Validation of a Machine Learning Model That Outperforms Clinical Risk Scoring Systems for Upper Gastrointestinal...

Area of Science:

• Gastroenterology
• Medical Informatics
• Predictive Analytics

Background:

• Current scoring systems for upper gastrointestinal bleeding (UGIB) risk assessment are suboptimal.
• Machine learning (ML) offers potential for improved risk stratification in UGIB patients.

Purpose of the Study:

• To develop and validate an ML model for predicting the risk of hospital-based intervention or death in UGIB patients.
• To compare the performance of the ML model against established clinical risk scoring systems.

Main Methods:

• A gradient-boosting ML model was derived and internally validated using data from 1958 UGIB patients across 4 countries.
• The ML model's performance was compared with the Glasgow-Blatchford score (GBS), Rockall score, and AIMS65.
• External validation was performed using data from 399 UGIB patients in 2 Asia-Pacific sites.

Main Results:

• The ML model achieved an area under the receiver operating characteristic curve (AUC) of 0.91 (internal) and 0.90 (external validation).
• The ML model demonstrated superior performance compared to GBS (AUC 0.88 internal, 0.87 external), Rockall score, and AIMS65.
• At 100% sensitivity, the ML model achieved significantly higher specificity (26%) than GBS (12%).

Conclusions:

• A developed machine learning model accurately identifies patients with UGIB at risk of intervention or death.
• The ML model outperforms existing clinical scoring systems in predictive accuracy and specificity.
• This ML tool can enhance the identification of low-risk UGIB patients suitable for outpatient management.