Comparison of logistic regression and artificial neural network models for predicting hypoglycemia in non-ICU...

To analyze data from non-intensive care unit (non-ICU) inpatients with diabetes to predict the risk of hypoglycemia using electronic health records (EHRs) and point-of-care (POC) blood glucose values.

Patient demographics, laboratory results, POC blood glucose, and procedures were performed during the hospital stays on Days 0-2 to predict hypoglycemic episodes (blood glucose ≤ 3.9 mmol/L) on Days 3-6. The dataset was randomly split into a training set and an independent verification set at a 7:3 ratio. Logistic Regression (LR) and Artificial Neural Network (ANN) were compared using the area under the curve (AUC). A nomogram plot was also constructed to display the predicted hypoglycemia probabilities.

Data from 16,593 diabetic patients (January 2017 to June 2022) were analyzed. Predictive factors from the LR model included the use of insulin; previous hypoglycemia in Days 0-2; respiratory rate; blood urea nitrogen; potassium; D-dimer levels; coefficient variation of blood glucose (BG CV) > 31%; and blood glucose gap (BG gap, maximum of blood glucose - minimum of blood glucose) > 10 mmol/L. In the verification set, the AUC of ANN was 0.762 and the AUC of LR was 0.763. There was no significant difference in the effects of the models built by the two methods. The results showed that the probability predicted by the nomogram using LR is similar to the clinical results. Decision curve analysis (DCA) indicated potential clinical application for the LR model.

The LR model demonstrated considerable value in predicting hypoglycemia risk, comparable to ANN. Trials of such models should be conducted to evaluate their utility in reducing inpatient hypoglycemia.

Not applicable.