| dc.description.abstract | Diabetes is a chronic disease that needs continuous blood glucose monitoring and self-management. The improper control of blood glucose levels in diabetic patients can lead to serious complications such as kidney and heart diseases, strokes, and blindness. The proper treatment of diabetes, on the other hand, can help a person live a long and normal life. On the other hand, tighter glycemic controls increase the risk of developing hypoglycemia, a sudden drop in a patients' blood glucose levels that can lead to coma and possibly death if proper action is not taken immediately. Continuous Glucose Monitoring (CGM) sensors placed on a patient body measure glucose levels every few minutes. They are also capable of detecting hypoglycemia. Yet detecting hypoglycemia sometimes is too late for a patient to take proper action, so a better approach is predicting the hypoglycemia event before it occurs. Recent research efforts have been made in predicting subcutaneous glucose levels at specific points in the future. Moreover, the models developed used are ill suited for predicting out-of-range glucose values, namely, hypoglycemia and hyperglycemia. Hence, in this research, we use machine learning techniques suitable for predicting hypoglycemia within a prediction horizon of thirty minutes. This period should be long enough to enable the diabetes patients to avoid hypoglycemia by taking proper action. In specific, we use and compare two approaches to perform the hypoglycemia prediction, namely, a time sensitive artificial neural networks (TS-ANN) and tree based temporal classification (TBTC) by applying feature extraction from the patient glucose signal. While the TS-ANN performed reasonably well (with average sensitivity= 80.19%, average specificity= 98.2%, and average accuracy= 97.6%), nevertheless, the TBTC approach outperformed the TS-ANN one with the ability to predict hypoglycemia events accurately (with average sensitivity= 93.9%, average specificity= 98.8, average accuracy= 98.16%) using three aggregate global features; mean, minimum, and difference, and two parameterized event primitives (PEPs), namely the negative slope and local minimum of the glucose signal. | en_US |
