Development and validation of machine learning models for predicting 28-day mortality in diabetic sepsis patients

Li Yuqian; Bayina Baterongsu; Cui Jian; Wang Yixi; Li Wenzhe

doi:10.3969/j.issn.1002-1949.2026.02.002

Abstract：Objective This study aimed to develop a 28-day mortality risk prediction model for diabetic sepsis patients using machine learning,with the goal of optimizing treatment and improving clinical outcomes.Methods Using the MIMIC-Ⅳ 2.2 database,a screening strategy based on the Sepsis-3.0 criteria was implemented.Diabetic sepsis patients were selected using ICD-9 and ICD-10 diagnosis codes,and clinical data were extracted.Patients were grouped by survival status at 28 days.The Boruta algorithm(bi-parameter),Logistic regression,and Lasso regression were used to select features.Eight machine learning models-Logistic regression,gradient boosting machine(GBM),LightGBM,AdaBoost,CatBoost,k-nearest neighbors,neural networks,and support vector machines-were used to predict the 28-day mortality risk.Hyperparameter optimization and cross-validation were performed to assess the risk of model overfitting.The best model was selected based on performance metrics and further explained using SHAP analysis.Results A total of 9 235 diabetic sepsis patients were included in the study.Ten important feature variables were identified[APACHE Ⅱ score,age,heart rate,respiratory rate,SOFA score,blood lactate,Charlson comorbidity index,oxygen saturation,activated partial thromboplastin time(APTT),and pH value].After hyperparameter optimization and model performance evaluation,the GBM model demonstrated the best stability and predictive ability.The area under the curve(AUC)of the GBM model in the training and test sets were 0.825 and 0.771 respectively.Decision curve and calibration curve analysis showed that the GBM model provided significant net clinical benefit and stability.SHAP analysis revealed that the APACHE Ⅱ score,age,and heart rate had the greatest impact on the GBM model's predictive performance.The model is user-friendly and can quickly predict the 28-day mortality risk of diabetic sepsis patients.Conclusions The 28-day mortality risk prediction model for diabetic sepsis patients,constructed and validated using the GBM algorithm,shows good clinical applicability and interpretability of pathological mechanisms.It may enable precise stratification for early intervention in patients,optimizing clinical treatment strategies.