Researchers present a way to personalize automated insulin delivery (AID) controllers by adjusting a small set of controller settings in real time using a person’s daily glucose data. The idea is to let the controller learn from the user’s actual blood-glucose outcomes each day, rather than rely only on settings chosen to be safe for the average person. The authors aim to reduce both high and low glucose while keeping treatment safe.

Technically, the team frames personalization as projected gradient descent on a daily risk score. In plain terms, the algorithm nudges controller parameters downhill on a measured “risk” surface so the system improves over days. To estimate the slope of that surface from noisy glucose data, they use a regularized weighted least squares estimator (RWLS). This estimator gives more weight to recent parameter settings near the current one and adds regularization to reduce the effect of measurement noise and day-to-day metabolic variability. They also add a small, bounded perturbation (dither) so the algorithm can explore and avoid getting stuck.

The authors tune two controller parameters. One parameter adjusts how conservatively the controller avoids stacking insulin doses after recent injections. The other shifts the controller’s operating point for background insulin delivery. They leave a global scaling parameter to existing methods. To justify that the adaptation will behave well, they use contraction theory, a mathematical tool that argues the glucose control system “forgets” past conditions and settles into a repeatable pattern under regular daily routines.

They tested the method in computer trials on the FDA‑accepted UVA/Padova Type 1 diabetes simulator using 100 virtual adults. The simulator includes variability in meal timing, meal size, and insulin sensitivity. Measured by time-in-range (TIR, glucose 70–180 mg/dL), their approach increased TIR by about 2% after 4 weeks, 3% after 8 weeks, and 4% after 17 weeks. The daily risk metric also includes penalties for low glucose and for treatments used to recover from hypoglycemia, so the adaptation balances avoiding both highs and dangerous lows.