Bayesian assessment of pre-endoscopic risk scores in upper gastrointestinal bleeding

Several pre-endoscopic scoring systems are used to stratify risk in upper gastrointestinal bleeding (UGIB), yet none is ideal. Traditional metrics such as sensitivity, specificity, or AUROC do not provide individualized estimates. A Bayesian approach may offer more clinically relevant insights, but data in UGIB are limited.

We conducted a retrospective analysis of a prospective registry of patients admitted with UGIB to a tertiary hospital between 2013-2024. Pre-test probabilities for intervention and mortality were calculated, and score performance was assessed using likelihood ratios to estimate post-test probabilities via a Bayesian approach. Sensitivity analyses were performed.

A total of 1,649 patients were included (mean age 64.4 years, 31.9% female). Intervention and mortality rates were 74.3% and 10.5%. MAP(ASH), ABC, and AIMS65 increased post-test probability for intervention in high-risk patients but showed limited ability to exclude it in low-risk individuals. These scores performed better at ruling out mortality than predicting it. In contrast, GBS was the most effective tool for identifying low-risk patients for both outcomes. Sensitivity analyses confirmed consistent findings across outpatients and bleeding-related mortality.

A Bayesian approach provides a more clinically relevant interpretation of UGIB scores by estimating post-test probabilities. This approach allows clinicians to move beyond statistical metrics like AUROC and apply risk scores in a more individualized and actionable manner. While useful, current tools have limitations and should be applied cautiously. More accurate models are needed to improve individualized risk assessment and early decision-making.