New machine-learning models outperform conventional risk assessment tools in Gastrointestinal bleeding

Area of Science:

• Medical Informatics
• Clinical Prediction Models
• Gastroenterology

Background:

• Accurate identification of high-risk patients with acute gastrointestinal bleeding (GIB) is critical for timely intervention.
• Existing clinical scoring systems may have limitations in predicting in-hospital mortality for GIB patients.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting in-hospital mortality in patients admitted with overt GIB.
• To compare the performance of ML models against established clinical risk scores.

Main Methods:

• Analysis of a prospective, multicenter Hungarian GIB Registry dataset including 1,021 patients.
• Development and validation of XGBoost and CatBoost ML models.
• Comparison of ML models with Glasgow-Blatchford (GBS), pre-endoscopic Rockall, and ABC scores using five-fold cross-validation and AUC analysis.

Main Results:

• XGBoost and CatBoost models demonstrated superior predictive performance for in-hospital mortality compared to GBS and Rockall scores.
• Area Under the Curve (AUC) values for XGBoost and CatBoost were 0.84 and 0.77, respectively, significantly higher than GBS (0.68) and Rockall (0.62).
• The ABC score achieved an AUC of 0.77, still lower than the ML models.

Conclusions:

• XGBoost and CatBoost ML models offer improved accuracy in assessing mortality risk for acute GIB patients.
• These ML models provide a more effective tool for risk stratification than conventional clinical scoring systems.
• The findings support the integration of ML-based tools for enhanced clinical decision-making in GIB management.