ShockSurv: A machine learning model to accurately predict 28-day mortality for septic shock patients in the intensive care unit

Abstract

ObjectiveSeptic shock has become the leading cause of morbidity and mortality in the ICU. The survival rate of septic shock patients can be improved if their mortality can be predicted in advance to take interventions early. However, currently there is no model to predict the mortality of septic shock patients. We aim to develop such a model. MethodsWe selected septic shock patients from MIMIC-IV according to Sepsis-3. From vital signs and clinical examination data, we extracted 46 features including markers of inflammation, organ dysfunction or injury. Using these features, we developed a model named ShockSurv basing on the XGBoost Algorithm to predict 28-day mortality for septic shock patients. We also compared the model with five machine learning methods and three clinical critical scoring systems and conduct external independent validation on a multi-center database. Results11,947 septic shock patients were enrolled in the training and independent test cohort, with a 28-day mortality of 19.03%. Our model achieves AUROC of 0.9161 (95% CI: 0.8703–0.9390) by the 5-fold cross-validation in the training cohort and 0.9027 (95% CI: 0.8802–0.9253) in the independent test cohort, which are superior to the comparative models. The top 15 important variables identified by our model are consistent with clinically observed variables important for patient outcomes. ConclusionThe proposed model can accurately predict 28-day mortality for septic shock patients. SignificanceBased on the prediction, clinicians can take appropriate clinical intervention timely for more critical patients to reduce their mortality.