Prediction-Based Alerting Methods Could Reduce Number and Duration of Hypoglycemic Events: An In Silico Quantification

Objective: Prevention of hypoglycemia is pivotal in type 1 diabetes treatment. Prediction of future glucose concentration based on continuous glucose monitoring (CGM) data, combined with algorithms for the generation of (preventive) hypoglycemia alerts, could help in mitigating, and sometimes even avoiding, hypoglycemic events. Here we quantify in silico the potential reduction in number and duration of hypoglycemic events obtainable by treating hypoglycemia (through ingestion of carbohydrates) on the basis of alerts generated by exploiting glucose predictions obtained by a recently presented neural network approach. Method: We generated 50 in silico CGM time series through the University of Virginia/Padova type 1 diabetes simulator (US2008/067725). The simulation scenario consisted of 2 days, three meals/day, with random variability on meal carbohydrates and insulin dosages. To quantify the potential reduction of hypoglycemia occurrence, we compared number and duration of hypoglycemic events when hypoglycemic treatment (15 g of carbohydrates) is administered on the basis of either the actual CGM value or the 30 min ahead-of-time glucose prediction obtained via the neural network. Result: Without any alert, each subject experienced, on average, two hypoglycemic events/day (with 121 min of duration, lowest glucose value 53 mg/dl). Hypoglycemia alerts generated by actual CGM reduce the number of hypoglycemic events to 1.6 per day (38 min each, lowest glucose 62 mg/dl), while hypoglycemia alerts generated using the predicted glucose further reduces the hypoglycemic events to less than 0.5 per day (21 min each, lowest glucose value 67 mg/dl). Conclusion: In silico simulation suggests that generation of hypoglycemia alerts based on glucose prediction algorithms significantly reduces number and duration of hypoglycemic events with respect to hypoglycemia alerts based on the actual CGM readings.