Predicting Mortality in Non-Variceal Upper Gastrointestinal Bleeding: Machine Learning Models Versus Conventional...

Area of Science:

• Emergency Medicine
• Clinical Informatics
• Data Science in Healthcare

Background:

• Non-variceal upper gastrointestinal bleeding (NVUGIB) presents significant mortality risks, especially in emergency settings.
• Current risk stratification scores (GBS, AIMS65, Rockall) have limited accuracy for predicting mortality in NVUGIB patients.
• There is a need for more precise tools to identify high-risk NVUGIB patients.

Purpose of the Study:

• To evaluate the predictive performance of supervised machine learning (ML) models for 30-day mortality in NVUGIB.
• To compare the accuracy of ML models against established clinical risk scores.
• To identify key predictors of mortality in NVUGIB using ML.

Main Methods:

• Retrospective cohort study of 1233 NVUGIB patients presenting to an emergency department.
• Training and evaluation of seven supervised ML algorithms with six feature selection techniques.
• Performance assessment using AUROC, F1-score, sensitivity, specificity, and calibration, compared to GBS, AIMS65, and Rockall scores.

Main Results:

• The best ML model (XGBoost) achieved an AUROC > 0.80 and F1-score of 0.909, significantly outperforming traditional scores (highest Rockall AUROC: 0.743).
• ML models demonstrated superior sensitivity, specificity, and calibration compared to conventional methods.
• Key predictors included age, comorbidities, hemodynamic status, and laboratory markers.

Conclusions:

• Supervised ML models show markedly superior apparent discrimination for predicting 30-day mortality in NVUGIB compared to traditional scores.
• These findings highlight the potential of ML for improving risk stratification in NVUGIB.
• External multicenter validation is crucial before clinical implementation of these ML models.