Machine learning prediction models and nomogram to predict the risk of in-hospital death for severe DKA: A clinical study based on MIMIC-IV, eICU databases, and a college hospital ICU

Abstract

Aim: To establish a prediction model and assess the risk factors for severe diabetic ketoacidosis (DKA) in adult patients during the ICU.

Introduction: With DKA hospitalization rates consistently increasing, in-hospital mortality has become a growing concern.

Methods: DKA patients aged >18 years old in the US-based critical care database (Medical Information Mart for Intensive Care (MIMIC-IV)) were considered. Independent risk factors for in-hospital mortality were screened using extreme gradient boosting (XGBoost) and the Bayesian information criterion (BIC) optimal subset regression. One predictive model was developed using machine learning extreme gradient boosting (XGBoost), and the other one was a nomogram based on logistic regression to estimate risks of in-hospital mortality with severe DKA. Established models were assessed by using internal validation and external validation. The MIMIC-IV was split into training and testing samples in a 7:3 ratio. The eICU Collaborative Research Database and admissions data from the department of critical care medicine of the first affiliated hospital of Harbin medical university were used for independent validation. The discriminatory ability of the model was determined by illustrating a receiver operating curve (ROC) and calculating the C-index. Meanwhile, the calibration plot and Hosmer-Lemeshow goodness-of-fit test (HL test) was conducted to evaluate the performance of our new build model. Decision curve analysis (DCA) was performed to assess the clinical net benefit. Net Reclassification Improvement (NRI) was used to compare the predictive power of the two models.

Results: A multivariable model that included acute physiology score III (APS III), the highest levels of blood plasma osmolality (osmolarity_max), minimum osmolarity (osmolarity_min)/osmolarity _max, vasopressor, and the highest levels of blood lactate was represented as the nomogram. The C- index of the nomogram model was 0.915 (95% CI: 0.966-0.864) in the training dataset and 0.971 (95% CI: 0.992-0.950) in the internal validation. The nomogram's sensitivity was well according to all data's HL test (P > 0.05). DCA showed that our model was clinically valuable. The XGB (extreme gradient boosting) model achieved an AUC (area under the curve) of 0.950 (95% CI, 0.920-0.980); however, the nomogram model made was more effective than XGB based on NRI.

Conclusion: The predictive XGB and nomogram models for predicting in-hospital patient deaths with DKA were effective. The forecast models can help clinical physicians promptly identify patients at high risk of DKA, prevent in-hospital deaths, and promptly intervene.

Keywords: Diabetes; Diabetic ketoacidosis; Machine learning; Nomogram.