Machine learning-based prediction models for liver-related events in patients with hepatitis B-related cirrhosis and clinically significant portal hypertension

Abstract

BACKGROUND

Hepatitis B-related cirrhosis represents a major contributor to liver-related events (LREs), with the development of clinically significant portal hypertension (CSPH) serving as a critical milestone in disease progression.

AIM

To establish predictive models based on multiple machine learning algorithms to improve the accuracy and clinical utility of LREs prediction.

METHODS

A total of 576 patients were retrospectively enrolled and randomly divided into training (n = 403) and validation (n = 173) cohorts. Features were selected through least absolute shrinkage and selection operator regression, random forest (RF), and support vector machine (SVM). Based on these features, five predictive models were constructed, including SVM, RF, logistic regression, extreme gradient boosting (XGBoost), and k-nearest neighbor. Model performance was evaluated using receiver operating characteristic and decision curve analysis, and feature importance and interactions were further explored using SHapley Additive exPlanations (SHAP).

RESULTS

Of the patients included, 313 (54.3%) developed LREs. Eight core predictive features were ultimately identified, with the liver stiffness measurement (LSM)-to-platelet ratio (LPR) contributing most significantly. The XGBoost and RF models demonstrated superior performance, achieving accuracies of 0.951 and areas under the curve of 0.975 and 0.965, respectively. SHAP analysis revealed that LPR, hemoglobin (HB), and LSM were key factors, with LPR exhibiting significant interactions with HB, international normalized ratio, and spleen thickness.

CONCLUSION

Machine learning-based prediction models, particularly XGBoost and RF, can effectively identify high-risk individuals among patients with compensated hepatitis B virus-related cirrhosis and CSPH. LPR that incorporates LSM is a valuable and robust predictive indicator.

Key Words: Hepatitis B; Liver cirrhosis; Clinically significant portal hypertension; Machine learning; Liver-related events; Prediction model

Core Tip: This study developed and validated machine learning models to predict liver-related events in patients with compensated hepatitis B virus-related cirrhosis and clinically significant portal hypertension. Among five models, extreme gradient boosting and random forest achieved the best accuracy and clinical utility. The liver stiffness measurement-to-platelet ratio (LPR) emerged as the most influential predictor, interacting with hemoglobin, international normalized ratio, and spleen thickness. These findings highlight machine learning based on LPR as a robust noninvasive method and provide a novel, interpretable tool for early risk stratification and personalized management in compensated cirrhosis.

INTRODUCTION

Liver cirrhosis is a chronic, progressive liver disease characterized by liver fibrosis and structural remodeling[1]. It is one of the leading causes of death and disability worldwide[2]. Hepatitis B virus (HBV) infection is the leading cause of cirrhosis and remains a major global public health challenge[3]. Patients with compensated cirrhosis experience relatively mild clinical symptoms, but as the disease progresses, they eventually develop liver-related events (LREs), including hepatic decompensation, hepatocellular carcinoma (HCC), and liver-related death, which severely affect quality of life and prognosis. However, most patients with HBV-related cirrhosis remain undiagnosed until LREs develop[4,5]. Clinically significant portal hypertension (CSPH) is a key pathophysiological change in the progression of cirrhosis that marks the onset of the high-risk phase[6]. The presence of CSPH not only increases the risk for decompensation but is also closely associated with liver-related mortality[7,8]. Therefore, risk stratification and prognostic assessment of patients with HBV-related cirrhosis and CSPH are of high clinical significance.

The traditional prognostic assessment of liver cirrhosis relies primarily on classic scoring systems, such as the Child-Pugh score and the model for end-stage liver disease score[9]. However, these scoring systems are primarily applicable to patients in the decompensated stage and have limited predictive ability for patients with compensated cirrhosis. In recent years, with the development of noninvasive liver fibrosis assessment technologies, some indicators have been widely used in the assessment of patients with liver cirrhosis. Ultrasound-assessed splenic size is valuable for first hepatic decompensation in patients with metabolic dysfunction-associated steatotic liver disease (MASLD)[10]. Some new plasma biomarkers, such as CHI3 L1, IGFBP1, SHBG, and TIMP2, have the potential to predict HCC in patients with HBV-related cirrhosis[11]. The dynamic fibrosis-4 index can assess fibrosis progression and predict the risk for all-cause mortality, cardiovascular events, and LREs in MASLD[12,13]. Liver stiffness measurement (LSM) using transient elastography has emerged as a highly accurate and non-invasive technique for diagnosing or assessing chronic liver disease. The integrated LSM model is more accurate than existing prediction models for predicting hepatic decompensation in patients with MASLD[14]. LSM is also associated with HCC risk in patients with MASLD and chronic hepatitis B[15,16]. LSM has diagnostic utility for liver fibrosis in patients with chronic hepatitis D. LSM values < 6 kPa can virtually rule out significant fibrosis[17]. Therefore, non-invasive LSM has utility in the diagnosis and prognosis of liver cirrhosis. Beyond LSM alone, several composite indices incorporating LSM have been developed to improve diagnostic and prognostic accuracy. The LSM-to-platelet (PLT) ratio (LPR), has demonstrated good performance in predicting the first hepatic decompensation in patients with MASLD and compensated advanced chronic liver disease[10]. Few studies have investigated the LSM-to-albumin (ALB) ratio (LAR). However, Liang et al[18] proposed that the spleen thickness-age-LSM-ALB algorithm has certain predictive value for high-risk gastroesophageal varices in patients with cirrhosis. In addition, LSM combined with the age-male-ALB-bilirubin-PLT score model can accurately evaluate the stage of liver fibrosis in patients with chronic hepatitis B undergoing treatment[19]. The LSM-spleen diameter-to-PLT ratio score (LSPS) also performed well in predicting varices requiring treatment and avoiding endoscopic examinations[20]. However, Ryu et al[21] found no correlation between LPR or LSPS and the hepatic venous pressure gradient (HVPG) in patients with alcoholic and viral cirrhosis, despite the value of LSM for predicting HVPG levels. These composite indices leverage the complementary information from LSM and other readily available clinical parameters, potentially enhancing risk stratification in patients with cirrhosis.

With the rapid development of artificial intelligence and machine learning technologies in medicine, the construction of predictive models based on big data has become a popular research topic[22]. Machine learning algorithms can process high-dimensional, complex data and identify nonlinear relationships and feature interactions that are difficult to detect using traditional statistical methods. Although previous studies have explored the application of machine learning to predict the prognosis of liver cirrhosis[23], research investigating predictive models for LREs in patients with HBV-related cirrhosis and CSPH remains limited.

Thus, this study aimed to develop a predictive model for LREs in patients with HBV-related cirrhosis and CSPH using multiple machine learning algorithms. This will not only help improve the accuracy of risk stratification for patients but also provide important insights for personalized treatment strategies.

MATERIALS AND METHODS

Participants

This study retrospectively enrolled patients with HBV-related cirrhosis and CSPH admitted to the Beijing Ditan Hospital between July 2015 and July 2024. Initially, 667 patients were screened for eligibility, of whom 576 were included in the analysis after applying the inclusion and exclusion criteria (Figure 1). All patients were followed up until July 31, 2024, the occurrence of LREs, or were lost to follow-up, whichever occurred first. Eligible participants were randomly assigned to the training and validation datasets at a ratio of 7:3. This study was approved by the Ethics Committee of the Beijing Ditan Hospital (Approval No. DTEC-KY2024-069-01). All procedures adhered to the principles of the Declaration of Helsinki.

Figure 1 Flowchart. HBV: Hepatitis B virus; LASSO: Least absolute shrinkage and selection operator; RF: Random forest; SVM: Support vector machine; LR: Logistic regression; XGBoost: Extreme gradient boosting; KNN: K-nearest neighbor; PPV: Positive predictive value; NPV: Negative predictive value; AUC: Area under the receiver operating characteristic curve; SHAP: SHapley Additive exPlanations.

Inclusion criteria: (1) Aged between 18 years and 70 years; (2) Compensated cirrhosis without previous ascites, variceal bleeding, or hepatic encephalopathy; (3) A clear diagnosis of chronic HBV and regular oral antiviral medication; (4) Meeting the diagnosis of CSPH; and (5) Having complete follow-up data. Chronic HBV is currently defined as hepatitis B surface antigen (HBsAg) positive, or HBsAg negative and hepatitis B core antibody positive, with a clear history of chronic HBV infection (HBsAg positive history > 6 months), and other causes have been excluded[24]. The diagnosis of liver cirrhosis was based on fulfillment of ≥ 1 of the following criteria[25,26]: (1) Pathological evidence from liver biopsy; and (2) Typical clinical, laboratory, and imaging features, including liver morphological changes, splenomegaly, thrombocytopenia, and portal vein dilatation. The definition of CSPH was based on the Baveno VII consensus[27,28], including LSM ≥ 25 kPa, or 20 kPa ≤ LSM < 25 kPa and PLT < 150 × 10⁹/L, or 15 kPa ≤ LSM < 20 kPa and PLT < 110 × 10⁹/L.

Exclusion criteria: (1) Previous hepatic decompensation event or diagnosis of HCC; (2) Severe cardiopulmonary and renal insufficiency; (3) With concurrent other liver diseases other than HBV (autoimmune liver disease, alcoholic liver disease, MASLD, etc.); (4) Pregnant or breastfeeding women; (5) Patients who had undergone liver transplantation, splenectomy or transjugular intrahepatic portosystemic shunt; or (6) Incomplete clinical data or follow-up time of less than 6 months.

Data collection

Baseline clinical data were collected, including demographic characteristics (age, gender), laboratory test indicators [aspartate aminotransferase (AST), alanine aminotransferase, total bilirubin (TB), direct bilirubin (DB), ALB, ALB/globulin ratio (A/G), gamma-glutamyltransferase, alkaline phosphatase (ALP), bile acids, creatinine, white blood cell count, hemoglobin (HB), PLT, neutrophil count, lymphocyte count, prothrombin time, international normalized ratio (INR), and Child-Pugh score], imaging findings (portal vein width, spleen thickness, spleen diameter), and LSM. The derivative indices were calculated as follows: LPR = LSM (kPa)/PLT (10⁹/L), LAR = LSM (kPa)/ALB (g/L), and LSPS = [LSM (kPa) × spleen diameter (cm)]/PLT (10⁹/L).

The primary outcome event was defined as the occurrence of LREs, including: (1) Hepatic decompensation (defined as the first occurrence of ascites, esophageal variceal bleeding, or hepatic encephalopathy); (2) HCC (diagnosed according to the American Association for the Study of Liver Diseases guidelines[29]); and (3) Liver-related death (death directly attributed to liver disease progression or complications).

Feature selection strategy

A total of 27 variables from multiple dimensions were collected. Three machine learning algorithms were further used for feature selection to ensure the robustness and reliability of the selected features. Firstly, least absolute shrinkage and selection operator (LASSO) regression was applied for feature screening. The optimal regularization parameter λ was determined using ten-fold cross-validation. As the λ value gradually increased, the coefficients of unimportant features gradually decreased to zero, and finally the feature variables with non-zero coefficients were retained. Secondly, the random forest (RF) algorithm was used to rank the importance of all features based on the mean Gini impurity reduction. Thirdly, the support vector machine (SVM) algorithm was used to evaluate the feature importance using the average rank. The intersection of the features selected by LASSO regression and the top 20 features ranked by the RF and SVM algorithms was analyzed to determine the final core prediction features.

Machine learning model construction

Based on the selected core features, five machine learning prediction models were constructed: (1) SVM: Using the radial basis function kernel and optimizing hyperparameters through grid search; (2) RF: Setting the number of decision trees to 100 and evaluating the model performance through out-of-bag error; (3) Logistic regression (LR): Using L2 regularization and optimizing parameters through maximum likelihood estimation; (4) Extreme gradient boosting (XGBoost): Setting the learning rate to 0.1 and the maximum depth to 6, and using the early stopping mechanism to prevent overfitting; and (5) K-nearest neighbor (KNN): Determining the optimal k value through cross-validation. All models underwent parameter tuning and model training in the training cohort, and their performance was evaluated in the validation cohort.

Model evaluation and validation

The predictive performance of each model was evaluated using multiple metrics, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and area under the receiver operating characteristic curve (AUC). Ten-fold cross-validation was performed on the training cohort to assess model stability and generalizability. Receiver operating characteristic curves were plotted to assess the discriminatory power of the models. Decision curve analysis (DCA) was used to assess the net clinical benefit of each model at different risk thresholds. The clinical utility of the models was assessed by comparing the net benefit of each model with the extreme strategies of “treating all patients” and “treating no patients”. To account for the time-to-event data of LREs and varying follow-up durations, we conducted a survival analysis using Cox proportional hazards regression and random survival forest (RSF) models. Model performance was assessed using Harrell's C-index and time-dependent AUC at 12-, 24-, 36-, and 48-months.

Model interpretability analysis

To improve the clinical interpretability of the models, a detailed interpretability analysis was performed using the SHapley Additive exPlanations (SHAP) method. SHAP, based on the SHapley value concept in game theory, quantifies the contribution of each feature to the model’s predictions. The impacts of feature importance ranking and feature value changes on the prediction results was analyzed by calculating the SHAP value for each feature. SHAP interaction analysis was further performed to explore the synergistic mechanisms among the key features, focusing on the interaction between the most important features and other clinical indicators.

Statistical analysis

Continuous variables are expressed as mean ± SD or median (IQR) depending on their distribution, whereas categorical variables are expressed as n (%). The normality of continuous variables was assessed using the Shapiro-Wilk test. Intergroup comparisons were performed using the t-test for continuous variables that exhibited a normal distribution, and the Mann-Whitney U test for those that did not. Categorical variables were compared using the χ² test or Fisher's exact test. All statistical analyses and machine learning modeling were performed using Python 3.10 software. Differences with P < 0.05 was considered to be statistically significant.

RESULTS

Baseline characteristics

This study enrolled 576 HBV-related cirrhosis patients with CSPH and randomly divided them into a training cohort (n = 403) and a validation cohort (n = 173). The median age of the overall cohort was 57 (51-62) years, and 56.9% of patients were male (Table 1). During the median follow-up of 2.83 (1.42-4.25) years, 313 (54.3%) patients developed LREs. The training and validation cohort exhibited comparable baseline characteristics, with no statistically significant differences in any of the measured variables (P > 0.05 for all comparisons; Supplementary Table 1). There were no significant differences in age or gender between the LREs group and non-LREs group (P > 0.05). However, significant differences were observed in the indicators of liver function and severity of portal hypertension. In the LREs group, the median LSM was significantly increased [30.6 (22-45) kPa vs 22.5 (18.2-30.8) kPa, P < 0.001], ALB [36.9 (28-42.3) g/L vs 42.3 (35.5-46.5) g/L, P < 0.001] and PLT [65 (45-89) × 10⁹/L vs 89 (65-117) × 10⁹/L, P < 0.001] was significantly decreased, and LPR [0.49 (0.33-0.72) vs 0.27 (0.19-0.39)], LAR [0.88 (0.60-1.24) vs 0.56 (0.44-1.83)], LSPS [7.69 (4.68-11.36) vs 3.64 (2.51-6.19)] were significantly increased (P < 0.001). In addition, there were differences in TB, DB, A/G, bile acid, creatinine, INR, blood cell count, Child-Pugh score, spleen size and portal vein width.

Table 1 Baseline characteristics of patients with hepatitis B virus-related cirrhosis and clinically significant portal hypertension.

	Overall (n = 576)	No LREs (n = 263)	LREs (n = 313)	P value
Age (year)	57 (51-62)	55 (48-60)	57 (51-62)	0.092
Sex of male	328 (56.9)	151 (57.4)	179 (57.2)	0.588
Follow-up time (year)	2.83 (1.42-4.25)	2.67 (1.46-4.50)	3.42 (1.42-4.25)	0.437
AST (U/L)	35.1 (26.6-52.0)	33.9 (26.4-49.7)	36.6 (26.6-52)	0.564
ALT (U/L)	30.5 (19.9-45.5)	31.2 (21.7-45.4)	29.3 (17.9-45.7)	0.183
TB (μmol/L)	21.05 (15.3-31.4)	18.5 (13.8-27.5)	23.2 (16.1-34)	< 0.001
DB (μmol/L)	8.65 (5.95-13.3)	7.2 (5.1-12.1)	9.5 (6.4-14.9)	< 0.001
ALB (g/L)	38.9 (29.9-44.8)	42.3 (35.5-46.5)	36.9 (28-42.3)	< 0.001
A/G	1.2 (0.8-1.5)	1.4 (1-1.6)	1.2 (0.8-1.4)	< 0.001
GGT (U/L)	39.9 (21.7-78.2)	38.5 (22-78.6)	41 (20.6-74.4)	0.852
ALP (U/L)	87.9 (65.4-116.9)	86.8 (66.0-111)	89.4 (65-121)	0.381
Bile acid (μmol/L)	19.6 (6.9-41.9)	10.9 (4.7-27.9)	31.5 (13.7-61.3)	< 0.001
Creatinine (μmol/L)	62.5 (52.5-72.4)	60.9 (51-71.1)	65.7 (54-76.3)	0.013
PT (second)	13.2 (12.4-14.4)	13.1 (12.4-14.5)	13.2 (12.4-14.4)	0.635
INR	1.22 (1.14-1.37)	1.19 (1.13-1.37)	1.26 (1.15-1.38)	0.003
WBC (× 109/L)	3.60 (2.66-4.84)	4.2 (3.1-5.34)	3.24 (2.37-4.36)	< 0.001
Neutrophil count (× 109/L)	1.94 (1.47-2.76)	2.28 (1.7-3.19)	1.81 (1.32-2.48)	< 0.001
Lymphocyte count (× 109/L)	1.10 (0.7-1.59)	1.29 (0.89-1.8)	0.91 (0.6-1.35)	< 0.001
HB (g/L)	134 (117-150)	142 (124-159)	129 (106-143)	< 0.001
PLT (× 109/L)	74 (52.8-99.6)	89 (65-117)	65 (45-89)	< 0.001
Child-Pugh score (point)	3 (3-5)	3 (3-3)	5 (4-6)	< 0.001
LSM (kPa)	25.4 (20.3-38)	22.5 (18.2-30.8)	30.6 (22-45)	< 0.001
LPR	0.37 (0.24-0.58)	0.27 (0.19-0.39)	0.49 (0.33-0.72)	< 0.001
LAR	0.71 (0.50-1.11)	0.56 (0.44, 1.83)	0.88 (0.60, 1.24)	< 0.001
LSPS	5.49 (3.19-9.03)	3.64 (2.51-6.19)	7.69 (4.68-11.36)	< 0.001
Portal vein width (mm)	12 (11-13)	12 (11-13)	12 (12-14)	< 0.001
Spleen thickness (mm)	44 (37-54)	42 (36-49)	48 (42-56)	< 0.001
Spleen diameter (cm)	14.8 (12.4-17.1)	13.8 (11.7-16.5)	15.4 (12.9-17.8)	< 0.001

Data are presented as median (IQR) for continuous variables and n (%) for categorical variables. Mann-Whitney U test was used for comparisons of continuous variables between groups; χ² test was used for categorical variables. LRE: Liver-related event; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; TB: Total bilirubin; DB: Direct bilirubin; ALB: Albumin; A/G: Albumin/globulin ratio; GGT: Gamma-glutamyltransferase; ALP: Alkaline phosphatase; PT: Prothrombin time; INR: International normalized ratio; WBC: White blood cell; HB: Hemoglobin; PLT: Platelet; LSM: Liver stiffness measurement; LPR: Liver stiffness measurement-to-platelet ratio; LAR: Liver stiffness measurement-to-albumin ratio; LSPS: Liver stiffness-spleen diameter-to-platelet ratio score.

Feature selection

This study used three machine learning algorithms for feature selection, including LASSO regression, RF, and SVM. LASSO regression determined the optimal regularization parameter λ through ten-fold cross-validation. As the λ value gradually increased, the coefficients of insignificant features gradually shrank to zero (Figure 2A). The trend in the partial likelihood deviation as a function of log(λ) indicated the optimal parameter selection for the model (Figure 2B). Ultimately, 15 predictive feature variables were selected, including LSM, LPR, gender, portal vein width, spleen thickness, lymphocyte count, HB, PLT, INR, AST, DB, A/G, ALP, bile acid, Child-Pugh score. The RF algorithm ranks feature importance according to the average Gini impurity reduction (Figure 2C). The larger the average Gini impurity reduction, the greater the feature importance. The SVM algorithm used the average ranking method to evaluate feature importance (Figure 2D). The lower the average ranking value, the greater the feature importance. The intersection of the 15 features screened by LASSO regression and the top 20 features of RF and SVM was analyzed (Figure 2E), and eight core predictive features were identified, including LSM, LPR, spleen thickness, HB, PLT, INR, AST, and ALP.

Figure 2 Feature selection. A: Curve showing the change in variable coefficients for different λ values in least absolute shrinkage and selection operator (LASSO) regression; B: Curve showing the relationship between partial likelihood deviation and log(λ); C: Feature importance ranking calculated by the random forest (RF) algorithm; D: Feature importance ranking evaluated by the support vector machine (SVM) algorithm; E: Intersection Venn diagram of features screened by LASSO, RF, and SVM. LSPS: Liver stiffness-spleen diameter-to-platelet ratio score; LPR: Liver stiffness measurement-to-platelet ratio; ALB: Albumin; LAR: Liver stiffness measurement-to-albumin ratio; LSM: Liver stiffness measurement; PLT: Platelet; HB: Hemoglobin; AST: Aspartate aminotransferase; GGT: Gamma-glutamyltransferase; TB: Total bilirubin; ALP: Alkaline phosphatase; INR: International normalized ratio; WBC: White blood cell; PT: Prothrombin time; DB: Direct bilirubin; ALT: Alanine aminotransferase; A/G: Albumin/globulin ratio; LASSO: Least absolute shrinkage and selection operator.

Prediction model construction and performance evaluation

Based on the eight core features identified, this study constructed five machine learning prediction models, including SVM, RF, LR, XGBoost, and KNN. In the training cohort (Figure 3A), the RF and XGBoost models demonstrated excellent predictive performance, with an accuracy of 0.951, sensitivity and specificity of 0.950 and 0.952, F1 scores of 0.953, and AUC values of 0.975 and 0.965, respectively. The SVM model showed moderate predictive performance, with an accuracy of 0.775, a sensitivity of 0.789, a specificity of 0.759, and an AUC of 0.834. The KNN model performed similarly to the SVM, with an accuracy of 0.769 and an AUC of 0.833. The LR model performed relatively poorly across all metrics, with an accuracy of 0.731 and an AUC of only 0.790 (Table 2). The cross-validation results yielded mean AUC values of 0.842 ± 0.045 for RF, 0.830 ± 0.061 for XGBoost, 0.787 ± 0.060 for SVM, 0.740 ± 0.055 for LR, and 0.737 ± 0.045 for KNN. These performance estimates demonstrated reasonable consistency across folds and were aligned with the performance of the subsequent validation set. In the validation cohort (Figure 3B), the predictive performance of some models declined compared with that of the training cohort but still maintained good discriminatory ability. RF achieved the highest AUC (0.922), followed by XGBoost (0.890), SVM (0.829), and LR and KNN (both 0.814). XGBoost showed a larger performance decrease from training to validation (0.075) than RF (0.053), indicating that RF had a superior generalization capability. The net clinical benefit of each model was assessed using DCA. In the training cohort (Figure 3C), the XGBoost and RF models demonstrated the highest net benefit within the risk threshold range of 0.1-0.8, significantly outperforming the extreme strategies of “treat all” and “treat none”. In the validation cohort (Figure 3D), the RF and XGBoost showed positive net benefits within the same risk threshold range, which were better than those of the SVM, LR, and KNN models.

Figure 3 Predictive performance of five machine learning models. A: Receiver operating characteristic (ROC) curves in the training cohort; B: ROC curves in the validation cohort; C: Decision curve analysis in the training cohort; D: Decision curve analysis in the validation cohort. TPR: True positive rate; FPR: False positive rate; XGBoost: Extreme gradient boosting; SVM: Support vector machine; AUC: Area under the receiver operating characteristic curve; KNN: K-nearest neighbor.

Table 2 Performances of the machine learning models for predicting liver related events.

Model	Accuracy	Recall	F1 score	Sensitivity	Specificity	PPV	NPV	AUC	CV-AUC
SVM	0.775	0.789	0.785	0.789	0.759	0.780	0.768	0.834	0.787 ± 0.060
RF	0.951	0.950	0.953	0.950	0.952	0.955	0.946	0.975	0.842 ± 0.045
LR	0.731	0.739	0.741	0.739	0.723	0.743	0.719	0.790	0.740 ± 0.055
XGBoost	0.951	0.950	0.953	0.950	0.952	0.955	0.946	0.965	0.830 ± 0.061
KNN	0.769	0.756	0.773	0.756	0.783	0.791	0.747	0.833	0.737 ± 0.045

PPV: Positive predictive value; NPV: Negative predictive value; AUC: Area under the receiver operating characteristic curve; CV: Cross-validation; SVM: Support vector machine; RF: Random forest; LR: Logistic regression; XGBoost: Extreme gradient boosting; KNN: K-nearest neighbor.

Survival analysis outcomes were presented in Supplementary Table 2. Within the training cohort, the RSF model yielded a C-index of 0.827, outperforming the Cox regression model which achieved 0.784. The time-dependent AUC values for the RSF model varied between 0.747 and 0.886, whereas the Cox regression model exhibited comparatively lower performance, with AUC values ranging from 0.745 to 0.823. When applied to the validation cohort, the RSF model demonstrated good predictive capability, achieving a concordance C-index of 0.791 and time-dependent AUC values between 0.642 and 0.881.

SHAP-based interpretability analysis

To gain a deeper understanding of model’s prediction mechanism and improve its clinical interpretability, this study used the SHAP method to conduct a detailed analysis of the model’s decision-making process. SHAP analysis results revealed the contribution of each feature to LREs prediction and its mode of action. Feature importance analysis (Figure 4A) revealed that LPR was the most influential predictor, with an average SHAP value of 1.53, markedly higher than that of HB (0.99), LSM (0.70), and all other variables. SHAP value distribution analysis further revealed association patterns between each feature and the risk for LREs (Figure 4B). High LPR values were primarily distributed within the positive SHAP value region, indicating that elevated LPR significantly increased the risk for LREs. HB exhibited the opposite pattern, with low HB values corresponding to positive SHAP values, suggesting that anemia is a significant risk factor for LREs. High LSM values were also associated with positive SHAP values, suggesting that LSM is a risk factor for LREs.

Figure 4 Model interpretability analysis based on SHapley Additive exPlanations. A: Feature importance ranking; B: SHapley Additive exPlanations value distribution. LPR: Liver stiffness measurement-to-platelet ratio; HB: Hemoglobin; LSM: Liver stiffness measurement; AST: Aspartate aminotransferase; ALP: Alkaline phosphatase; INR: International normalized ratio; PLT: Platelet; SHAP: SHapley Additive exPlanations.

SHAP-based feature interaction analysis

SHAP interaction analysis of the key features focused on the synergistic interaction between the most important features. In the LPR-HB interaction analysis (Figure 5A), the effect of HB was significantly amplified with increasing LPR values. In particular, within the LPR range > 0.5, lower HB values were associated with higher positive SHAP values, indicating that the coexistence of elevated LPR and anemia significantly increased the risk for LREs. The LPR-INR interaction analysis (Figure 5B) demonstrated that higher INR values produced a stronger risk prediction signal in the context of elevated LPR, reflecting the mutually reinforcing effect between INR and LPR. The LPR-spleen thickness interaction analysis (Figure 5C) showed that the synergistic effect of spleen thickening and elevated LPR played an important role in predicting LREs. The PLT-spleen thickness interaction analysis (Figure 5D) further demonstrated a significant negative interaction effect between increased spleen thickness and thrombocytopenia.

Figure 5 Feature interaction analysis based on SHapley Additive exPlanations. A: Interaction between liver stiffness measurement-to-platelet ratio (LPR) and hemoglobin; B: Interaction between LPR and international normalized ratio; C: Interaction between LPR and spleen thickness; D: Interaction between platelet and spleen thickness. LPR: Liver stiffness measurement-to-platelet ratio; HB: Hemoglobin; INR: International normalized ratio; PLT: Platelet; SHAP: SHapley Additive exPlanations.

DISCUSSION

This retrospective analysis of patients with compensated HBV-related cirrhosis and CSPH showed that the incidence of LREs was 54.3% during a median follow-up period of 2.83 years. This finding reflected the high-risk nature of the cohort. LSM obtained using transient elastography has become a widely validated, non-invasive tool for assessing liver fibrosis and portal hypertension. In this study, we incorporated LSM into a machine learning framework combined with multidimensional clinical variables to achieve more accurate risk prediction. Eight core predictive variables (LSM, LPR, spleen thickness, HB, PLT, INR, AST, and ALP), determined after feature selection, showed good predictive efficacy. The RF and XGBoost models performed the best with AUC values of 0.975 and 0.965, respectively. LPR was identified as the most important predictive feature. These findings provide a new tool for early risk stratification and personalized management of patients with HBV-related cirrhosis and CSPH.

LSM is not only closely related to the degree of fibrosis, but can also effectively predict the risk for hepatic decompensation and HCC in patients with chronic liver disease[14-16]. Dynamic LSM monitoring can also predict the clinical efficacy in patients with portal hypertension and primary biliary cholangitis[30,31]. The Baveno VII consensus even combined LSM with PLT for the non-invasive diagnosis of CSPH to replace invasive tests[27]. In addition, LSM-derived indices, such as LSPS, exhibited a stronger ability to predict high-risk esophageal varices and variceal rebleeding[32,33]. However, traditional prediction models often ignored the potential nonlinearities and interactions between different clinical variables. As such, the present study incorporated LSM and its derivative indicators into a machine learning framework, combined with multidimensional clinical variables, to achieve more accurate risk prediction and explore its potential application in compensated patients.

We selected eight important clinical features and further constructed and compared the performance of five machine learning algorithms for predicting LREs. Results revealed that XGBoost and RF significantly outperformed the other models, with accuracies of 0.951 and an AUC of 0.975 and 0.965, respectively. Both models demonstrated balanced sensitivity (0.950) and specificity (0.952), demonstrating their superior ability to discriminate between high-risk and low-risk patients. Although both XGBoost and RF achieved excellent performance in the training cohort, RF demonstrated better generalization in the validation cohort. The larger AUC decrease for XGBoost (0.075 vs 0.053 for RF) may be attributed to its gradient boosting mechanism, which sequentially corrects errors and overfits to training-specific patterns. In contrast, RF parallel tree construction with bootstrap aggregation provides better variance reduction and more robust generalization. Notably, SVM showed excellent stability (a decrease of only 0.005), but its lower absolute performance (AUC = 0.829) makes it less suitable for clinical use. Although cross-validation results suggested an overfitting tendency, the robust performance of the independent validation cohort confirmed the clinical applicability of these models. These findings support the use of RF and XGBoost as the optimal models for clinical implementation, offering the best combination of high predictive accuracy and acceptable generalization stability. This advantage may stem from the fact that both are ensemble learning algorithms, which effectively reduce the risk for overfitting by integrating multiple decision trees and can capture complex nonlinear relationships and high-order interactions. These findings are consistent with those of previous studies demonstrating the utility of RF and XGBoost in identifying liver cirrhosis, assessing fibrosis severity, and predicting clinical outcomes in chronic liver disease[34-37]. The RF is recommended as the optimal model for clinical implementation because it offers the best combination of high predictive accuracy and acceptable generalization stability. Importantly, the enhanced time-dependent performance of our RSF model highlights the benefits of integrating temporal data into the LREs risk assessment. The consistently elevated time-dependent AUC values observed throughout the follow-up period suggest that this model can categorize patients according to appropriate surveillance schedules, which may facilitate the development of more individualized monitoring approaches.

SHAP analysis revealed that LPR was the most important predictor, with high LPR values predominantly distributed in the positive SHAP value region, demonstrating a clear dose-response relationship. As the ratio of LSM to PLT, LPR comprehensively reflects the two key pathological processes of liver fibrosis and the severity of portal hypertension, making it more biologically reasonable than a single indicator. High LPR values indicate severe fibrosis and portal hypertension with significantly elevated LSM and decreased PLT. This fusion of dual pathological signals makes it an ideal indicator for predicting LREs. Notably, virtually all patients with LPR > 0.5 were classified in the high-risk category, providing a clear threshold for clinical risk stratification. LSM exhibited a nonlinear pattern, with SHAP values remaining relatively stable below 25 kPa but increasing significantly beyond 30 kPa. This threshold effect is aligned with the pathophysiology of cirrhosis, in which exceeding critical LSM values substantially increases the risk for adverse events. HB demonstrated an inverse pattern, with low values associated with positive SHAP values, reflecting mechanisms, such as hypersplenism-induced red blood cell destruction, malnutrition, chronic inflammation, impaired hepatic synthesis, and the potential risk for chronic gastrointestinal blood loss[38]. Therefore, decreased HB levels often precede the onset of clinical symptoms and serve as important indicators for the early identification of high-risk patients.

Complex feature interactions and pathophysiological insights exist. SHAP interaction analysis revealed complex synergistic effects among the predictive features, reflecting multisystemic and multifaceted pathophysiological changes. The LPR-HB interaction demonstrated pronounced amplification, particularly at LPR > 0.5 combined with low HB. This interactive effect reflects the superposition of three pathological mechanisms: Hypersplenism-induced cytopenia[39], chronic inflammation suppressing erythropoiesis, and impaired hepatic synthetic function affecting hematopoietic factors. Bothou et al[40] found that HB concentration was negatively correlated with portal vein blood flow and was a strong independent predictor of hepatic decompensation in outpatients with cirrhosis. When LPR coexists with anemia, the patient has entered a high-risk stage of hepatic decompensation, and multisystem dysfunction significantly increases the risk for adverse events. The LPR-spleen thickness interaction demonstrated a correlation between anatomical changes and functional abnormalities in portal hypertension, whereas the negative PLT-spleen thickness interaction directly quantified the severity of hypersplenism. Spleen thickness is an independent risk factor for high-risk esophageal varices, and combining LSM and spleen stiffness measurements can significantly reduce the rate of missed diagnoses of high-risk esophageal varices[41]. The spleen thickness-PLT-ALB algorithm can also improve the detection rate of high-risk esophageal and gastric varices[18]. These complex interaction patterns revealed the interdependence and mutual reinforcement of various biological systems in liver cirrhosis, providing a theoretical basis for the development of more accurate individualized predictive models.

Although this study achieved relatively promising results, there were some limitations. First, this was a single-center, retrospective study, and sample selection may have been biased, which could have affected the generalizability of the model. Second, despite the preventive measures implemented, a potential tendency for overfitting still exists, reflecting our modest sample size and the complexity of predicting heterogeneous clinical outcomes. This underscores the need for prospective external validation in independent multicenter cohorts. Third, the simultaneous inclusion of LSM, PLT, and their derived LPR ratio introduced moderate multicollinearity, but our tree-based machine learning algorithms demonstrated robust predictive performance without evidence of harmful overfitting. Fourth, the limited follow-up period may underestimate long-term predictive efficacy because some patients did not experience LREs during observation. Fifth, this study only included patients with HBV-related cirrhosis, and it is unclear whether the model is applicable to cirrhosis of other etiologies, such as alcohol or MASLD. Furthermore, CSPH diagnosis relied on noninvasive Baveno VII criteria rather than the gold-standard HVPG measurement, which may have affected the accuracy of CSPH classification in our cohort, and the model's performance may differ in populations in which CSPH is confirmed by invasive HVPG measurement. Future studies should validate our predictive model in cohorts with HVPG-confirmed CSPH to enhance its clinical applicability and generalizability. Future studies are needed to perform prospective external validation in multicenter populations with different etiologies, explore the incorporation of dynamic indicators (such as changes of LSM trends and dynamic monitoring of hematological parameters) into the model, and integrate them with clinical decision support systems to link the prediction results with intervention strategies to achieve precision medicine.

CONCLUSION

This study constructed and validated a machine learning-based prediction model that effectively identified individuals at risk for LREs among patients with compensated HBV-related cirrhosis and CSPH. The XGBoost and RF models demonstrated the best performance and stability. LPR has emerged as a pivotal predictor of complex interactions with other clinical indicators. These findings introduce a novel tool for the early risk stratification and individualized management of patients with compensated cirrhosis.

ACKNOWLEDGEMENTS

Authors are grateful to all members of Center for Integrative Medicine of Beijing Ditan Hospital for their contributions to the manuscript preparation.