Moderator of 1 Session
Presenter of 1 Presentation
Integrating metabolic expenditure data from wearable sensors into an automated insulin delivery system: clinical study results
Abstract
Abstract Body
Background: The objective was to evaluate an automated insulin delivery (AID) system that responds automatically to physical activity.
Methods: We evaluated an exercise-aware model predictive control (ExMPC) AID using iPancreas developed at OHSU, which includes a Dexcom G6 CGM, an Insulet Omnipod, a control algorithm running on a Samsung S9 smart-phone, and a Polar M600 smart watch. Another exercise-aware algorithm, fading-memory-proportional-derivative (FMPD) was also evaluated. Heart rate and accelerometer data from the smart-watch were combined to calculate metabolic equivalent of task (MET). METs were continuously used in ExMPC to adjust insulin delivery. FMPD notified when METS exceeded a threshold of 4 METs and then shut off insulin for 30 minutes, then reduced insulin by 50% for 1-hour. We compared ExMPC with FMPD in a 2-arm, randomized 3-day outpatient study that included an in-clinic 30-minute aerobic exercise video on day 1. Wilcoxon rank-sum test determined difference in % time-in-range (TIR: 70-180 mg/dL), % time-low (TL: <70 mg/dL), and % time very low (TVL: <54 mg/dL) between algorithms during in clinic exercise and across the entire study.
Results: From start of in-clinic exercise to 2-hours post-exercise, ExMPC (n=18) had higher TIR than FMPD (n=20), (87.5% vs. 76.3%, P=.046) and trended towards less TVL (0.0% vs. 1.5%, P=.09). Across the entire study, TIR (74.5% vs. 75.7%) and TL (1.0% vs. 1.4%) were comparable between algorithms.
Conclusions: An exercise-aware MPC AID can safely control glucose levels during exercise and under free-living conditions without the need for notifications and confirmations from a user.