A Randomized Clinical Trial for Meal Bolus Decision Using Learning-based Control in Adults With Type 2 Diabetes

We propose an artificial-pancreas-like algorithm (AP-A) that could automatically determine the preprandial insulin dose based on intermittently scanned continuous glucose monitoring (isCGM) data trajectories in multiple dose injection (MDI) therapy.

We aim to determine whether preprandial insulin dose adjustments guided by the AP-A are as effective and safe as physician decisions.

We performed a randomized, single-blind, clinical trial at a tertiary, referral hospital in Beijing, China. Type 2 diabetes participants were eligible if they were aged 18 years or older, with a glycated hemoglobin A1c of 8.0% or higher. Eligible participants were randomly assigned (1:1) to the AP-A arm supervised by physician and the conventional physician treatment arm. The primary objective was to compare percentage time spent with sensor glucose level in 3.9 to 10.0 mmol/L (TIR) between the 2 study arms. Safety was assessed by the percentage time spent with sensor glucose level below 3.0 mmol/L (TBR).

A total of 140 participants were screened, of whom 119 were randomly assigned to the AP-A arm (n = 59) or physician arm (n = 60). The TIR achieved by the AP-A arm was statistically noninferior compared with the control arm (72.4% [63.3%-82.1%] vs 71.2% [54.9%-81.4%]), with a median difference of 1.33% (95% CI, -6.00 to 10.94, noninferiority margin -7.5%). TBR was also statistically noninferior between the AP-A and control arms (0.0% [0.0%-0.0%] vs 0.0% [0.0%-0.0%]), respectively; median difference (95% CI, 0.00% [0.00%-0.00%], noninferiority margin 2.0%).

The AP-A-supported physician titration of preprandial insulin dosage offers noninferior glycemic control compared with optimal physician care in type 2 diabetes.