Study On Data-Driven Glucose Prediction Methods

In recent years, the number of diabetic patients is increasingly developed, so the researchers all over the world pay more and more attention on finding the treatment therapies for diabetes. So far, one of the most effective ways to remedy the diabetes is exogenous insulin infusion. In order to avoid hypoglycemia and provide useful information for glycemic control, the algorithms need past and present information to predict blood glucose level. Extreme learning machine (ELM) and its modified version (RELM), have been firstly used for blood glucose prediction and hypoglycemia alarm by using CGMS data. Using three different prediction horizons (10min,20min,30min), these two algorithms are evaluated systematically in terms of sensitivity and specificity. The area under the ROC curve is used to evaluate the comprehensive performance. From the experiment results, it can be concluded that the bigger the prediction horizon is, the worse the prediction performance is. Both two algorithms can obtain good specificity and acceptable sensitivity. Compared with ELM, RELM has worse sensitivity but better specificity. These two methods have comparable comprehensive performance.Second order glucose model has been implemented to describe the glucose kinetic. Compared with the existing three order model, the proposed one combined with Kalman filter is more suitable for predict blood glucose. The clinical usability is judged by RMSE and Clarke grid error analysis.