Deng F, Cao Y, Zhao S. Upper gastrointestinal haemorrhage patients' survival: A causal inference and prediction study.
Eur J Clin Invest 2024;
54:e14180. [PMID:
38376066 DOI:
10.1111/eci.14180]
[Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Key Words] [MESH Headings] [Track Full Text] [Journal Information] [Submit a Manuscript] [Subscribe] [Scholar Register] [Received: 07/23/2023] [Revised: 01/11/2024] [Accepted: 02/03/2024] [Indexed: 02/21/2024]
Abstract
BACKGROUND
Upper gastrointestinal (GI) bleeding is a common medical emergency. This study aimed to develop models to predict critically ill patients with upper GI bleeding in-hospital and 30-day survival, identify the correlation factor and infer the causality.
METHODS
A total of 2898 patients with upper GI bleeding were included from the Medical Information Mart for Intensive Care-IV and eICU-Collaborative Research Database, respectively. To identify the most critical factors contributing to the prognostic model, we used SHAP (SHapley Additive exPlanations) for machine learning interpretability. We performed causal inference using inverse probability weighting for survival-associated prognostic factors.
RESULTS
The optimal model using the light GBM (gradient boosting algorithm) algorithm achieved an AUC of .93 for in-hospital survival, .81 for 30-day survival in internal testing and .87 for in-hospital survival in external testing. Important factors for in-hospital survival, according to SHAP, were SOFA (Sequential organ failure assessment score), GCS (Glasgow coma scale) motor score and length of stay in ICU (Intensive critical care). In contrast, essential factors for 30-day survival were SOFA, length of stay in ICU, total bilirubin and GCS verbal score. Our model showed improved performance compared to SOFA alone.
CONCLUSIONS
Our interpretable machine learning model for predicting in-hospital and 30-day mortality in critically ill patients with upper gastrointestinal bleeding showed excellent accuracy and high generalizability. This model can assist clinicians in managing these patients to improve the discrimination of high-risk patients.
Collapse