Development of an automated machine learning-based prognostic prediction model for emergency septic shock patients

Objective To overcome the limitations of static assessment, linear modeling, and lack of individualization in traditional scoring systems for prognostic evaluation of emergency septic shock patients, we developed an admission-based risk prediction model with real-time calculation capability to achieve precise stratification of 28-day all-cause mortality risk using initial clinical presentations. Methods We retrospectively enrolled 913 septic shock patients admitted to a tertiary hospital between January 2019 and January 2025. An AutoML framework was constructed using an Improved Red Fox Optimization (IRFO) algorithm, which optimized feature selection and hyperparameter tuning through chaotic mapping initialization and dynamic Lévy flight strategies. The dataset was split into training ( n = 730, 80%) and testing ( n = 183, 20%) sets. Six benchmark models were synchronously developed for comparison, including Logistic Regression (LR), Support Vector Machine (SVM), Adaptive Boosting (AdaBoost), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM). Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance, and model performance was evaluated via five-fold cross-validation using key metrics (AUC, sensitivity, specificity, Brier score). Clinical utility was validated via Decision Curve Analysis (DCA) and SHapley Additive exPlanations (SHAP) interpretability models. Results In the training set cross-validation, the AutoML model exhibited robust discriminative capacity (ROC-AUC=0.9127, PR-AUC=0.9059). When applied to the independent test set, it maintained optimal performance: ROC-AUC=0.8911, PR-AUC=0.8790. At the default threshold of 0.5, the model achieved sensitivity=0.9022 and specificity=0.7143; at the clinically recommended operating threshold of 0.35, sensitivity increased to 0.951 (specificity:0.648) to prioritize minimizing missed fatalities. Critical predictive features identified included Neutrophil-to-Lymphocyte Ratio (NLR), Hemoglobin (Hb), Procalcitonin (PCT), D-dimer, Mean Platelet Volume (MPV), and age. Decision Curve Analysis demonstrated significantly superior net clinical benefit over conventional intervention strategies when risk threshold exceeded 10%. For the test set, the established AutoML model demonstrated significantly superior predictive performance compared to conventional scores (SOFA, APACHE II, and qSOFA), highlighting substantially enhanced predictive accuracy. Conclusion The IRFO-optimized AutoML framework constructed a high-accuracy, interpretable prognostic model for septic shock based on admission routine blood parameters, capturing nonlinear interactions that traditional linear approaches may overlook. On the independent test set, the model demonstrated favorable predictive performance compared to traditional scoring systems (SOFA, APACHE II, qSOFA), and the accompanying visualized decision-support interface offers a potentially useful tool for emergency risk stratification.