Background and Aims

We designed a safety-focused methodology for automatically dosing prandial insulin by utilizing a glycemic disturbance detector.

Methods

By using regularized deconvolution to solve for unknown disturbances, we simulated how changes in insulin doses would affect glycemia in a clinical dataset of 14 patients over a month (ClinicalTrials.gov NCT03859401). Our automatic bolus strategy relied on an algorithm that determined the probability of glycemic disturbances requiring insulin to prevent hyperglycemia. Percentages of the individual’s total daily insulin (TDI) were delivered at different probability thresholds. After injection, blood glucose was simulated for 120 minutes and the number of hypoglycemic events was counted. We determined an escalating dosing strategy based on what led to no more than one additional hypoglycemic event per day. We then simulated the dataset using automatic doses to augment or replace boluses in the record.

Results

From our criteria, we chose thresholds of 3%, 4%, 5%, 6%, and 9% TDI at probabilities of 0.3, 0.5, 0.7, 0.8, and 0.9, respectively. When the full dataset was simulated, there was an increase of 0.88% percent time <70 mg/dL, a 2.85% increase in percent time 70-140 mg/dL, a 2.09% increase in percent time 70-180 mg/dL, a reduction of 2.97% percent time >180 mg/dL, and 1.47% less percent time >250 mg/dL when compared to the observed hybrid closed-loop (HCL) data.

Figure 1 - Number of hypoglycemic events per day at different TDI percentages and thresholds.

Conclusions

This automatic bolusing strategy increased euglycemia and decreased hyperglycemia in simulation when compared to HCL, while only increasing hypoglycemia slightly.