Human-machine co-adaptation to automated insulin delivery: a randomised clinical trial using digital twin technology
Boris Kovatchev, Patricio Colmegna, Jacopo Pavan, Jenny L. Diaz Castaneda, María F. Villa-Tamayo, Chaitanya L.K. Koravi, Giulio Santini, C. Alix, Meaghan Stumpf, Sue A. Brown
Abstract
Most automated insulin delivery (AID) algorithms do not adapt to the changing physiology of their users, and none provide interactive means for user adaptation to the actions of AID. This randomised clinical trial tested human-machine co-adaptation to AID using new 'digital twin' replay simulation technology. Seventy-two individuals with T1D completed the 6-month study. The two study arms differed by the order of administration of information feedback (widely used metrics and graphs) and in silico co-adaptation routine, which: (i) transmitted AID data to a cloud application; (ii) mapped each person to their digital twin; (iii) optimized AID control parameters bi-weekly, and (iv) enabled users to experiment with what-if scenarios replayed via their own digital twins. In silico co-adaptation improved the primary outcome, time-in-range (3.9-10 mmol/L), from 72 to 77 percent (p < 0.01) and reduced glycated haemoglobin from 6.8 to 6.6 percent. Information feedback did not have additional effect to AID alone. (Clinical Trials Registration: NCT05610111 (November 10, 2022)).