Exploring Machine Learning Methods to Predict Hypoglycemic States in Diabetes Type I Patients

Hypoglycemia is a serious condition associated with increased mortality in patients with type 1 diabetes, which is an incurable autoimmune disease. Hypoglycemia is defined by blood glucose levels below 70 mg/dL. The causes can include excessive insulin injections, skipping meals, or increased physical activity. It can occur suddenly and may be asymptomatic, impeding timely preventive measures. Thus, innovative technologies, such as machine learning, can help to predict the state before it occurs. Prediction models are mainly classified as short- and long-term prediction horizons (PHs) of up to 2 hours and up to 24 hours, respectively. Most research conducted in the field of diabetes forecasts blood glucose values. Still, the obtained accuracy may not be su!cient to prevent hypoglycemia due to the possible time lag of CGM devices. Moreover, most studies focus on one PH only. This thesis included short- and long-term PHs in the same classification model to consider multiple use cases and to enable better decision support. The predicted times are 5-15 min, 15-30 min, 30 min-1 h, 1-2 h, 2-4 h, 4-8 h, 8-12 h, 12-24 h before hypoglycemia. The input features are prior glucose measurements, the administered basal and bolus insulin dosages, and acceleration data. First, a correlation analysis between the input features and the classes is conducted. Thereafter, RNN and CNN are explored to classify the onset of hypoglycemia based on the proposed nine classes. Furthermore, training with six classes classifying up to 4 hours before the onset is compared. Finally, subject-specific models are tested. The population-based correlation analysis reveals a very weak association between basal insulin and glucose, and between basal insulin and acceleration data. An individual correlation analysis showed stronger relationships, but the scores varied significantly among the subjects. For the classification model with nine classes, the best results are obtained with a LSTM model. Subject-specific models improve the performance. However, only classes 0-2 could be well classified with recalls of 98%, 72%, and 50%, respectively. A population-based model with only six classes obtains better results with recalls of 99%, 73%, and 56% for classes 0, 1, and 2, respectively. In conclusion, the proposed system that includes short- and long-term PHs is not feasible with the data or models used. Whereas, a model classifying multiple short-term horizons up to 4 hours before hypoglycemia produces promising results with improved precision, and F1-measure and indicates that at least 60% of events can be predicted which is increased to approximately 70% in subject 563.