Interpretable Prediction of Mortality Risk in Elderly Patients With Type 2 Diabetes Mellitus and Cerebral Infarction...

Objective To develop an interpretable machine learning model for predicting mortality risk in elderly intensive care unit (ICU) patients with type 2 diabetes mellitus (T2DM) and cerebral infarction,and to identify critical prognostic factors. Methods We extracted data of 514 elderly patients with T2DM and cerebral infarction from the Medical Information Mart for Intensive Care-Ⅳ database.The dataset was partitioned into training and test sets (7∶3 ratio) via scikit-learn.Within the training set,collinearity analysis was conducted,and features with variance inflation factor >5 were excluded.Lasso regression was further adopted to refine the feature selection.Six machine learning models-eXtreme Gradient Boosting (XGBoost),Logistic regression,LightGBM,AdaBoost,decision tree,and gradient boosting decision tree-were constructed and subjected to rigorous five-fold cross-validation.The optimal model was interpreted by SHAP analysis on the test set to determine the hierarchy of mortality-associated predictors and their nonlinear interactions. Results The XGBoost model demonstrated the best training performance and prediction generalization ability.The area under the curve for 30-day and 365-day mortality risk were 0.928 (95%CI=0.853-0.995) and 0.882 (95%CI=0.800-0.963),respectively.SHAP analysis revealed that the Oxford Acute Severity of Illness Score,length of hospital stay,congestive heart failure,length of ICU stay,peripheral capillary oxygen saturation,and heart rate were the top six predictive factors for 30-day mortality risk,while blood urea nitrogen,Oxford Acute Severity of Illness Score,peripheral capillary oxygen saturation,age,heart rate,and respiratory rate were the top six predictive factors for 365-day mortality risk. Conclusion The XGBoost model shows significant potential in predicting mortality risk in elderly ICU patients with T2DM and cerebral infarction,underscoring the importance of key clinical predictors.