order of importance from top to bottom. For example, in the bee swarm plot for the XGBoost model (Figure 4D), we observed the impact of different features on the predicted probability of EGVs. A positive SHAP value for a feature indicated that the feature contributed to increasing the predicted probability of EGVs, while a negative SHAP value indicated decrease. Specifically, for XGBoost, the plot suggested that lower album in levels, higher prothrombin time (PT) values, lower pvspeed, and higher spleen stiffness were associated with an increased EGV risk.

Figure 4 Overall bee swarm plot for the top five models ranked by area under curve value. A: Random forest; B: Extra tree; C: Light gradient boosting machine; D: Extreme gradient boost; E: Adaptive boosting. SHAP: SHapley Additive exPlanation; pvspeed: Portal vein speed; Alb: Albumin; PT: Prothrombin time; pv: Portal vein; spwide: Spleen width; INR: International normalized ratio; splong: Spleen length.

To enhance interpretability, SHAP force plots were generated to illustrate the impact of individual features on the predictions of the model. Figure 5A displays force plots for all patients, providing a comprehensive view of the contribution of each feature to the model output. These plots demonstrated how features such as pvspeed, spwide, spleen stiffness and PT influenced the risk predictions for EGVs. Red segments indicate factors that increased the predictive score (higher risk), while blue segments represent factors that decreased it (lower risk). Figure 5B and C highlights the application of SHAP force plots in specific cases, with separate plots for accurately and inaccurately predicted patients. Figure 5B focuses on incorrectly predicted cases, emphasizing features like spleen length, albumin and PT, which were associated with deviations from the correct outcome. Figure 5C depicts correctly predicted cases, showing the contributions of features such as spleen stiffness, age and spwide, and demonstrating how these factors collectively led to accurate risk assessments. To further analyze the predictions, Figure 6 presents waterfall plots and decision plots for accurately and inaccurately predicted patients. In the waterfall plots, the individual contribution of each feature is displayed, illustrating how cumulative effects led to either corrector incorrect predictions. For correctly predicted patients (Figure 6A), features such as pvspeed, spleen stiffness and spwide strongly influenced the final risk score. Conversely, in inaccurately predicted cases (Figure 6B), misalignment in the contributions of these features resulted in predictive errors. The decision plots (Figure 6C and D) offer a cumulative perspective, showing how the inclusion of each feature incrementally affected the prediction. In accurately predicted patients (Figure 6C), the steady accumulation of significant features contributed to precise risk assessment. However, for inaccurately predicted cases (Figure 6D), the combined effects of the features failed to align correctly with the true outcome, highlighting areas where the model could be refined. These visualizations emphasized the utility of SHAP analysis in providing both global and local interpretability for the predictive model, allowing for a detailed understanding of how individual features interact to influence outcomes. Moreover, they highlight specific cases where feature contributions may require further investigation to improve model performance. Through recursive feature elimination ranking, we determined the importance of all variables and plotted curves for sensitivity, specificity and other metrics corresponding to different numbers of variables (Figure 7). Other results are shown in Supplementary Figure 1.

Figure 5 SHapley Additive exPlanation force plots demonstrating model predictions for actual patients. A: The force plot for all patients illustrates the SHapley Additive exPlanation (SHAP) values, showing how each feature contributes to the predictions of the model for each individual. The x-axis represents the model output value, with red indicating an increase in the prediction score (higher risk) and blue indicating a decrease (lower risk). This comprehensive view demonstrates the impact of features like portal vein speed, spleen width and spleen stiffness on the predictions for all patients; B: The force plot application for actual patients further interprets model predictions. The force plot for inaccurately predicted patients displays the SHAP values for a case where the model prediction was incorrect. Features such as spleen length, prothrombin time and albumin are highlighted, showing their contributions to the incorrect prediction. Red bars indicate features that increased the prediction score, leading to a higher risk, while blue bars indicate features that decreased the score, indicating a lower risk; C: The force plot for accurately predicted patients shows the SHAP values for a correctly predicted case. Significant features like spleen stiffness, age and spleen width are shown with their respective contributions, demonstrating how these features combined to provide an accurate risk assessment, with red bars increasing the prediction score and blue bars decreasing it. splong: Spleen length; PT: Prothrombin time; Alb: Albumin; pvspeed: Portal vein speed; pv: Portal vein; spwide: Spleen width; INR: International normalized ratio.

Figure 6 Waterfall and decision plots demonstrating model predictions for actual patients. A: Waterfall plot for accurately predicted patients shows the contribution of each feature to the final prediction. Features such as portal vein speed, spleen width and spleen stiffness were key contributors. Blue bars decrease the predictive score (indicating lower risk), and red bars increase the predictive score (indicating higher risk). This plot demonstrates how the model effectively used these features to make correct predictions; B: Waterfall plot for inaccurately predicted patients displays the contributions for cases where the predictions of the model were incorrect. The same features were considered, but their combined contributions led to an incorrect prediction. This plot helped identify which features might have been misinterpreted by the model; C: Decision plot for accurately predicted patients shows how the predictive score of the model changed as each feature was considered. The cumulative impact of features such as portal vein speed, spleen width and spleen stiffness led to a correct prediction. Each point on the line represents the contribution of an additional feature to the final score; D: Decision plot for inaccurately predicted patients illustrates the cumulative effect of each feature for patients with incorrect predictions. Despite the inclusion of significant features, the cumulative contributions did not lead to an accurate prediction. This helped identify potential areas for model improvement. spwide: Spleen width; PT: Prothrombin time; pv: Portal vein; spv: Splenic vein; splong: Spleen length; Alb: Albumin; pvspeed: Portal vein speed; INR: International normalized ratio.

Figure 7 The changes in various predictive metrics for the extreme gradient boosting model when using different numbers of variables. It was observed that the model achieved an optimal area under the curve with just four variables. XGB: Extreme gradient boosting; AUC: Area under the curve; PPV: Positive predictive value; NPV: Negative predictive value.

The top four variables identified were spleen stiffness, pvspeed, PT and albumin. Dependency plots for these factors illustrate the relationships between these features and their respective SHAP values, providing insights into their impact on the predictions (Figure 8). Spleen stiffness demonstrated a strong positive correlation with SHAP values, indicating that higher spleen stiffness contributed significantly to the risk prediction for EGVs (Figure 8A). Figure 8B shows an inverse relationship between pvspeed and SHAP values, where lower speeds were associated with a higher risk score, reflecting the role of portal hemodynamics in the development of varices. Similarly, Figure 8C highlights a positive correlation between PT and SHAP values, suggesting that prolonged PT - an indicator of impaired liver function - elevated the risk prediction. Figure 8D depicts a negative correlation between albumin levels and SHAP values, with lower albumin levels, indicative of reduced liver synthetic function, contributing to higher risk scores. These plots collectively emphasize the critical role of these variables in influencing the predictions, offering a detailed understanding of their contributions to the identification of high-risk patients. Additional dependency plots for the other nine variables are shown in Supplementary Figure 2. Interaction plots for continuous variables are depicted in Supplementary Figure 3. SHAP heatmap is depicted in Supplementary Figure 4.

Figure 8 SHapley Additive exPlanation dependency plots for key predictive features. This figure illustrates the SHapley Additive exPlanation (SHAP) dependency plots for key predictive features, showing how each feature influenced the predictions of the model for esophagogastric varices (EGVs). A: Relationship between spleen stiffness and its SHAP value, revealing a positive correlation where higher spleen stiffness increased the predictive score for EGVs; B: Relationship between portal vein speed and its SHAP value, indicating a negative correlation, with higher speed decreasing the predictive score; C: Relationship between prothrombin time and its SHAP value, where higher PT values were associated with higher SHAP values, suggesting an increased predictive score for EGV; D: Relationship between albumin levels and their SHAP values, showing a negative correlation where higher albumin levels decrease the predictive score. These SHAP dependency plots provide a comprehensive visualization of how individual features affected the predictions of the extreme gradient boosting model, highlighting the most influential factors in predicting high-risk EGVs and aiding in understanding the decision-making process of the model. pvspeed: Portal vein speed; PT: Prothrombin time; Alb: Albumin.

Based on the selected variables, wedeveloped a web-based calculator, accessible at https://pectgew2rqefrdqyjgqcrh.streamlit.app/. This tool allows users to input patient-specific examination results, such as albumin, PT, pvspeed and spleen stiffness, to generate individualized predictions for the likelihood of high-risk EGVs. As illustrated in Figure 9, the web-based tool provides the predicted EGV risk rate - in this case, 94.15% - and visualizes the contribution of each variable to the final prediction using a SHAP force plot. The red bars represent features that increase the risk score, such as PT and spleen stiffness, while blue bars indicate features that decrease the risk score, such as higher pvspeed. This interactive platform enables clinicians to intuitively understand the key factors driving the prediction, making it a practical and transparent tool for clinical decision-making in managing patients at risk of EGVs.

Figure 9 Web-based predictive model. The image displays a web-based application for predicting esophagogastric varices (EGVs) using the extreme gradient boosting model. The left side of the interface allowed users to input patient characteristics, such as albumin, prothrombin time, portal vein speed and spleen stiffness. These input fields could be adjusted to reflect the specific values for each patient. The main part of the interface shows the prediction of the model, with the EGV risk rate given as 94.15%. Below this, the SHapley Additive exPlanation force plot provides an explanation of the prediction. The force plot visually represents the contribution of each feature to the final predictive score. Features such as spleen stiffness, prothrombin time and portal vein speed are highlighted, showing how they made the prediction higher or lower. The pink bars indicate features that increased the predictive score (higher risk), while blue bars indicate features that decreased it (lower risk). This tool helps clinicians understand the risk factors contributing to EGVs for individual patients by providing a clear and interactive explanation of the predictions of the model. The SHapley Additive exPlanation force plot ensures transparency in the decision-making process, making it easier for healthcare providers to trust and act on the predictions. EGV: Esophagogastric varices; Alb: Albumin; PT: Prothrombin time; pvspeed: Portal vein speed; spwide: Spleen width; pv: Portal vein; spv: Splenic vein; splong: Spleen length; INR: International normalized ratio.

DISCUSSION

In this study, we developed and validated a noninvasive predictive model using 11 ML algorithms for EGVs in hepatitis B patients. The XGBoost model demonstrated superior performance, achieving an AUC of 0.96 in the validation dataset. The model was effectively interpreted using SHAP, identifying key predictors such as spleen stiffness, pvspeed, PT and albumin levels. These findings suggest that our model can reliably predict the risk of severe EGVs, facilitating targeted endoscopic screening and improving clinical decision-making.

Our study distinguishes itself from previous research primarily through its methodological rigor and comprehensive approach. While earlier studies typically used a single ML algorithm or a few algorithms to predict EGVs[13,14], we utilized an extensive suite of 11 different ML algorithms. This broad evaluation allowed us to meticulously compare and identify the most effective models. The XGBoost model emerged as the superior algorithm, achieving an impressive AUC of 0.96 in our validation dataset. This contrasts sharply with previous studies, which often reported lower AUC values because of their limited algorithmic scope and less rigorous validation processes[15]. Our approach ensures greater reliability and robustness in prediction but also highlights the enhanced accuracy achievable through comprehensive algorithmic evaluation.

In examining the role of albumin in predicting EGVs, our study confirms and extends the findings of previous research. Earlier studies, such as those by Li et al[16] and Majid et al[17], have documented the association between lower albumin levels and increased risk of varices. These studies typically relied on traditional statistical methods and did not fully exploit the predictive potential of albumin within a comprehensive predictive framework. Our study integrated albumin with other significant predictors to demonstrate how albumin, as a marker of liver synthetic function, significantly contributes to the prediction of severe EGVs. The decreased albumin levels reflect impaired liver function and portal hypertension, which are closely associated with the development of varices. By incorporating albumin into a multi-parameter model, we provided a more nuanced and accurate predictive tool, highlighting the importance of advanced analytical techniques in medical research.

PT has long been recognized as a significant indicator of liver function and a predictor of varices. Previous studies, such as those by Li et al[18] have highlighted its relevance but often used conventional statistical analyses that may overlook complex interactions between PT and other predictors. PT reflects the ability of the liver to produce clotting factors, and prolonged PT indicates liver dysfunction and a higher risk of bleeding varices. Our study utilized PT as one of the critical predictors, capturing the intricate relationships between it and other variables. This comprehensive approach enhanced the accuracy of varices prediction, demonstrating the advantage of understanding and utilizing clinical predictors in a multifaceted framework.

The inclusion of pvspeed in our predictive model represents a significant advancement over previous imaging techniques. Unlike other studies that used limited or indirect measures, we used high-resolution ultrasound to directly assess pv hemodynamics. The specific measurement of pvspeed offers a level of precision not typically seen in other imaging studies. Pvspeed is a direct indicator of portal hypertension, a major risk factor for varices. Lower speeds suggest increased resistance in the portal system, indicative of severe varices. This parameter proved to be a crucial predictor, significantly enhancing the performance of the model. The use of ultrasound to measure pvspeed provided a noninvasive yet highly accurate method for assessing EGV risk, setting our study apart from others and highlighting the potential of advanced imaging techniques in clinical prediction models.

Spleen stiffness has been identified as an important marker in previous studies[19,20], which noted its predictive value for portal hypertension and varices. However, these studies often did not integrate spleen stiffness into a comprehensive predictive model, limiting their predictive power. Spleen stiffness reflects the degree of portal hypertension and splenic congestion, which are directly related to the presence of varices. Our study confirmed spleen stiffness as a critical predictor and integrated it with other significant factors, improving the accuracy and reliability of the model. By focusing on spleen stiffness and utilizing advanced techniques, we were able to create a more effective and interpretable predictive model, demonstrating the enhanced capability of our comprehensive approach in accurately predicting severe EGVs. Despite the promising results, this study had some limitations. The retrospective design may have introduced selection bias, and the findings may not be generalized to other populations. Additionally, while the model performed well in our cohort, external validation in different settings is necessary to confirm its utility.

CONCLUSION

Our study demonstrates that a high-accuracy noninvasive predictive model using ML algorithms for EGVs in hepatitis B patients is feasible. This model, supported by a user-friendly web application, holds promise for improving patient care and optimizing resource allocation in clinical practice. Continued research and external validation will be crucial in realizing the full potential of this predictive tool.

ACKNOWLEDGEMENTS

We express our gratitude to the patients who participated in this research.