Probabilistic abnormal glycemic event alert for T1DM patients

Predicted glucose value based on continuous glucose monitoring data can help to provide an early hyper/hypoglycemic alert allowing patients to make necessary decision to avoid abnormal glycemic event. Traditional alert strategy does not consider the existence and influences of prediction errors resulting from modeling accuracy, external disturbances, etc. Hyper/hypoglycemic alert is generated by simply comparing the prediction value against the threshold, which may result in high false alert rate and missing alert rate, and thus incorrect therapy. To solve the above problem, a probability based method is proposed in this paper to quantify the probabilistic alert of hyper/hypoglycemia, which will tell the possibility of the blood glucose going into abnormal glycemic episode. It is achieved by quantitative analysis of uncertainty based on prediction errors. Here, the Gaussian mixture model (GMM) is used to estimate the probability density function of the prediction errors. Then the probability of the hyper/hypoglycemia alert is calculated. The proposed method is investigated based on thirty in silico subjects.