Development and validation of machine learning models for 30-day mortality prediction in septic shock

Abstract

Background Septic shock carries a high risk of death, yet traditional prognostic tools have limitations. This study aimed to develop and validate machine learning (ML) models for predicting 30-day mortality in septic shock patients using comprehensive admission variables. Methods A retrospective analysis was performed on 695 patients with septic shock. Demographic characteristics, vital signs, and 33 laboratory parameters were included as candidate predictors. LASSO regression was used for feature selection from age, gender and 33 laboratory parameters. Eleven ML algorithms were developed using a two-way split with internal cross-validation (60% training with 5-fold CV for hyperparameter tuning, 40% held-out validation cohort for final evaluation) and their performances were evaluated using AUC with 95% confidence intervals (CIs), sensitivity, specificity, calibration curves, and decision curve analysis. SHAP method was applied for model interpretation. Results LASSO selected six predictors: lactate, platelet count, neutrophil percentage, monocyte percentage, albumin, and C-reactive protein. Following rigorous hyperparameter tuning, the Random Forest (RF) model demonstrated the best and most stable performance in the validation cohort (AUC = 0.820, 95% CI: 0.765–0.875). The RF model demonstrated good calibration and net clinical benefit. A web-based risk calculator was subsequently developed. Conclusion The RF model based on six readily available laboratory parameters accurately predicts 30-day mortality in septic shock patients. However, the findings are limited by the single-center retrospective design and lack of external validation, warranting further multi-center studies to confirm generalizability.