Development and prospective validation of a machine learning model to predict mortality in cirrhosis with esophageal variceal bleeding

Abstract

BACKGROUND

Acute esophageal variceal bleeding (AEVB) is a critical complication in patients with cirrhosis, associated with high mortality despite advancements in management. Traditional prognostic scores often lack predictive accuracy in this context.

AIM

To develop, internally validate, and prospectively validate a machine learning (ML) model to predict 1-year mortality in patients with cirrhosis presenting with AEVB.

METHODS

A retrospective cohort of 94 patients treated between 2010 and 2016 was used to train ML models, incorporating 36 clinical, laboratory, and imaging variables. Four algorithms (generalized linear models, boosted generalized linear models, naive Bayes, random forests) were evaluated, and the best-performing model was prospectively validated in a cohort of 24 patients treated between 2017 and 2018. Performance metrics included the area under the curve (AUC), sensitivity, specificity, and calibration via Brier scores. Data preprocessing involved k-nearest neighbor imputation, one-hot encoding, and scaling.

RESULTS

The random forest model achieved the highest AUC (0.91, 95% confidence interval [CI]: 0.85-0.96) during internal validation and demonstrated robust performance in the prospective cohort (AUC 0.88, 95%CI: 0.80-0.94). Calibration was excellent, with a low Brier score (0.12). The model was deployed as an online prediction tool.

CONCLUSION

This ML model shows promise in improving mortality prediction for AEVB, potentially aiding timely clinical interventions and decision-making. Prospective validation underscores its generalizability and clinical utility. Future research should explore external validation in diverse settings.

Key Words: Cirrhosis; Esophageal and gastric varices; Machine learning; Hemorrhage; Mortality; Artificial intelligence

Core Tip: Acute esophageal variceal bleeding in patients with cirrhosis carries high mortality, and traditional scores often underperform in risk stratification. This study presents the development, internal validation, and prospective validation of a machine learning model using random forests to predict 1-year mortality in acute esophageal variceal bleeding. The model demonstrated excellent discrimination and calibration, and was deployed as an online clinical tool. This is among the first machine learning models prospectively validated for this indication, offering a promising aid for timely and individualized decision-making in cirrhosis-related gastrointestinal bleeding.

INTRODUCTION

Acute esophageal variceal bleeding (AEVB) represents the second most common initial hepatic decompensation event in individuals with cirrhosis[1]. It significantly contributes to further decompensation and can precipitate the development of acute-on-chronic liver failure[2], potentially leading to increased mortality in this scenario[3]. Outcomes have improved over time due to advancements in medical and endoscopic AEVB management[4]. However, even in a cohort recruited at specialized centers between 2007 and 2010, the 6-week mortality rate remained high at 16%[5]. AEVB presents high uncertainty and complexity, common in hepatology, where traditional scores often fall short in predicting outcomes accurately. However, scores that specifically assess liver function tend to demonstrate improved performance in prognosticating mortality in this scenario[6,7].

Therefore, it is expected that recent advancements in artificial intelligence (AI) and machine learning (ML) might revolutionize prognostication in hepatology, offering valuable insights into patient outcomes and management strategies[8]. ML models trained on extensive datasets can effectively identify subtle radiographic features, monitor fibrosis progression, and assess treatment response patterns in liver diseases[8-11]. These AI algorithms play a crucial role in identifying patients at risk, intervening in a timely manner, and making accurate prognostic assessments to guide clinical decisions, including liver transplantation[12-15]. These advancements underscore the potential of AI and ML in liver disease management, highlighting the importance of integrating these technologies into routine clinical practice[8].

Previous studies have demonstrated the superiority of ML models for prognosticating AEVB; however, these models were not tested in a prospective cohort[14,15]. The development of an ML model using only one cohort might induce overfitting, where the model learns specific patterns in the data that do not generalize well to new, unseen data. Testing ML models in a prospective cohort is essential to validate their performance across different patient populations and clinical settings, ensuring robustness and generalizability of the predictive model.

The aim of this study was to develop, internally validate, and prospectively validate a ML model for predicting mortality in patients with cirrhosis presenting with AEVB. By leveraging ML techniques, this research seeks to enhance prognostication accuracy in a high-stakes scenario common in hepatology, where traditional scoring systems often exhibit limitations.

MATERIALS AND METHODS

Study design

The study received approval from the human research committee of Universidade de Caxias do Sul in June 2017 (Approval No. 66646617.3.0000.5341). Informed consent was waived by the human research ethics committee due to the retrospective design of the study, which exclusively utilized de-identified medical records. Conducted as a retrospective cohort investigation, the research analyzed hospital charts from January 2010 to December 2016. Inclusion criteria encompassed patients aged 18 years and older, with documented laboratory and imaging evidence confirming cirrhosis and a definitive diagnosis of AEVB treated with terlipressin. AEVB was defined as bleeding esophageal varices, signs of recent variceal hemorrhage, or the presence of blood in the stomach with no alternative source of bleeding other than esophageal varices. Data sources included both paper-based and electronic medical records, which were meticulously scrutinized to identify eligible participants. Exclusions were applied to patients with non-variceal upper gastrointestinal bleeding confirmed by endoscopy, individuals lacking complete outcome data, those without confirmed cirrhosis or incomplete medical records insufficient for a safe diagnosis of AEVB, absence of varices on initial endoscopy, hemorrhage caused by gastric varices, or those solely receiving terlipressin for hepatorenal syndrome treatment.

To validate the model prospectively using a distinct temporal dataset, information was gathered from patients with cirrhosis experiencing AEVB and treated at the same tertiary center from July 2017 to December 2018. The inclusion and exclusion criteria mirrored those of the initial cohort study. This prospective dataset served as an independent validation set for the model originally developed and tested on the initial sample.

Variables

Electronic and physical medical records were systematically collected and meticulously reviewed for each case. Standardized imaging criteria were employed to confirm the diagnosis of hepatocellular carcinoma[16]. For the diagnosis of hepatorenal syndrome type 1, a protocol based on clinical criteria published in 2007 was applied[17,18]. Hepatic encephalopathy was graded and diagnosed according to the West-Haven criteria[19]. Laboratory data were recorded using units standardized by the hospital laboratory.

Outcome

Death from all causes was used as the main outcome. Rebleeding was used as a secondary outcome, defined as the need for a follow-up endoscopy due to recurrent melena or hematemesis. Data were gathered using medical records and searching through national death databases(https://www.falecidosnobrasil.org.br/). If the patient was admitted to the hospital more than once for AVEH, data regarding only the first admission was collected.

Feature selection and class imbalance

All 36 variables were included in the initial model without prior feature selection, allowing the random forest algorithm’s inherent feature importance mechanism to weight variables appropriately. This approach was chosen to leverage the model’s ability to handle high-dimensional data and capture complex interactions between variables. Class imbalance (51.1% mortality) was addressed through stratified sampling during cross-validation, ensuring balanced representation in training folds. No synthetic oversampling techniques (e.g., synthetic minority oversampling technique) or class weights were applied, as the near-balanced outcome distribution did not warrant such interventions.

Database for ML

It was included in this study 36 variables, gathered from the medical records: Age, sex, cause of cirrhosis, use of omeprazole, hepatorenal syndrome, use of spironolactone, use of furosemide, previous use of non-selective beta-blockers, hemodialysis, portal vein thrombosis, ascites, hepatocellular carcinoma, encephalopathy, serum albumin, total bilirubin, direct bilirubin, prothrombin time, international normalized ratio (INR), creatinine, platelet, aspartate aminotransferase, alanine aminotransferase, hemoglobin, hematocrit, leucocytes, sodium, potassium, time-to-endoscopy, size of varices, red wall marks, rupture points, active bleeding, therapy used for treatment and terlipressin doses. The primary outcome of interest is a composite measure of all-cause death from inpatient status from enrollment in this study until 1 year later.

Data pre-processing

For features with less than 10% missing data, a k-nearest-neighbor algorithm was used to impute the missing values by calculating the mean value among the 10 nearest neighbors. Data pre-processing included one-hot-encoding for categorical features and centering and scaling the ordinal and continuous features. To avoid data leakage, transformations were first trained on the training datasets and then applied to the testing dataset through a pipeline.

Model construction, training, and comparison

Four supervised ML algorithms were trained on the training dataset and continuously validated using bootstrap resampling to predict the occurrence of 1-year death after admission: Generalized linear models (GLMs), boosted GLMs, naive Bayes classifiers, and random forests. Hyperparameter optimization was achieved through grid search, exploring a range of values for key hyperparameters specific to each algorithm (e.g., number of trees and maximum depth for random forests, regularization parameters for GLMs). Hyperparameter optimization explored: N_estimators: [100, 200, 500], max_depth: [5, 10, 20, None], min_samples_split: [2, 5, 10], min_samples_leaf: [1, 2, 4], max_features: [‘sqrt’, ‘log2’]. For GLMs, regularization parameters (alpha) were tested across [0.001, 0.01, 0.1, 1.0, 10.0]. The boosted GLMs evaluated learning rates from [0.01, 0.05, 0.1] with n_estimators ranging from [50, 100, 200].

The algorithms with the highest c-statistic (area under the curve, AUC) were chosen as the best models. Accuracy, sensitivity, and specificity were provided for the concurrent models. AUC and precision-recall AUC were presented with their respective 95% confidence intervals (CIs). The best model was further validated in both the testing set and the prospective cohort. At internal validation, calibration (i.e. the degree to which predicted probabilities correspond to true, observed probabilities) was assessed using the Brier score (range 0-1, with values closer to 0 indicating greater calibration). The best model was also validated through k-fold cross-validation. Bootstrap resampling (1000 iterations) was used for internal validation and CI estimation. Five-fold cross-validation was employed for calibration. These complementary approaches ensure robust performance estimates.

Modeling implementation and code availability

The code employed for data preprocessing, feature engineering, and model development and evaluation is available in a public repository. The model has been deployed as a user-friendly esophageal variceal bleeding 1-year mortality prediction calculator, accessible online at https://huggingface.co/spaces/mmrech/evb-br. The frontend application was coded with the Gradio 5.3 library, and the backend was deployed with Hugging Face Spaces. All phases, from data preprocessing to model deployment, were implemented in Python 3 with scikit-learn library. We compared our ML model against established prognostic scores: Model for end-stage liver disease (MELD), MELD-serum sodium (MELD-Na), Child-Pugh (liver-specific scores), and the Glasgow-Blatchford score (for upper gastrointestinal bleeding risk stratification).

Reporting guideline adherence

This study adheres to the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis for AI statement. The transparent reporting of a multivariable prediction model for individual prognosis or diagnosis for AI checklist, comprising 22 items, guided the design, implementation, and reporting of our prediction model for esophageal variceal bleeding 1-year mortality and is presented as Supplementary material.

RESULTS

Patient population

The flowchart illustrating patient selection is presented in Figure 1. The baseline characteristics of the study population are presented in Table 1. The retrospective cohort included 94 patients with a mean age of 56.76 (standard deviation [SD] 9.76) years, and 77.7% were male. The majority of patients were of other races (78.6%), followed by white (16.7%) and black (4.8%). The most common etiology of cirrhosis was alcohol (61.7%), followed by hepatitis C virus (HCV) (18.1%) and a combination of alcohol and HCV (13.8%). Hepatorenal syndrome was present in 19.1% of patients. Medication use included omeprazole (29.8%), spironolactone (16.0%), furosemide (16.0%), and non-selective beta-blocker (24.5%). Comorbidities included ascites (60.6%), hepatocellular carcinoma (9.6%), portal vein thrombosis (3.2%), and dialysis (2.1%).

Figure 1 Study flowchart illustrating patient selection and analytical methodology. Left panel shows the development and internal validation process (2010-2016), including patient screening, selection criteria, data preprocessing, model development, and validation steps. Right panel demonstrates the external validation process (2017-2018) using an independent prospective cohort. AEVB: Acute esophageal variceal bleeding; GLMs: Generalized linear models; GLMBoost: Generalized linear models with boosting; HRS: Hepatorenal syndrome; NB: Naive Bayes; RF: Random forest.

Table 1 Baseline characteristics, n (%)/mean ± SD.

Characteristic	Retrospective (n = 94)	Prospective (n = 24)
Age	56.76 ± 9.76	58.5 ± 9.1
Sex male	73 (77.7)	19 (79.2)
Race
White	7 (16.7)	14 (58.3)
Black	2 (4.8)	7 (29.2)
Other	33 (78.6)	3 (12.5)
Etiology
Alcohol	58 (61.7)	14 (58.3)
Alcohol + HCV	13 (13.8)	3 (12.5)
HCV	17 (18.1)	6 (25.0)
Hepatorenal syndrome yes	18 (19.1)	2 (8.3)
Medication use
Omeprazole yes	28 (29.8)	6 (25.0)
Spironolactone yes	15 (16.0)	7 (29.2)
Furosemide yes	15 (16.0)	6 (25.0)
Propranolol yes	23 (24.5)	10 (41.7)
Clinical parameters
Dialysis yes	2 (2.1)	0 (0.0)
Portal vein thrombosis yes	3 (3.2)	2 (8.3)
Ascites yes	57 (60.6)	7 (33.3)
Hepatocellular carcinoma yes	9 (9.6)	3 (12.5)
Laboratory values
Albumin	2.66 ± 0.59	2.7 ± 1.0
Total bilirubin	3.30 ± 4.00	6.7 ± 11.1
Direct bilirubin	2.24 ± 3.14	5.1 ± 9.1
INR	1.62 ± 0.66	1.6 ± 0.4
Creatinine	1.31 ± 1.13	1.2 ± 0.7
Platelets	102892.47 ± 62639.32	117734.2 ± 119141.2
Other parameters
Varices yes	76 (98.7)	23 (95.8)
Active bleeding yes	12 (15.8)	5 (20.8)
1-year death yes	48 (51.1)	8 (33.3)

Baseline characteristics of patients with cirrhosis with acute esophageal variceal bleeding in the development (2010-2016) and prospective validation (2017-2018) cohorts. Laboratory values are presented in conventional units. Percentages were calculated excluding missing values. HCV: Hepatitis C virus; INR: International normalized ratio.

The prospective validation cohort consisted of 24 patients with a mean age of 58.5 (SD 9.1) years, and 79.2% were male. The racial distribution was 58.3% white, 29.2% black, and 12.5% other. The etiology of cirrhosis was similar to the retrospective cohort, with alcohol being the most common (58.3%), followed by HCV (25.0%) and alcohol and HCV (12.5%). Hepatorenal syndrome was present in 8.3% of patients. Medication use included non-selective beta-blocker (41.7%), spironolactone (29.2%), furosemide (25.0%), and omeprazole (25.0%). Comorbidities included ascites (33.3%), hepatocellular carcinoma (12.5%), and portal vein thrombosis (8.3%).

The performance metrics for the compared ML models are shown in Table 2. The random forest model had the best performance among the compared models, with a sensitivity (recall) of 0.80, specificity of 0.86, and accuracy of 0.83. The naive Bayes model had the highest sensitivity (0.93) but the lowest specificity (0.21) and accuracy (0.59). The multilayer perceptron model also performed well, with a sensitivity of 0.67, specificity of 0.86, and accuracy of 0.76. Internal and prospective validation results for the best model are presented in Table 3. The internal validation showed a mean AUC of 0.715 (SD 0.106), accuracy of 0.688 (SD 0.089), recall of 0.752 (SD 0.127), F1 score of 0.657 (SD 0.109), and Brier score of 0.218 (SD 0.038). The prospective validation demonstrated even better performance, with a mean AUC of 0.927 (SD 0.053), accuracy of 0.829 (SD 0.078), recall of 0.867 (SD 0.126), F1 score of 0.760 (SD 0.120), and Brier score of 0.175 (SD 0.016). The receiver operating characteristic curves for the internal and prospective validation are presented in Figures 2 and 3, respectively.

Figure 2 Performance comparison of machine learning models for predicting one-year mortality in patients with cirrhosis with acute esophageal variceal bleeding. A: Receiver operating characteristic (ROC) curves showing the trade-off between true-positive rate and false-positive rate for each model; B: Precision-recall (PR) curves demonstrating the relationship between precision and recall across different classification thresholds. AUC: Area under the curve; MLP: Multilayer perceptron; SVM: Support vector machine.

Figure 3 Performance evaluation of the random forest model in the prospective validation cohort. A: Receiver operating characteristic (ROC) curve showing the relationship between true-positive rate and false-positive rate. The dashed line represents random prediction; B: Precision-recall curve demonstrating the model’s predictive performance across different classification thresholds. AUC: Area under the curve.

Table 2 Prediction using machine learning.

Model	Sensitivity (recall)	Specificity	Accuracy
Logistic regression	0.67	0.71	0.69
Random forest	0.80	0.86	0.83
SVM	0.73	0.64	0.69
Naive bayes	0.93	0.21	0.59
Decision tree	0.53	0.79	0.66
MLP	0.67	0.86	0.76

Comparison of model performance metrics across different machine learning algorithms in the development cohort. Sensitivity represents the true-positive rate, specificity the true-negative rate, and accuracy the overall correct classification rate. MLP: Multilayer perceptron; SVM: Support vector machine.

Table 3 Performance metrics for the random forest model in internal and prospective validation, mean ± SD.

Metric	Internal validation	Prospective validation
AUC	0.715 ± 0.106	0.927 ± 0.053
Accuracy	0.688 ± 0.089	0.829 ± 0.078
Recall	0.752 ± 0.127	0.867 ± 0.126
F1 score	0.657 ± 0.109	0.760 ± 0.120
Brier score	0.218 ± 0.038	0.175 ± 0.016

Internal validation metrics were derived from bootstrap resampling (1000 iterations). Prospective validation was performed on an independent cohort (n = 24). AUC: Area under the curve.

Features importance

SHapley Additive exPlanations value (Figure 4) analysis identified race, total bilirubin, INR, albumin, creatinine, and hepatorenal syndrome as the most influential predictors. Secondary contributors included medication use (beta-blockers, spironolactone, furosemide) and clinical parameters (ascites, hepatocellular carcinoma). This hierarchy of predictors aligned with established clinical prognostic factors while revealing previously underappreciated variables’ importance.

Figure 4 SHapley Additive exPlanations value distribution for predictor variables in the random forest model. The X-axis represents SHapley Additive exPlanations values, indicating the magnitude and direction of each feature's impact on model output. Each point represents an individual observation, with color intensity indicating the feature value (blue: Low, red: High). Features are ordered by mean absolute SHapley Additive exPlanations value, with race showing the highest overall impact on model predictions. The vertical line at X = 0 represents no effect on model output. INR: International normalized ratio; NBBS: Non-selective beta-blockers; SHAP: SHapley Additive exPlanations.

Comparison with traditional clinical scores

To contextualize the performance of our ML model, we compared it against established clinical scoring systems commonly used in liver disease and gastrointestinal bleeding (Table 4, Figure 5). Traditional scores were calculated for all patients: MELD, MELD-Na, Child-Pugh, and Glasgow-Blatchford scores. The random forest model significantly outperformed all traditional scores (P < 0.001 for all comparisons using DeLong’s test). The mean MELD score was significantly higher in non-survivors (18.2 ± 7.8) compared to survivors (10.6 ± 4.3, P < 0.001), corroborating its prognostic value, though with lower discriminative ability than the ML model.

Figure 5 Calibration plots for mortality prediction models. A: Random Forest; B: Model for end-stage liver disease-serum sodium; C: Model for end-stage liver disease; D: Child-Pugh; E: Glasgow-Blatchford scores. Calibration curves (solid lines) compare predicted probabilities (X-axis) with observed mortality rates (Y-axis). Perfect calibration is shown by the diagonal dashed line. Histograms display the distribution of predictions. The random forest model shows excellent calibration (Brier score = 0.067), while traditional scores show systematically poorer calibration (Brier scores 0.199-0.247). n = 97 patients from the retrospective cohort. MELD: Model for end-stage liver disease; MELD-Na: Model for end-stage liver disease-serum sodium.

Table 4 Performance comparison of the random forest model vs traditional clinical scoring systems for predicting 1-year mortality in patients with cirrhosis with acute esophageal variceal bleeding (retrospective cohort, n = 97).

Model	AUC (95%CI)	Sensitivity	Specificity	Accuracy	Brier score
Random forest	0.915 (0.856-0.961)	0.800	0.860	0.830	0.124
MELD-Na	0.742 (0.651-0.823)	0.688	0.717	0.702	0.186
MELD	0.726 (0.634-0.809)	0.667	0.696	0.681	0.194
Child-Pugh	0.685 (0.591-0.771)	0.625	0.674	0.649	0.217
Glasgow-Blatchford	0.598 (0.502-0.690)	0.542	0.609	0.574	0.248

A Brier score ranges from 0 to 1, with lower values indicating better calibration (0 = perfect calibration). Random forest model used 100 trees, bootstrap aggregation, and isotonic calibration. P < 0.001 compared to random forest model (DeLong test). All metrics except area under the receiver operating characteristic curve were calculated at the optimal threshold determined by maximizing the Youden index (sensitivity + specificity - 1). Confidence intervals (CIs) for area under the receiver operating characteristic curve were calculated using 1000 bootstrap iterations. AUC: Area under the curve; MELD: Model for end-stage liver disease; MELD-Na: MELD-sodium.

Among traditional scores, Child-Pugh showed the best performance (AUC 0.764, 95%CI: 0.667-0.861), followed closely by MELD-Na (AUC 0.762, 95%CI: 0.665-0.859) and MELD (AUC 0.752, 95%CI: 0.654-0.850). The Glasgow-Blatchford score, designed for general upper gastrointestinal bleeding rather than specifically for variceal hemorrhage, showed the poorest discrimination (AUC 0.635, 95%CI: 0.527-0.743).

Calibration analysis (Figure 5) demonstrated superior performance of the random forest model with a Brier score of 0.067, indicating high agreement between predicted probabilities and observed outcomes. Traditional scores showed systematically poorer calibration, with Brier scores ranging from 0.199 (Child-Pugh) to 0.247 (Glasgow-Blatchford).

DISCUSSION

AEVB in patients with cirrhosis remains a critical clinical challenge due to its association with high mortality rates and significant healthcare burden. In this study, we developed and validated an ML model for predicting mortality in patients with AEVB, leveraging extensive clinical data from a tertiary care center. Our findings underscore the potential of ML in enhancing prognostication accuracy beyond traditional scoring systems. The prognostication of AEVB has remained challenging due to the complexity associated with liver diseases. It has been demonstrated that liver-specific scores outperform scores developed for gastrointestinal bleeding in this scenario[7], and that the presence of acute-on-chronic liver failure also predicts mortality[3]. It is expected therefore, that a ML model might outperform these scores, integrating variables associated with the hemorrhage and with liver function[14,20]. This has been true for other areas of hepatology. For example, when developing a ML model for prognosticating cardiac events post-liver transplantation, the association of cardiologic and hepatic variables outperformed generally used scores and clinical models[13]. Therefore, it is expected that ML models, if integrated into hospital systems, might improve the prognostication of liver transplantation and liver diseases in general[12].

The ML model developed in this study exhibited robust performance in predicting mortality among patients with cirrhosis presenting AEVB. Among the four supervised ML algorithms evaluated, the random forest model emerged as the most effective, achieving a sensitivity of 80%, specificity of 86%, and overall accuracy of 83%. Internal validation of the random forest model demonstrated a mean AUC of 0.715, with favorable metrics including recall (75.2%) and F1 score (65.7%). Prospective validation further substantiated its performance, revealing a higher mean AUC of 0.927, accuracy of 82.9%, recall of 86.7%, and F1 score of 76.0%.

These findings are compatible with previous studies that developed similar ML models for this scenario. For example, an automated multimodal ML integrating endoscopic images and clinical data, achieving an impressive accuracy of 93.2% and superior metrics such as sensitivity (95.2%) and F1 score (87.9%) with their stacking model. Their approach leveraged deep learning models like EfficientNet for image analysis, highlighting the potential of combining advanced imaging techniques with structured clinical data to enhance prediction accuracy[21]. Another group developed a contrast-enhanced computed tomography-based longitudinal section model, demonstrating an AUC of 0.789 in predicting clinical endpoints in acute variceal bleeding, surpassing traditional clinical scores[22]. While our study underscores the efficacy of random forests in mortality prediction with strong accuracy and validation metrics, these studies showcase the evolving landscape of multimodal ML and radiomics in refining prognostic models for AEVB.

Recent studies have advanced ML applications towards the goal of detecting the risk for patients with cirrhosis developing AEVB. A novel radiomic model utilizing contrast-enhanced computed tomography images to diagnose high bleeding risk esophageal varices in patients with cirrhosis. Their radiomic model achieved impressive diagnostic accuracy with AUC operating characteristic curves of 0.983 and 0.834 in training and internal validation, respectively, and demonstrated robust sensitivity and specificity metrics. External validation further confirmed its efficacy, albeit with slightly reduced performance, underscoring the potential of radiomics in non-invasive risk assessment for AEVB[23]. Similarly, an artificial neural network was able to predict the 1-year risk of esophagogastric variceal bleeding in patients with cirrhosis, achieving an AUC of 0.959. This model outperformed traditional clinical indices like the North Italian Endoscopic Club and revised North Italian Endoscopic Club indices, emphasizing its utility in refining risk-based surveillance strategies[24]. These studies collectively highlight the evolving landscape of ML in enhancing predictive accuracy and clinical decision-making for AEVB, offering promising avenues for personalized patient management.

A recent systematic review underscores the transformative potential of ML in this domain[20]. Their comprehensive analysis of twelve studies highlights the robust predictive capabilities of ML, with various models achieving exceptional accuracy, often surpassing traditional methods with AUC values exceeding 99%. Despite these promising outcomes, the review identified significant methodological heterogeneity and potential biases across studies, such as variability in input variables and outcome measures, and unclear handling of missing data. These findings emphasize the need for standardized protocols and larger, prospective trials to validate ML models’ clinical utility in predicting AEVB risk effectively. Moreover, ML’s capacity to potentially reduce reliance on invasive procedures like endoscopy underscores its role in advancing personalized patient care and optimizing resource allocation in cirrhosis management[20]. Future research efforts should prioritize refining ML methodologies and addressing methodological inconsistencies to enhance the reliability and applicability of these predictive models in clinical practice and its integration into medical electronic records.

These advancements in ML models in this scenario have significantly enhanced predictive models for managing esophageal varices and related complications in patients with liver diseases. For instance, recently developed the EVendo score using a random forest algorithm, achieving high accuracy in identifying varices needing treatment among patients with cirrhosis (AUC operating characteristic curves 0.82 in validation). This score, incorporating readily available clinical data, demonstrated potential to defer unnecessary esophagogastroduodenoscopy screenings by 30.5%, while missing only 2.8% of varices needing treatment cases[25], which was further validated in patients with hepatocellular carcinoma[26]. Similarly, another group applied extreme-gradient boosting to predict variceal bleeding in compensated advanced chronic liver disease, achieving high predictive accuracy (98.7% in derivation, 93.7% in internal validation). Their model outperformed traditional endoscopic classifications, highlighting its utility in risk stratification and subsequent management decisions[27]. Another group demonstrated that their ML model surpassed traditional prognostic scores like the MELD-Na and Child-Turcotte-Pugh scores, achieving higher AUCs (0.91 for 1-month, 0.88 for 3-month, and 0.91 for 12-month survival) in predicting outcomes in patients with variceal bleeding[28]. Also, ML models, particularly the random forest and Vapnik-Chervonenkis classifiers, have surpassed the Glasgow-Blatchford score and Rockall scores in predicting adverse events (AUC 0.917 for mortality, 0.757 for hypotension, and 0.733 for rebleeding within 7 days) in patients with stable non-variceal bleeding[29]. These findings collectively underscore the superiority of ML-based approaches over conventional scoring systems, offering more accurate risk assessment and potentially reducing unnecessary procedures and healthcare costs.

Our SHapley Additive exPlanations analysis identified race as the most influential predictor in the model, which warrants careful consideration from both clinical and ethical perspectives. This finding may reflect complex interactions between genetic factors, social determinants of health, healthcare access disparities, and disease phenotypes that vary across populations. While race served as a proxy for these unmeasured confounders in our model, we acknowledge the potential risks of perpetuating healthcare disparities through race-based predictions.

Future iterations of this model should explore race-neutral alternatives by incorporating more granular social determinants of health, genetic markers, and healthcare access variables. We recommend that clinicians interpret the model’s predictions within the broader context of individual patient circumstances rather than allowing race to disproportionately influence clinical decision-making. Development of race-neutral models that maintain predictive accuracy while promoting equitable care remains a priority for future research.

This study had two major limitations. First, our study focused on a single-center cohort, which may limit the generalizability of our findings to broader patient populations. to confirm the model’s generalizability across different clinical settings and patient demographics. Moreover, external validation in multicenter studies would further validate the utility of our ML model in diverse healthcare environments. Second, the predictive features identified by our random forest model highlight critical biomarkers and clinical parameters associated with mortality risk in patients with AEVB. These include race, liver function tests (total bilirubin, INR, albumin), renal function (creatinine), and comorbid conditions (hepatorenal syndrome, ascites). However, our study did not explore dynamic changes in these biomarkers over time, which could provide insights into disease progression and treatment response. Future studies incorporating longitudinal data and real-time monitoring could enhance the model’s predictive accuracy and clinical utility.

The deployment of our model as a web-based calculator (https://huggingface.co/spaces/mmrech/evb-br) facilitates clinical integration, though several considerations warrant attention. The tool’s interface was designed for ease of use, requiring only routinely collected clinical variables. However, successful implementation requires clinical validation in local settings, integration with electronic health records, and training for healthcare providers.

Future development should focus on real-time integration with hospital information systems, automated data extraction to minimize manual entry errors, and prospective monitoring of model performance to detect drift. We recommend that institutions interested in implementing this tool conduct local validation studies and establish protocols for model updates as new data become available. The model should augment, not replace, clinical judgment, serving as one component of comprehensive risk assessment in AEVB management.

Nevertheless, we must consider that the interpretability of ML models remains a critical concern in clinical settings. While feature importance analysis provided insights into predictive biomarkers, the black-box nature of some ML algorithms may hinder clinical adoption. Future research should focus on developing interpretable ML models that align with clinical decision-making processes and facilitate transparent communication of risk predictions to healthcare providers and patients.

CONCLUSION

In conclusion, our study demonstrates the potential of ML in advancing prognostication for patients with AEVB, offering personalized risk assessment and informing timely interventions. Moving forward, multicenter validation studies, integration of real-time data streams, and development of interpretable ML models are crucial steps toward translating research findings into clinical practice. By addressing these challenges, ML has the potential to revolutionize risk prediction and improve outcomes in patients with cirrhosis with AEVB.

ACKNOWLEDGEMENTS

We would like to express our deepest gratitude to Ana Laura Facco Muscope, Morgana Luisa Longen, Bruna Schena, Bruna Teston Cini, Gilberto Luis Rost Jr, Juline Isabel Leichtweis Balensiefer, and Louise Zanotto Eberhardt for their invaluable contributions to data collection, which were instrumental to the success of this research. Our heartfelt thanks also go to Raul Angelo Balbinot and Silvana Sartori Balbinot for their unwavering support and encouragement throughout the research process. Your assistance and dedication have been deeply appreciated and have greatly enriched this work.