Establishment and external validation of a prediction model for 90-day mortality in ICU patients with liver cirrhosis complicated by esophagogastric variceal bleeding

Patients with liver cirrhosis complicated by esophagogastric variceal bleeding (EVB) exhibit considerable prognostic variability, highlighting the need for precise predictive tools. Conventional critical illness scores (e.g., SOFA, SAPSII) are not tailored for this specific cohort. This study aims to develop a dedicated prognostic model for EVB, compare it internally with traditional scores, and rigorously assess its generalizability in an independent external cohort. A total of 473 patients with liver cirrhosis and EVB were retrieved from the MIMIC-IV database and randomly assigned to a training set and an internal validation set in a 7:3 ratio. A model predicting 90-day all-cause mortality was developed through feature selection. The new model was compared against the LODS, SOFA, OASIS, SAPSII, APSIII, and MELD scores. An external validation cohort of 76 patients from the ICU of the Affiliated Hospital of Southwest Medical University was used for further testing. The final model incorporated activated partial thromboplastin time (APTT), total bilirubin (TB), and neutrophil count-to-albumin ratio (NAR). The model exhibited discrimination comparable to SAPSII (AUC = 0.80), APS-III (AUC = 0.80), MELD (AUC = 0.77), and superior to SOFA (AUC = 0.75), LODS (AUC = 0.76), and OASIS (AUC = 0.63) in the training dataset (AUC = 0.81, 95% CI: 0.76–0.86). In the internal validation set, the model demonstrated excellent discrimination (AUC = 0.827). In the external validation set, the AUC of the model decreased to 0.715. However, the calibration curve demonstrated high consistency between predicted and observed probabilities (maximum bias < 0.01), and DCA showed that the net benefit of the model surpassed those of the “treat-all” or “treat-none” strategies. The TB-APTT-NAR model is a simple and potentially useful tool for predicting mortality risk in ICU patients with cirrhosis complicated by EVB. It demonstrates excellent calibration and clinical utility, along with moderate discriminative ability. However, given the limited external validation sample size, the model should currently be considered a preliminary risk stratification aid rather than a definitive clinical decision model. Future multicenter studies with larger sample sizes are warranted to further validate its generalizability.