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

BACKGROUND

Hyperglycemic crisis is one of the most common and severe complications of diabetes mellitus, associated with a high motarlity rate. Emergency admissions due to hyperglycemic crisis remain prevalent and challenging. This study aimed to develop and validate predictive models for in-hospital mortality risk among patients with hyperglycemic crisis admitted to the emergency department using various machine learning (ML) methods.

METHODS

A multi-center retrospective study was conducted across six large general adult hospitals in Chongqing, western China. Patients diagnosed with hyperglycemic crisis were identified using an electronic medical record (EMR) database. Demographics, comorbidities, clinical characteristics, laboratory results, complications, and therapeutic interventions were extracted from the medical records to construct the prognostic prediction model. Seven machine learning algorithms, including 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) were compared with logistic regression (LR) for predicting the risk of in-hospital mortality 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 on the training set to optimize model hyperparameters. The sensitivity, specificity, positive and negative predictive values, area under the curve (AUC) and accuracy of all models were computed for comparative analysis.

RESULTS

A total of 1668 patients were eligible for the present study. The in-hospital mortality rate was 7.3% (121/1668). In the training set, feature importance scores were calculated for each of the eight models, and the top 10 significant features were identified. In the validation set, all models demonstrated good predictive capability, with areas under the curve value exceeding 0.9 with a F1 score between 0.632 and 0.81, except the MARS model. Six machine learning algorithm models outperformed the referred logistic regression algorithm except the MARS model. Among the selected models, RPART, RF, and SVM achieved the best performance in the selected models (AUC values were 0.970, 0.968 and 0.968, F1 score were 0.652, 0.762, 0.762 respectively). Feature importance analysis identified novel predictors including mechanical ventilation, age, Charlson Comorbidity Index, blood gas index, first 24-hour insulin dosage, and first 24-hour fluid intake.

CONCLUSION

Most machine learning algorithms exhibited excellent performance predicting in-hospital mortality among patients with hyperglycemic crisis except the MARS model, and the best one was RPART model. These algorithms identified overlapping but different, up to 10 predictors. Early identification of high-risk patients using these models could support clinical decision-making and potentially improve the prognosis of hyperglycemic crisis patients.

CLINICAL TRIAL NUMBER

Not applicable.