Critical care (e.g. trauma and cardiothoracic surgical) and diabetic patients are prone to variability in glucose concentration on a daily basis. Hypoglycemic and hyperglycemic glucose values in these patient populations have been associated with decreased patient outcomes. In diabetic patients, persistently elevated glucose values are associated with development of long term complications such as, but not limited to retinopathy, neuropathy, and nephropathy. In the critical care patient population, elevated glucose has been correlated to increases in mortality, length of stay in the intensive care unit (ICU), and morbidities. The maintenance of tight glycemic control in these patients without severe hypoglycemia or glycemic variability appears to improve outcomes in these patients.
Various factors are associated with future glycemic excursions such as, but not limited to: lifestyle/activities (e.g. sleep-wake cycles), emotional factors (e.g. stress), nutritional intake, medication dosages, and ICU medical records (in critical care patients). In the field of diabetes research, models for prediction of glucose and/or models used to maintain tight glycemic control have been the focus of research. In the critical care patient population, very little research into development of such models has been completed to date.
Multiple factors affect or are indicators of future glucose concentration. A suitable modeling technique needs to incorporate the effect of such factors for accurate prediction of glucose. A modeling technique well suited for this task is a neural network model.A neural network is an adapative modeling technique, which learns and updates model parameters based on determining patterns/trends existent in input data.This adapative capability, makes neural network modeling well suited for prediction of glucose where multiple factors impact future glycemic excursions.
This dissertation summarizes the development and optimization of various neural network model architectures for the real-time prediction of glucose in diabetic and critical care patients. Neural network models were configured to predict glucose using prediction horizons >60 minutes, which have not been attained in many predictive models to date. The performance of the neural network model is assessed via determination of overall model error, percentage of glycemic extremes predicted, and clinical acceptability of model predictions as determined via Clarke Error Grid Analysis.