Hypoglycemia prediction using extreme learning machine (ELM) and regularized ELM

Abstract

Hypoglycemia prediction plays an important role for diabetes management. Along with the development of continuous glucose monitoring (CGM) technology, blood glucose prediction becomes possible. Using CGM readings, extreme learning machines (ELM) and regularized ELM (RELM) are implemented in this paper to predict hypoglycemia. Under three different prediction horizons, 10, 20, and 30 min, these two methods are compared systematically in terms of root mean square error (RMSE), sensitivity, and specificity. In addtion, receiver operating characteristic (ROC) curve as a function of sensitivity and specificity is applied to evaluate the performace of ELM and RELM. The area under curve (AUC) value was used the evaluate the ROC performance for different test accurately. The experiment results demonstrate that these two methods can predict hypoglycemia pretty good. As expect, the bigger prediction horizon (PH), induce the worse performance. As hypoglycemia threshold is increased, sensitivity impoves at cost of spcificity. Both methods can get good specificity and acceptable sensitivity. Good specificity can make sure each alarm is effective for patients to take correct actions. In terms of AUC, ELM and RELM have comparable performance for hypoglycemia prediction.