📄 medicine

Development and Validation of a Predictive Model for 6 Month Rebleeding in Cirrhotic Patients with Esophagogastric Variceal Bleeding

This study developed and validated a high-performing XGBoost predictive model, utilizing five core clinical predictors identified via LASSO regression, to accurately assess the 6-month rebleeding risk in cirrhotic patients with esophagogastric variceal bleeding.

Original authors: Jisheng Gu, Yanru Li, Xinyue Ma, Pengfei Zhou, Sujuan Fei

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jisheng Gu, Yanru Li, Xinyue Ma, Pengfei Zhou, Sujuan Fei

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the liver as a busy, aging city. Over time, due to issues like chronic hepatitis or alcohol use, the city's roads (blood vessels) get clogged and damaged, causing traffic jams. This is called cirrhosis. Because the roads are blocked, the pressure builds up in the main highway leading into the city (the portal vein). Eventually, the pressure gets so high that the weaker side streets (varices) burst open, causing a massive flood of blood. This is esophagogastric variceal bleeding.

Doctors are very good at stopping the flood immediately (hemostasis), but the real challenge is predicting which cities are likely to have a second flood within the next six months.

Here is how the researchers at Xuzhou Medical University built a new "weather forecast" for this specific problem, explained simply:

1. The Goal: A Better Crystal Ball

For a long time, doctors have used old, general maps (like the Child-Pugh or MELD scores) to guess if a patient will bleed again. The researchers felt these maps were a bit blurry—they couldn't tell the difference between a city that was safe and one that was about to flood. They wanted to build a high-tech, 6-month weather forecast specifically for cirrhosis patients who have just had a bleed.

2. Gathering the Data: The Detective Work

The team looked back at the medical records of 308 patients from their hospital who had this type of bleeding between 2019 and 2025.

They split these patients into two groups: a Training Group (216 people) to teach the computer, and a Test Group (92 people) to see if the computer actually learned anything.
They checked who had a second bleed within 6 months (118 people) and who didn't (190 people).

3. Finding the Clues: The "Five Fingerprints"

The researchers started with a huge pile of clues: age, blood tests, ultrasound images, and medical history. It was like having a bag full of 50 different puzzle pieces. To find the right pieces, they used a smart digital filter called LASSO regression.

Think of LASSO as a strict editor who says, "We don't need all these words; let's keep only the ones that actually tell the story." This process whittled the list down to just five key indicators that mattered most:

Ascites: Is there fluid building up in the belly? (Yes = higher risk).
Portal Vein Diameter: How wide is the main highway? (Wider = higher risk).
Albumin-to-Bilirubin Ratio: A score measuring how well the liver is working and cleaning toxins. (Lower score = higher risk).
Prothrombin Time: How long it takes for blood to clot. (Slower clotting = higher risk).
Fibrinogen: A specific protein needed to form clots. (Lower levels = higher risk).

4. Building the Engine: The "Smart Coach"

The team didn't just use one way to predict the future. They built five different computer models (like five different coaches trying to predict the game outcome):

Logistic Regression (The traditional coach)
Support Vector Machine (The geometric coach)
Random Forest (The team of decision-makers)
Gradient Boosting (The iterative coach)
XGBoost (The super-advanced, high-speed coach)

They let all five coaches practice on the Training Group, then tested them on the unseen Test Group.

The Winner: The XGBoost model was the clear champion.

It was the most accurate at distinguishing between patients who would bleed again and those who wouldn't.
It didn't just guess; it was "calibrated," meaning if it said there was a 70% chance of bleeding, it was right about 70% of the time.
It outperformed the other four models and the old traditional methods.

5. Making Sense of the "Black Box"

Usually, advanced computer models are like a "black box"—you put data in, and a result comes out, but you don't know why. To fix this, the researchers used a tool called SHAP.

The Analogy: Imagine the model is a mystery movie. SHAP is the detective who explains exactly why the suspect was caught. It showed that the width of the portal vein and the presence of ascites were the biggest "villains" driving the risk of rebleeding.

6. The Final Product: A Web-Based Calculator

To make this useful for real doctors, they turned the winning XGBoost model into a web-based calculator.

A doctor can type in those five simple numbers (fluid in belly, vein width, etc.).
The website instantly calculates the patient's specific risk of bleeding again in the next six months.
This helps doctors decide who needs extra care and who can go home with standard follow-up.

Summary

The researchers created a new, highly accurate tool to predict if a cirrhosis patient will have a second bleed within six months. By using a smart computer algorithm (XGBoost) and focusing on just five easy-to-measure factors, they built a system that is more accurate than previous methods. They also made sure the system is transparent (doctors know why it made a prediction) and accessible (it's a website anyone can use).

Important Note: The study was done at a single hospital in China. While the results are promising, the researchers acknowledge that the tool needs to be tested in other hospitals and with different groups of people before it becomes a standard rule everywhere.

Technical Summary: Development and Validation of a Predictive Model for 6-Month Rebleeding in Cirrhotic Patients with Esophagogastric Variceal Bleeding

Problem Statement
Esophagogastric variceal bleeding (EGVB) is a life-threatening complication of cirrhosis with high mortality and recurrence rates. Despite advances in acute hemostasis, mid-to-long-term rebleeding remains a significant clinical challenge. Current risk stratification relies on conventional scores (e.g., Child-Pugh, MELD, ALBI) that primarily reflect hepatic reserve but lack the sensitivity and accuracy to precisely predict rebleeding risk, leading to either overtreatment of low-risk patients or missed high-risk cases. Furthermore, existing machine learning (ML) studies in this domain are often limited by small sample sizes, non-rigorous variable selection, lack of validation, and poor interpretability. This study addresses the need for a well-calibrated, interpretable, and practical tool to predict the risk of rebleeding within 6 months—a critical transitional window where acute factors subside and chronic portal hypertension becomes the dominant determinant of risk.

Methodology
The study employed a retrospective cohort design involving 308 patients with cirrhosis and EGVB admitted to the Affiliated Hospital of Xuzhou Medical University between January 2019 and June 2025.

Data Processing: Patients were randomly stratified into a training set ( $n=216$ ) and an internal validation set ( $n=92$ ) at a 7:3 ratio. Missing data were handled using multiple imputation by chained equations (MICE) with a random forest algorithm.
Feature Selection: A two-step screening process was utilized. First, univariate logistic regression identified variables with $P < 0.05$ . Second, Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation was applied to the training set to eliminate redundancy, handle multicollinearity, and select core predictors.
Model Construction: Five machine learning algorithms were constructed and compared: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Machine (GBM), and Extreme Gradient Boosting (XGBoost). Hyperparameters were tuned using nested 10-fold cross-validation.
Evaluation Metrics: Model performance was assessed using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) for discrimination. Calibration was evaluated via calibration curves and Brier scores. Clinical utility was assessed using Decision Curve Analysis (DCA).
Interpretability: The optimal model was interpreted using SHapley Additive exPlanations (SHAP) to quantify variable contributions and visualize nonlinear relationships.
Deployment: A web-based calculator was developed using the R shiny package to visualize risk predictions.

Key Contributions

Identification of Core Predictors: The study identified five clinically accessible and pathophysiologically relevant predictors: ascites status, portal vein diameter, albumin-to-bilirubin (ALBI) ratio, prothrombin time (PT), and fibrinogen.
Model Optimization: The XGBoost algorithm demonstrated superior performance compared to LR, SVM, RF, and GBM.
Interpretability Framework: The integration of SHAP analysis addressed the "black-box" nature of ML, providing clear insights into variable importance and directional effects.
Clinical Tool: The development of a web-based risk calculator translates the complex algorithm into a user-friendly tool for individualized risk assessment.

Results

Variable Selection: LASSO regression successfully reduced the initial pool of variables to the five core predictors. Ascites, larger portal vein diameter, and prolonged PT were identified as risk factors, while higher ALBI ratio and fibrinogen levels were protective.
Discrimination: In the training set, the XGBoost model achieved the highest AUC of 0.929 (95% CI: 0.892–0.965). In the internal validation set, it maintained robust performance with an AUC of 0.855 (95% CI: 0.780–0.929), outperforming all other models (LR: 0.844, SVM: 0.847, RF: 0.848, GBM: 0.853).
Calibration and Utility: The XGBoost model exhibited excellent calibration with the lowest Brier scores (0.097 in training, 0.165 in validation). Decision Curve Analysis confirmed that the XGBoost model provided a higher net clinical benefit across a wide range of threshold probabilities compared to alternative models.
SHAP Analysis: SHAP values indicated that portal vein diameter and ascites were the most influential predictors, revealing nonlinear relationships between these variables and rebleeding risk.

Significance and Claims
The authors claim that this study successfully developed and validated a LASSO-derived XGBoost model that effectively predicts 6-month rebleeding risk in cirrhotic patients with EGVB. The model is characterized by excellent discrimination, satisfactory calibration, and favorable clinical utility. By focusing on the 6-month interval, the study aims to fill a gap between acute management (6-week outcomes) and long-term surveillance (1-year outcomes), providing a clinically actionable window for secondary prevention.

The paper emphasizes that the selected predictors are not only statistically significant but also pathophysiologically meaningful, capturing the interplay between portal hypertension severity, hepatic functional reserve, and coagulation dysfunction. The authors assert that the integration of SHAP analysis and the deployment of a web-based calculator enhance the model's transparency and practicality, offering clinicians a reliable tool for early identification of high-risk patients and individualized management.

The study acknowledges limitations, including its retrospective, single-center nature and the exclusion of certain hemodynamic variables (e.g., HVPG) due to data availability. Consequently, the authors modestly conclude that while the model shows promise for risk-stratified clinical management, multicenter prospective studies are required to further validate its generalizability before widespread clinical implementation.