Establishment and internal-external validation of a 28-day mortality prediction model for septic shock patients with...

Septic shock remains a major cause of death in the ICU with high 28-day mortality, and concomitant left ventricular systolic dysfunction (LVSD) further worsens patient outcomes, while existing prediction models lack adequate specificity. This study aimed to establish and perform internal-external validation of a 28-day mortality prediction model specifically for septic shock patients with LVSD as a key risk factor.

This retrospective multicenter cohort study enrolled 226 septic shock patients from September 2021 to October 2024 as the training cohort and 132 patients from November 2024 to December 2025 as the validation cohort, from ICUs of General Hospital and Cardiovascular Hospital of Ningxia Medical University. Baseline LVEF was measured at admission, with LVSD defined as LVEF < 50% or > 70%. Baseline data, clinical variables, myocardial biomarkers, and 28-day mortality were collected. Univariate and multivariate logistic regression identified independent predictors. A predictive nomogram was constructed and assessed using ROC, calibration curves, and DCA. Internal validation was performed via 1,000 bootstrap resamples, followed by external validation in the validation cohort.

Independent predictors of 28-day mortality in patients with septic shock were: left ventricular systolic dysfunction, decreased pH, atrial arrhythmia, dopamine use, and reduced PaO₂/FiO₂. Based on these predictors, a predictive model for 28-day mortality in patients with septic shock was constructed, and a nomogram was developed. The area under the ROC curve (AUC) of the model in the training cohort was 0.767 (95% CI: 0.703-0.831). After internal validation via Bootstrap sampling, the mean AUC was 0.779 (95% CI: 0.713-0.841). The calibration curve approached the ideal line (Hosmer-Lemeshow test P = 0.476), and the DCA indicated clinically net benefit within the 0.05-0.85 probability threshold range. External validation confirmed the reliability of the predictive model. Further comparison revealed that the predictive performance of this model was significantly superior to that of the APACHE II score for predicting mortality (0.767 vs. 0.652, P = 0.006). This nomogram has been converted into a web-based dynamic nomogram calculator available for free public use (https://linjia.shinyapps.io/dynnomapp/).

The 28-day mortality prediction model demonstrated excellent discriminative power, stability, and clinical applicability.