Development and validation of inpatient mortality prediction models for patients with hyperglycemic crisis using machine learning approaches

Abstract

Abstract Background：Hyperglycemic crisis is one of the most common complications of diabetes mellitus with a high motarlity rate. Emergency admissions for hyperglycemic crisis are still very common and challenging. The study aimed to develop and validate models for predicting the inpatient mortality risk of patients with hyperglycemic crisis admitted in emergency department using different machine learning(ML) methods. Methods: We carried out a multi-center retrospective study within six large general adult hospitals in Chongqing, western China. Patients diagnosed with hyperglycemic crisis were included based on an electronic medical record (EMR) database. The patients’ medical records along with demographics, comorbidities, clinical characteristics, laboratory results, complications, and therapeutic measures were extracted to construct theprognostic prediction model. We applied seven machine learning algorithms (support vector machines (SVM), random forest (RF), recursive partitioning and regression trees (RPART), extreme gradient boosting with dart booster (XGBoost), multivariate adaptive regression splines (MARS), neural network (NNET), and adaptive boost (AdaBoost)) compared with logistic regression (LR) to predict the risk of in-hospital death in patients with hyperglycemic crisis. Stratified random sampling was used to split the data into training (80%) and validation (20%) sets. Ten-fold cross validation was performed in train to optimize model hyper-parameters. The sensitivity, specificity, positive and negative predictive values, area under the curve (AUC) and accuracy of all models were computed in order to compare them. Results: A total of 1668 patients were eligible for the present study. The mortality rate during hospitalization was 7.3%(121/1668). In the training set, we calculated importance scores for each feature for eight models, and themost significant 10 features for all models were listed. In the validation set, all models showed good predictive capability with areas under the curve above 0.9 except the MARS model. Six machine learning algorithm models outperformed the referred logistic regression algorithm except the MARS model. RPART, RF, and SVM have better performance in the selected models (AUC values were 0.970, 0.968 and 0.968, respectively). Variable importance revealed newly detected predictors including mechanical ventilation, age, Charlson Comorbidity Index, blood gas index, first 24-hour insulin dosage, and first 24-hour fluid intake. Conclusion: All machine learning algorithms performed well to predict inpatient mortality in patients with hyperglycemic crisis except the MARS model, and the best was RPART model. These algorithms identified overlapping but different, up to 10 predictors. These models identify high-risk patients early, can assist in clinical decision-making, and can improve the prognosis of hyperglycemic crisis patients to some extent.