Machine learning models for predicting active bleeding in patients presenting to the emergency department with upper gastrointestinal bleeding: a retrospective cross-sectional study
Retrospective cross-sectional study reveals machine learning models predict active upper gastrointestinal bleeding, indicating potential to safely rule out urgent endoscopy.
Key Points
To develop and internally validate supervised machine learning models using routine admission data to predict endoscopically confirmed active upper gastrointestinal bleeding in emergency department patients.
Retrospective observational cross-sectional study analyzing 1,009 adult patients presenting with suspected upper gastrointestinal bleeding who underwent endoscopy at a tertiary academic hospital.
Patients were divided into a training cohort (N=666) and an internal validation cohort (N=343) to evaluate five machine learning algorithms against the Glasgow–Blatchford Score using clinical, laboratory, and blood gas variables.
In the test cohort, 8.7% (30/343) exhibited active bleeding (Forrest Ia/Ib), presenting with significantly higher Glasgow–Blatchford scores (13.07 ± 1.93 vs. 12.19 ± 2.36; p = 0.039) and lower platelet counts (224.37 ± 95.58 vs. 267.65 ± 118.78 × 10³/µL; p = 0.049).
Random Forest yielded the highest discrimination with an AUC of 71.2% (95% CI, 60.4–81.9%), sensitivity of 80.0% (95% CI, 62.7–90.5%), and negative predictive value of 97.0% (95% CI, 93.6–98.6%), outperforming the Glasgow–Blatchford Score at cutoff ≥ 13 (AUC 61.3% [95% CI, 50.2–72.4%], sensitivity 66.7%, NPV 94.4%).