Ilkka Launonen, Norway
University Hospital of North Norway Kvalitets- og utviklingssenteret (KVALUT)Presenter of 1 Presentation
LOGISTIC REGRESSION FOR EARLY DETECTION OF HYPOGLYCEMIA
Abstract
Background and Aims
Detection of a potential hypoglycemia early enough is crucial to prevent its adverse effects ahead in time. Blood glucose prediction has been typically based on an ARX time series model or machine learning methods by training of neural networks, for example. Our goal is to investigate how logistic regression, where the probability of the future hypoglycemic event is modelled directly, performs in the task. The effect of variables other than CGM readings or insulin have been less studied with respect to predictive accuracy. Our aim is to also study how different predictor variables, particularly in relation to exercise, affect the predictions.
Methods
We use logistic regression to model the log-odds of the hypoglycemic event. The model is estimated from an individual T1DM patient’s data, and is thus personalized. We approach the early detection of hypoglycemia as a binary classification problem. At each time point, the classifier predicts, based on the values of predictor variables, that blood glucose will be either lower or higher than the hypoglycemic threshold at the end of the prediction time horizon. The classification is based on hard thresholding the estimated probability of hypoglycemia.
Results
We estimated our models on the OhioT1DM dataset, which contains six patients' data over a period of 8 weeks. We considered several hard threshold values for the classifier to compute the models' ROC curves, and compared predictive accuracy using different sets of predictor variables found in the dataset.
Conclusions
The results show that logistic regression is a viable alternative for the early detection of hypoglycemia.