Closed-Loop Insulin Delivery: What Autonomy Actually Looks Like
Closed-Loop Insulin Delivery: What Autonomy Actually Looks Like
Closed-Loop Insulin Delivery: What Autonomy Actually Looks Like
The marketing language around automated insulin delivery systems promises something close to a cure. "Let the system do the work." "Freedom from diabetes management." The Medtronic 780G, Omnipod 5, and Tandem Control-IQ are real engineering achievements. They are also, in daily clinical use, far less autonomous than most patients expect when they start.
I want to be precise about what "closed-loop" means here, because the term does real work in shaping patient expectations.
The loop, in practice
A closed-loop system reads continuous glucose monitor (CGM) data, runs it through a control algorithm, and adjusts basal insulin delivery via pump. The Medtronic 780G uses a PID (proportional-integral-derivative) algorithm with meal-detection logic. Control-IQ from Tandem uses a model-predictive control approach, forecasting glucose 30 minutes ahead. Omnipod 5 runs a version of the algorithm developed in the Type Zero collaboration at the University of Virginia.
All three systems modulate basal rates. All three can deliver automatic correction boluses. None of them bolus for meals on their own.
That last point matters enormously. The patient still has to count carbohydrates, estimate them correctly, and deliver a manual meal bolus. Carb-counting errors are the single largest source of postprandial glucose spikes in people using these systems. A 2023 study by Breton et al. in Diabetes Care showed that even with Control-IQ running, time-in-range dropped from 78% to 61% when carb estimates were off by more than 30%.
Where the autonomy breaks down
There are other gaps. CGM sensors drift. Medtronic's Guardian 4 sensor still requires periodic calibration, though less often than its predecessors. Dexcom G7, used with Omnipod 5 and Control-IQ, is factory-calibrated, but accuracy degrades on days 8 through 10 of a 10-day sensor. The algorithm trusts the CGM value. If the CGM reads 6.2 mmol/L and the actual glucose is 8.5, the system will under-deliver insulin.
Site absorption variability is another real problem. Lipohypertrophy at overused infusion sites slows insulin absorption. The algorithm expects a pharmacokinetic profile that no longer matches reality. There is no sensor for that. The patient has to rotate sites and recognize when absorption has changed.
Then there are exercise and illness. Current systems have "activity" modes that raise the glucose target, but they are blunt instruments. A patient going for a 20-minute walk and a patient doing high-intensity intervals need very different adjustments. The algorithm cannot distinguish between the two.
What clinicians should tell patients
I think we do patients a disservice when we frame these devices as "artificial pancreas" systems without qualification. They are hybrid closed-loop systems, and the "hybrid" part is doing most of the heavy lifting in daily life. The algorithm handles overnight basal management well. Time-in-range from midnight to 6 AM is consistently above 85% across published trials. That is genuinely life-changing for people who previously woke up at 3 AM to treat lows.
But daytime management, with meals and activity and stress, is still a partnership between the patient and the machine. Patients who expect full autonomy get frustrated when their glucose spikes after a meal they did not bolus for correctly. That frustration sometimes leads to device abandonment, which is worse than the alternative.
The honest framing is this: these systems are very good copilots. They are not yet pilots. The regulatory and engineering path toward fully closed-loop, meal-announcement-free systems is years away, pending dual-hormone (insulin-glucagon) delivery and faster-acting insulin formulations. Until then, the word "autonomous" needs an asterisk, and clinicians should be the ones explaining what that asterisk means.