Nocturnal Hypoglycemia Detection using Optimal Bayesian Algorithm in an EEG Spectral Moments Based System

This paper presents a hypoglycemia detection system using electroencephalogram (EEG) spectral moments from 8 patients with type 1 diabetes (T1D) at night time. Four channels (C3, C4, O1, and O2) associated with glycemic episodes were analyzed. Spectral moments were applied to EEG signal and its corresponding speed and acceleration. During hypoglycemia, theta moments increased significantly (P<0.001) and alpha moments decreased significantly (P<0.001). The system used an optimal Bayesian neural network for detecting hypoglycemic episodes. Based on the optimal network architecture with the highest log evidence, the final classification results for the test set were 79% and 51% in sensitivity and specificity, respectively. Essentially, the estimated blood glucose profiles correlated significantly to actual values in the test set (P<0.0001). Using error grid analysis, 93% of the estimated values were clinically acceptable.