HTH-5 Could machine learning (ML) improve indices for predicting outcome of AUGIB?

IntroductionRisk stratification scores for acute upper gastrointestinal bleeding (AUGIB) have limitations. ML models using multiple variables have the potential to improve predictive value.MethodsPatients undergoing at least one endoscopy were selected from the 2007 UK upper gastrointestinal bleeding (UGIB) audit1 for developing ML models predicting 28-day all-cause mortality and rebleeding. Input variables were divided into pre-endoscopy & endoscopy features (demographics, presentation, co-morbidity, concomitant drugs, biochemical parameters, pre-endoscopic management and first endoscopy findings). Random forest with 100 estimators or trees was used as a prediction model. Shapely additive explanations (SHAP)2 identified the most relevant features in each model. MissForest3 imputed missing data. Outcome prediction from ML models for pre-endoscopy and combined pre-endoscopy & endoscopy were compared to Rockall scores. 5-fold cross-validation compared performance, reported as average areas under receiver operating characteristic curve (AUC).Results5004 patients were included (mean age 66yr, 60% male). 28-day all-cause mortality was 6.7% (n=335) with 9.3% (n=465) rebleeding. All-cause mortality predicted by ML was independent of endoscopy (AUC 0.84, 95% CI 0.82-0.85, pre-endoscopy; 0.84, 95% CI 0.82-0.85, pre- & endoscopy) vs 0.76, 95% CI:0.75-0.77 for Rockall. Rebleeding predicted by ML was enhanced by endoscopy (AUC 0.64, 95% CI: 0.63-0.66 pre-endoscopy; AUC 0.75, 95% CI:0.75-0.77 pre- & endoscopy) vs AUC 0.67, 95% CI:0.66-0.68 for Rockall. The SHAP tool reported 10 most relevant features (Figure 1).Abstract HTH-5 Figure 1ConclusionPre-endoscopy ML models performed better than Rockall scores for predicting mortality. Endoscopy appeared to enhance ML prediction of rebleeding. Comparison with other risk scores was not possible from the dataset, but ML models from electronic records could develop dynamic decision support tools.ReferencesHearnshaw et al Gut 2011.Lundberg et al Nat Mach Intell 2020.Stekhoven et al Bioinformatics 2012.