Causal Effects of Endoscopy Timing on Outcomes in Critically Ill Patients With Upper Gastrointestinal Bleeding: A Double Machine Learning Analysis.

backgroundWhen endoscopy should be performed in critically ill patients presenting with upper gastrointestinal bleeding is still a matter of debate. Although early intervention is broadly endorsed in guidelines, robust causal data demonstrating a survival advantage specifically within ICU populations remain sparse.

methodsThis is a retrospective cohort analysis drawing on the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Eligible participants were adult ICU patients (aged ≥18 years) admitted with upper gastrointestinal bleeding who received endoscopy during the first 72 hours following ICU admission. We defined the principal exposure as early endoscopy, dichotomized at ≤12 hours versus >12 hours. In-hospital death served as the main endpoint, while 30-day mortality, recurrent bleeding, and need for blood transfusion were examined as secondary endpoints. To derive causal estimates of treatment effect, we applied double machine learning with cross-fitting, controlling for 11 baseline confounders spanning age, vital signs, laboratory parameters, and comorbid conditions. We additionally evaluated whether treatment effects varied across patients and performed a broad set of sensitivity analyses using different time cutoffs, endpoints, and patient subgroups.

resultsThe cohort comprised 584 ICU patients with upper gastrointestinal bleeding (mean age 64.4 years; 59.8% underwent early endoscopy), with an overall death rate of 9.8%. The double machine learning analysis identified no meaningful impact of early endoscopy on in-hospital mortality (average treatment effect (ATE): +1.45 percentage points; 95% CI: -3.08 to +5.99; p=0.530). Sensitivity analyses, however, surfaced worrying safety findings: early endoscopy correlated with a significant rise in both recurrent bleeding (+6.18 percentage points; 95% CI: +1.75 to +10.61; p=0.006) and transfusion needs (+9.77 percentage points; 95% CI: +3.78 to +15.76; p=0.001). Analysis of effect heterogeneity revealed a notable dependence on age: patients at or above the median age fared worse with early endoscopy, showing elevated mortality (+7.77 percentage points; 95% CI: +0.44 to +15.10; p=0.038), whereas younger patients trended toward benefit. A post-hoc power calculation indicated that the study had merely 9% power to detect the observed mortality effect, with roughly 29,000 patients needed to reach 80% power.

conclusionsThis hypothesis-generating analysis found that early endoscopy in critically ill patients with upper gastrointestinal bleeding conferred no significant survival benefit and was instead linked to greater rates of rebleeding and transfusion. The harm seen among older patients indicates that a blanket policy of early endoscopy may be unsuitable across the entire ICU population. These results call into question prevailing clinical conventions and underscore the importance of risk-stratified strategies and sufficiently powered prospective trials.