Full Oral

Aim: Multiple clinical risk scores have been developed for acute upper gastrointestinal bleeding (AUGIB) with limited sensitivity and specificity.We aim to develop a machine learning model to identify high risk AUGIB patients associated with poor outcomes and to compare its performance with existing risk scores.Methods: A population-wide database was developed with all AUGIB patients in 6 major hospitals in Hong Kong.The study endpoint is a composite endpoint of poor outcomes including mortality (all-cause mortality and bleeding-related mortality), rebleeding within 30 days, need for ICU, re-endoscopy, surgery and embolization.15 factors were included.Machine learningbased predictive models were developed to estimate the probability of poor outcomes in patients with AUGIB.Several machine learning algorithms (decision trees, random forests, support vector machines and gradient boosting machines [GBM]) were trained.Model performance was assessed using AUROC, sensitivity and specificity.The results were validated with a 10-fold split and cross validation.Results: 6503 AUGIB patients were included, in whom 1500 patients were with poor outcomes.The AUROC, sensitivity and specificity of the model with GBM were 0.9233, 0.8313 and 0.8563, respectively.Moreover, the GBM model outperforms the existing clinical risk scores in terms of the AUROC, F1 score and specificity of GBM.The AUROC, F1, sensitivity, specificity of GBS ≥ 12 and Rockall Score ≥ 4 were 0.751, 0.497, 0.588, 0.747 and 0.714, 0.483, 0.744, 0.565, respectively.Conclusion: GBM model outperforms other clinical risk scores.This new model may replace existing clinical risk scores with higher accuracy.