Background  Patients with sepsis have a high mortality rate. Accurate predicting the occurrence of adverse outcomes will benefit the effective allocation of medical resources.

  Objective  To construct machine learning model to predict the risk of early mortality in patients with sepsis and provide evidences for clinical decision-making.

  Methods  Patients with sepsis meeting the Sepsis-3 criteria were screened from the Medical Information Mart for Intensive Care (MIMIC)- Ⅲ database. We randomly selected 70% of the data as the training set to build the model, and the remaining 30% of the data were used as the validation set. Extreme gradient boosting (XGBoost) model was applied to integrate relevant parameters to predict the risk of death in intensive care unit (ICU) of patients with sepsis. The predictive ability of the model was evaluated by receiver operating characteristic (ROC) curve, and the model was compared with simplified acute physiology score (SAPSⅡ, sequential organ failure assessment (SOFA), system inflammatory reaction syndrome (SIRS), quick sequential organ failure assessment (qSOFA) to verify the predictive value.

  Results  A total of 4939 sepsis patients meeting the inclusion criteria were included, of which 551 died in the ICU and 4388 survived. Based on whether the patient died in ICU as the outcome, the area under ROC curve (AUC) obtained by the established XGBoost model was 0.848, with sensitivity of 0.841, specificity of 0.711 and accuracy of 0.726, all of which were higher than other scores(P<0.05). The top 10 features of the model were lactate level, vasoactive-inotropic score (VIS), malignant tumor, aspartate aminotransferase (AST), anion gap (AG), mechanical ventilation, international normalized ratio (INR), Glasgow Coma Scale (GCS), and the type of intensive care unit.

  Conclusion  The performance of the XGBoost model is superior to other commonly used clinical scores for accurate prediction of the early death risk of patients with sepsis. Application this model can assist in making clinical decision and allocating medical resources.