Machine learning based prediction models for the prognosis of COVID-19 patients with DKA

Abstract

Patients with Diabetic ketoacidosis (DKA) have increased critical illness and mortality during coronavirus diseases 2019 (COVID-19). The aim of our study was to develop a predictive model for the occurrence of critical illness and mortality in COVID-19 patients with DKA utilizing machine learning. Blood samples and clinical data from 242 COVID-19 patients with DKA collected from December 2022 to January 2023 at Second Xiangya Hospital. Patients were categorized into non-death (n = 202) and death (n = 38) groups, and non-severe (n = 146) and severe (n = 96) groups. We developed five machine learning-based prediction models-Extreme Gradient Boosting (XGB), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP)-to evaluate the prognosis of COVID-19 patients with DKA. We employed 5-fold cross-validation for model evaluation and used the Shapley Additive Explanations (SHAP) algorithm for result interpretation to ensure reliability. The LR model demonstrated the highest accuracy (AUC = 0.933) in predicting mortality. Additionally, the LR model excelled (AUC = 0.898) in predicting progression to severe disease. This study developed a machine learning-based predictive model for the progression to severe disease or death in COVID-19 patients with DKA, which can serve as a valuable tool to guide clinical treatment decisions.

Keywords: COVID-19; Diabetes; Ketoacidosis; Machine learning; Prediction.

Fig. 1

Process flowchart for handling sample data information.

Fig. 2

Evaluation of the five machine learning algorithms based on the AUC of the ROC curve. (a) The AUC about the death group versus the non-death group; (b) The AUC about the severe group versus the non-severe group.

Fig. 3

Shapley Additive Explanations (SHAP) summary plot in the LR model. (a) SHAP beeswarm plot showed the distribution of SHAP values of each feature to predict mortality in patients with DKA. Red represents higher feature values, and blue represents lower feature values. (b) SHAP beeswarm plot showed the distribution of SHAP values of each feature to predict the severe illness. Red represents higher feature values, and blue represents lower feature values.