Development and external validation of risk prediction models for gastrointestinal bleeding after aortic dissection surgery
Year: 2026
Type: article
Abstract: Background: Gastrointestinal bleeding (GIB) following surgery for aortic dissection (AD) is a life-threatening complication associated with elevated mortality and increased healthcare resource utilization. However, reliable predictive tools remain scarce. This study aimed to develop prediction models for postoperative GIB following AD surgery and to externally evaluate an XGBoost (XGB) model using an independent multicenter database. Methods: Patients who underwent AD surgery were identified from the Medical Information Mart for Intensive Care IV database (MIMIC-IV; development cohort) and the eICU Collaborative Research Database (eICU-CRD; external evaluation cohort). The primary outcome was postoperative GIB within 30 days, defined using a composite database-derived algorithm requiring either (I) a hemoglobin (Hb) decrease of ≥2 g/dL accompanied by at least one supporting indicator, including a transfusion record, hemostatic medication use, or endoscopic examination; or (II) the combination of a transfusion record and a GIB diagnosis. An isolated Hb decrease or transfusion record was not sufficient for outcome classification. Routinely available demographic, physiologic, laboratory, and disease-severity variables were extracted. Records with missing values in variables required for the final analysis were excluded before model development. Multivariable logistic regression served as the baseline model, with XGB and random forest employed as comparator models. Model performance was evaluated using the area under the curve (AUC), calibration curves, the Brier score, and decision-curve analysis (DCA). Results: The MIMIC-IV development cohort included 935 patients (GIB incidence, 10.7%), whereas the eICU-CRD validation cohort comprised 542 patients (GIB incidence, 12.7%). In the development cohort, the AUCs were 0.887 for logistic regression, 0.962 for XGB, and 0.853 for random forest. In the eICU-CRD cohort, XGB yielded an AUC point estimate of 0.873. Conclusions: This retrospective two-database analysis developed prediction models for postoperative GIB following AD surgery. The models showed discrimination in the analyzed data; however, prospective validation using predictors measured within prespecified pre-event windows is required before clinical implementation.
Source: Journal of Thoracic Disease
Institution: First Affiliated Hospital of Zhengzhou University
Language: English
Subfield: Cardiology and Cardiovascular Medicine
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7221808119