Early prediction of septic shock in ICU patients using machine learning: development, external validation, and explainability with SHAP.

BACKGROUND

Septic shock is a severe and life-threatening complication of sepsis associated with high mortality. Early identification remains challenging due to the heterogeneous clinical presentation of patients, incomplete real-world ICU data, and the dynamic pathophysiological progression of sepsis. This study aimed to develop and externally validate machine learning (ML) models to predict the progression to septic shock in intensive care unit (ICU) patients with sepsis.

METHODS

Data were extracted from two large critical care databases: MIMIC-IV (training set) and eICU-CRD (validation set). Variable selection was performed using LASSO regression. Six ML algorithms-random forest (RF), XGBoost, support vector machine, light gradient boosting machine, logistic regression, and naïve Bayes-were trained on the MIMIC-IV dataset and externally validated on the eICU dataset. Model performance was evaluated using AUC, F1 score, sensitivity, specificity, and balanced accuracy. Model interpretability was assessed using SHapley Additive exPlanations (SHAP).

RESULTS

A total of 11,383 septic patients were included, of whom 2,259 in the training cohort and 1,212 in the validation cohort developed septic shock. The RF model achieved the best performance, with an AUC of 0.785, balanced accuracy of 0.717, and an F1 score of 0.511. SHAP analysis identified SOFA score, heart rate, creatinine, SAPS II, and OASIS as the most influential predictors.

CONCLUSION

The proposed ML model enables early prediction of septic shock using routinely collected ICU data. SHAP-based interpretation enhances transparency and clinical interpretability. This approach may assist in timely risk stratification, support data-driven decision-making, and ultimately improve outcomes in patients with sepsis.