Aims The need for endotherapy in acute upper gastrointestinal bleeding (AUGIB) is influenced
by pre-endoscopy and endoscopy findings. Identifying differences between patients
requiring endotherapy and determining predictors is crucial for improving care.
Methods We analysed data for non-variceal AUGIB patients from the 2022 UK audit1. Missing data were handled by imputation. Pre-endoscopy variables included demographics,
symptoms, vital signs, and laboratory parameters. Endoscopy findings included Forrest
classifications and lesion types. Baseline characteristics were compared between those
needing endotherapy and those not. Logistic regression (LR), LASSO, Random Forest
(RF), and XGBoost models predicted endotherapy, with performance assessed by AUC.
Results Of 2,386 patients, 34% (n=817) required endotherapy. Patients undergoing endotherapy
were older (71 vs. 68 years, p=0.008) and had lower systolic blood pressure (118 mmHg
vs. 123 mmHg, p<0.001). Melaena (68% vs. 57%, p<0.001) and shock (11% vs. 7%, p=0.003)
were more frequent in the endotherapy group, while coffee ground vomiting was less
common (14% vs. 26%, p<0.001), and haematemesis was similar (29% vs. 30%, p=0.7).
Haemoglobin was lower (85 g/L vs. 98 g/L, p<0.001) and urea higher (12 mmol/L vs.
10 mmol/L, p<0.001). Early RBC transfusion was common (52% vs. 34%, p<0.001). Endoscopic
findings associated with endotherapy included Forrest 1a (6.3% vs. 0%, p<0.001), Forrest
1b (44% vs. 30%, p<0.001), Forrest 2a (29% vs. 2.1%, p<0.001), gastric ulcers (22%
vs. 18%, p=0.014), duodenal ulcers (42% vs. 22%, p<0.001), and Dieulafoy lesions (4.7%
vs. 0.2%, p<0.001). Using pre-endoscopy variables, LR and LASSO achieved the highest
AUC (0.635), followed by RF (0.631) and XGBoost (0.607). Significant pre-endoscopy
predictors included urea (OR 1.02, 95% CI 1.01–1.03), and early RBC transfusion (OR
1.36, 95% CI 1.06–1.75). Including endoscopy findings improved discrimination: LASSO
achieved the highest AUC (0.809), followed by LR (0.807), RF (0.805), and XGBoost
(0.771). Key predictors included Forrest 1a (OR 11.64, 95% CI 5.28–25.65), Forrest
1b (OR 2.49, 95% CI 1.90–3.26), Forrest 2a (OR 8.21, 95% CI 5.85–11.51), Dieulafoy
lesions (OR 47.42, 95% CI 10.45–215.24), duodenal ulcers (OR 2.24, 95% CI 1.62–3.09),
Mallory-Weiss tears (OR 2.09, 95% CI 1.10–3.97) and telangiectasia (OR 2.84, 95% CI
1.49–5.41) [1].
Conclusions Patients needing endotherapy for non-variceal AUGIB had distinct clinical, laboratory,
and endoscopic features, including lower haemoglobin, higher urea, and Forrest 1a,
1b, and 2a lesions. Pre-endoscopy variables alone provided moderate discriminatory
ability (AUC ~0.64), while including endoscopy findings significantly improved prediction
(AUC ~0.81). LASSO and LR performed strongly and are suitable for clinical use. Future
work should develop risk scores for guiding interventions.
