Fully automated insulin (Lyumjev) + pramlintide closed loop without carb counting

What was tested and why it matters

Every automated insulin system sold today still needs you to announce your food. The dream is a loop that just handles meals. One route to that is a second hormone: pramlintide, a lab-made copy of amylin, which a healthy pancreas releases alongside insulin at every meal and which people with type 1 diabetes lose almost completely. Pramlintide slows stomach emptying and blunts the post-meal glucagon surge, so the glucose rise after a meal is smaller and slower — in principle gentle enough for an algorithm to chase without ever being told what you ate.¹

This trial is the biggest and longest test of that idea so far, and it was designed to ask the question that actually matters: not "does adding pramlintide help?" but "can a fully automated pramlintide loop, with no carb counting at all, match a hybrid loop where you do count carbs?"²

Who it was for

Adults (18 and over) with type 1 diabetes for at least a year, already on an insulin pump for at least three months. People with gastroparesis, on gut-motility drugs, on recent GLP-1 or SGLT2 inhibitor therapy, or pregnant were excluded — sensible, because pramlintide's whole mechanism runs through the stomach.² Twenty-six adults were analysed.¹

Design

A randomized crossover trial at McGill University Health Centre, funded with Breakthrough T1D (JDRF).² Each participant did three 3-week arms in random order:

• fully automated Lyumjev (insulin lispro-aabc) + pramlintide at 8 µg per unit of insulin, no meal announcement at all;
• the same, at 10 µg per unit;
• control: automated Lyumjev + placebo, used as a hybrid loop with carbohydrate-matched boluses — i.e. the user counts carbs.

The rig was two YpsoPumps (one per hormone, because pramlintide and insulin cannot share a reservoir), a Dexcom G6 sensor, and McGill's EuGlide algorithm.² The primary endpoint was percentage of time in range, 70-180 mg/dL (3.9-10.0 mmol/L), and — importantly — this was a non-inferiority trial. The question was how much you give up by dropping carb counting.²

Key results — a narrow miss

Time in range was 71.1% (8 µg/U) and 71.2% (10 µg/U) against 72.1% on the carb-counting control. That is about a percentage point, or roughly 15 minutes a day. But the pre-specified non-inferiority margin was not met — p=0.0787 and p=0.0584.¹ Numerically these arms are almost a tie; statistically, the trial failed to prove the fully automated loop is not worse. It is a narrow miss, but it is a miss, and it should be reported as one.

Three further findings matter:

• No hypoglycemia benefit — if anything the reverse. Time below 54 mg/dL (3.0 mmol/L) was higher on the 10 µg/U arm than on control (0.3% vs 0.1%, p=0.008), and time below 70 mg/dL (3.9 mmol/L) was numerically higher on both pramlintide arms (1.1% and 1.4%) than on control (0.9%).¹ This is the clearest thing the trial says: amylin is not a hypoglycemia defence. Unlike glucagon, it cannot raise glucose — it only flattens the meal curve.
• Less insulin. Daily dose fell to 42.8 U and 41.3 U from 47.4 U on control (p<0.001).¹
• Gut side effects were the real cost. 7 and 8 of the 26 participants had gastrointestinal adverse events on the pramlintide arms, versus none on control, and two people stopped because of pramlintide nausea.¹

The one bright spot

In a subgroup of 10 participants whose baseline time in range was below 70% — the people whose control is furthest from target — the pramlintide arms clearly beat the control loop: 67.9% vs 60.6% (p=0.011).¹ That is a real, sizeable difference, and it is where the investigators say the approach should be tested next.

Read it cautiously all the same. This is a subgroup of 10 people inside a 26-person trial, and a subgroup that looks good after the primary endpoint fails is a hypothesis, not a result. It is a good reason to run another trial; it is not a reason to believe the drug works in that group yet.

What it means

The honest summary: a fully automated insulin-plus-pramlintide loop got you to roughly the same place as counting carbs, and charged you nausea for the privilege. No time-in-range gain, no hypoglycemia gain, a lower insulin dose, a second pump, and a second consumable. For someone with good control already, that is a bad trade. For someone stuck below 70% time in range, it might be a good one — and that is the trial that has not been run.

Two caveats to hold onto. This is a conference presentation, not a peer-reviewed publication — the numbers here come from the ADA meeting report and may change on publication.¹ And 26 people is a small trial: a one-point difference in time in range is well inside the noise you would expect at that size, which is precisely why the non-inferiority test could not resolve it either way.

The trial bears directly on the missing context and algorithm intelligence gaps — it is a genuine attempt to remove meal announcement entirely — and it settles something about hypo defence: the second hormone that helps there is glucagon, not amylin.

One housekeeping note for anyone checking the source: the registry entry has not been updated since before the readout, so it does not yet reflect that the study finished. We mark it completed on the strength of the ADA presentation.²¹

References

1. Doumat J, et al. (McGill University), presented at the American Diabetes Association 86th Scientific Sessions, New Orleans, June 2026. As reported by Healio: "Fully closed-loop system with insulin lispro plus pramlintide may benefit type 1 diabetes" (6 July 2026). Conference presentation — not yet peer-reviewed or published. https://www.healio.com/news/endocrinology/20260706/fully-closedloop-system-with-insulin-lispro-plus-pramlintide-may-benefit-type-1-diabetes ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

2. ClinicalTrials.gov. NCT06046417 — A Fully Automated Lyumjev and Pramlintide Delivery System for Adults With Type 1 Diabetes. McGill University Health Centre (Research Institute of the MUHC), collaborator Breakthrough T1D / JDRF. Registry record: design, arms, eligibility, primary endpoint (% time in 3.9-10.0 mmol/L). The record has not been updated since before the trial's readout. https://clinicaltrials.gov/study/NCT06046417 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶