Did a change of glucagon derail the Inreda artificial pancreas trial?

The Inreda AP is a Dutch dual-hormone fully closed loop. It doses insulin and glucagon from two Guardian 3 sensors with no meal announcements at all, and earlier studies had reported time in range of around 80%. DARE was the randomised trial meant to take it into the Dutch basic insurance package. Its results were published in The Lancet Regional Health Europe this autumn, and they weren’t what anyone following the system had hoped for: about 65% time in range, and no benefit over hybrid closed loop.

In September Zorginstituut Nederland (ZIN) concluded that effectiveness hadn’t been sufficiently demonstrated. According to Inreda, around 127 existing users now face paying for it themselves or moving to something else.

There’s a complication. Shortly before the trial started, the glucagon the system had been built around, GlucaGen from Novo Nordisk, wasn’t available, and the trial used Glugon from United Biotech in India instead. The authors and ZIN both suggest that held the results back. I wanted to know how far that goes towards explaining what happened, so I went through the paper, its supplement, the ZIN assessment and Inreda’s own real-world data.

Running a year-long trial of a novel device across 14 centres is hard, and losing a key ingredient just before you start is something nobody plans for. But the answer matters to the people using the system, and I’ve ended up with some questions.

What DARE found

DARE randomised 243 adults with type 1 diabetes. One cohort compared the Inreda AP with injections plus CGM, the other with hybrid closed loop, mostly the 780G. Everyone wore the same blinded FreeStyle Libre Pro IQ to measure outcomes, whichever system they were using.

So it beat injections, and it didn’t beat hybrid closed loop. ZIN’s verdict comes from assessing the two cohorts together, which gives a difference of 2.4 points (−0.7 to 5.5) against the 5 points it treats as clinically relevant, and it rated the evidence as very low quality because of the glucagon switch and the imprecision. Both cohorts’ figures are a long way below the 80% of the earlier studies.

Time in range wasn’t the only problem. 59 of the 119 people allocated to the Inreda AP stopped using it during the year, most of them in the first three months. There were nine severe hypos in the dual-hormone arm against two in usual care, and serious adverse events of any kind affected 14.3% of people on the Inreda AP against 4.8% in usual care. ZIN rated that safety evidence as very low quality too, partly because it thinks the glucagon switch may have inflated the risk, and said the events other than severe hypos didn’t appear to be related to the device. It regards safety as uncertain.

The glucagon switch

ZIN’s assessment says the two glucagons were expected to behave the same, but that during the study Glugon turned out to act more slowly and less strongly than GlucaGen. During the trial Inreda released two software updates to compensate, giving glucagon earlier and raising the daily maintenance dose, and later a new infusion set to reduce occlusions.

The clearest evidence of what the switch cost comes from Inreda itself. In three years of real-world data on 102 users, published in September by Jansen and colleagues, the people who stayed on GlucaGen held time in range at around 80%. In the 51 who switched to Glugon, it fell from 79.3% to 75.8% over three months, a drop of 3.5 points. Over the same period their glucagon use went up by about half, from 0.55 to 0.82 mg a day, and they still had more lows. That’s what you’d expect if each dose was doing less, or doing it later.

This is a retrospective before-and-after comparison by the manufacturer, so it can’t rule out other changes over the same period. It’s still the most direct evidence there is.

What happens if you add those 3.5 points back to DARE? I applied them to the people still using the system at 12 months, as a scenario rather than an estimate. The difference against injections grows from about 8 points to 10, and the difference against hybrid closed loop goes from −0.4 to about +1.3. Even on that generous assumption the Inreda AP doesn’t clearly beat hybrid closed loop. The scenario gives nothing to the people who stopped, though, and if a slower and less effective glucagon made people give up, that’s where some of its effect would be.

Good nights, harder days

The supplementary tables split time in range by day and night, and this is the part I found most interesting. Overnight the dual-hormone arm reached 77.6% and 79.6% in the two cohorts, beating hybrid closed loop by 5.2 points (1.1 to 9.2). During the day it managed 62.8% and 60.8%, about level with hybrid closed loop.

These figures include only people with sensor data, so they don’t match the headline exactly, but the pattern is clear enough. The shortfall is during the day, when people eat. A loop that’s told nothing about meals has to react to every one after it’s under way, and the low that can follow a large reactive dose of insulin has to be caught by cutting insulin, giving glucagon or both. A weaker, slower glucagon could make that harder, particularly when the system is trying to counteract insulin already delivered. How much of DARE’s daytime shortfall this explains isn’t clear, but the overnight figures suggest the underlying approach works.

What the glucagon may not explain

Not everything traces back to the glucagon. People could give more than one reason for stopping, and the commonest two were dissatisfaction with glucose levels and device burden, which included the fingersticks needed to calibrate two sensors and the size of the device. The first may well be the glucagon. A different glucagon won’t change the second.

There’s also a gap of about 11 points between Inreda’s real-world users after the switch (75.8%) and the DARE participants who stayed on the system (64.4%). Both groups were on Glugon. Inreda’s users were experienced and had chosen to stay on the system, while DARE’s were new to it. And Inreda’s figures come from the system’s own sensors, while DARE measured everyone on an independent blinded Libre. A loop treats its own sensor’s errors as real glucose and steers them into range, so it may well look better on its own sensor than on someone else’s. I don’t think the published data can say how much of the 11 points each of these accounts for. Because both arms in DARE wore the same blinded sensor, the comparison within the trial is protected from that problem.

There’s a wider lesson here too. When open-source AID studies report impressive results from experienced users, the usual criticism is that they’re self-selecting cohorts. That’s fair, and the same scrutiny should apply to commercial systems. Inreda’s real-world data showed around 80% time in range among established users who’d chosen to keep using the system. DARE found something rather different when new users were randomised to it, and half of them stopped during the year. The glucagon change and the different outcome sensors complicate the comparison, so DARE doesn’t show that self-selection caused the gap. It does show the risk of assuming that observational results from one group will carry over to people starting the system.

Questions I’d like answered

• Inreda refers to an internal report on the effect of changing glucagon. Could it be published, with any pharmacokinetic or clamp data comparing Glugon with GlucaGen?
• Could DARE’s outcomes be analysed by the software version each person was actually running, so we can see whether the updates helped?
• Could the paired Guardian and Libre Pro IQ data from the dual-hormone arm be reported, to show how much the result depends on which sensor measures it?
• Of the people who stopped because of device burden, how many were troubled mainly by calibration and the second sensor?
• Inreda says a generic GlucaGen has been submitted for registration and a new trial is planned. Will it compare against current hybrid closed loop on an independent blinded sensor, as DARE did?

What the next trial needs to show

The Inreda AP set out to take meal management out of people’s lives, and the overnight results suggest it can do a lot. I don’t think the next trial needs to beat hybrid closed loop on time in range, because the point of the system is that you don’t announce meals. Time in range and safety no worse than hybrid closed loop, with people still using it at a year and a measurable reduction in day-to-day burden, would be a real result. Showing that properly means a non-inferiority trial with the margin set in advance and enough people to test it. I’ll be interested to see what it finds.

Notes for the specialists

These are the details behind the summary above, for anyone who wants to check them.

Pooled against per-cohort analysis. ZIN assessed the two cohorts together: 65.7% against 62.8%, a difference of 2.4 points (−0.7 to 5.5), against the 5 points it treats as clinically relevant. That pooled figure is weighted towards the hybrid closed loop cohort, which is more than twice the size. The protocol specified and powered each cohort separately, for 7 points against hybrid closed loop and 15 against injections, so the result against injections stands in the trial’s own terms, though its interval runs down to 1.0 point.

People who carried on. The per-protocol analysis gives 63.9% and 64.9%, slightly below intention to treat, and a post hoc analysis excluding measurements from people who stopped gives −0.9 points against hybrid closed loop (−5.1 to 3.3). Those who kept using it didn’t do better on average, although a benefit of around 3 points isn’t ruled out.

No recovery over the year. Dual-hormone time in range drifted from 67.8% and 68.1% at 3 months to 64.8% and 66.1% at 12 months. About 55% of the arm received the December 2024 update and 27 people the new infusion set, at different dates, so the aggregate can’t show whether either helped.

Safety. The risk ratio for serious adverse events was 2.95 (1.21 to 7.23); four of the 17 people affected had a level 3 hypo. All nine severe hypos in the dual-hormone arm were in people on the Inreda AP in the cohort compared against hybrid closed loop, seven people in all, between February and October 2024, before the December update. Two involved glucagon delivery failing. The update column in Table S18 lists updates that postdate several events, so it probably records what people received over the whole trial.

Nights. Overnight time below 3.9 mmol/L was 5.0% in the injections cohort (3.6% on injections) and 4.1% in the hybrid cohort (4.3% on hybrid closed loop), and three severe hypos happened at night. Table S10 includes only people with Libre Pro IQ data (30 of 37 and 70 of 82 in the dual-hormone arms).

Glucagon dose. DARE reports 71 to 86 units a day at every visit (Table S14), and the protocol treats more than 70 as a criterion for switching insulin to Lyumjev. I couldn’t find a definition of a unit. If it’s 10 micrograms, that’s 0.71 to 0.86 mg a day, close to Inreda’s users on Glugon (0.82 mg) and above those on GlucaGen.

The Jansen denominator. 60 of the 102 stayed on GlucaGen, which leaves 42, yet the switch comparison has 51. Time below range in that comparison rose from 1.83% to 2.34% and the share above 70% time in range fell from 92.2% to 74.5%. No interval is given for the 3.5-point drop.

Timeline. Protocol versions 2 to 4 (June to December 2022) name GlucaGen; version 5 (4 May 2023) changes the description to “Glucagon USP, glucagonhydrochloride”. Enrolment began in October 2023. That suggests the change was being considered months before the trial, not that Glugon had been chosen. I found no published formulation for Glugon.

The scenario’s assumptions. It assumes an observational drop in experienced users applies equally to new users on different software, and that it adds to a baseline-adjusted randomised difference. It uses 57% and 48% still on the system at 12 months.

Glucagon’s expected contribution. A 2022 meta-analysis put the benefit of adding glucagon at 2.7 points (−0.4 to 5.8). Those figures describe glucagon working as its algorithm expects, which is the thing in question here.

Did the Libre Pro IQ read low? A GMI calculated from mean sensor glucose matched laboratory HbA1c within 0.2 points in every arm, where a 14.5% low bias would have put it about 0.7 points below. So there’s no sign of a large uniform low bias on average. The populations don’t fully overlap, and a correct mean doesn’t guarantee a correct time in range near the edges of the range.

My own sensors. From August to October my AAPS loop ran on a G7 and on Libre 2 Plus and 3 Plus sensors in turn, with both recorded throughout. Time in range on the two differed by 1.1 to 5.2 points depending on the period and on which sensor drove the loop, and not consistently in the direction an own-sensor effect predicts. That’s one person on a different system and a different Libre, so it shows only that two sensors can disagree by several points.

Sensor wear. The protocol counts a 14-day blinded wear as valid with 168 hours of data. How much the 12-month wears actually had isn’t reported.

Quality of life. EQ-VAS differed by 2.8 points (−0.9 to 6.5) against ZIN’s threshold of 10. Fear of hypos was lower on the Inreda AP in the hybrid cohort and distress lower in the injections cohort, but none of the measures captures the work of managing meals.

Sources

• Jancev M, et al. Dual-hormone fully closed-loop therapy versus multiple daily insulin injections or hybrid closed-loop therapy in adults with type 1 diabetes. Lancet Reg Health Eur 2026;101872. doi:10.1016/j.lanepe.2026.101872, including its supplementary appendix and protocols.
• Zorginstituut Nederland. Standpunt bihormonale kunstalvleesklier bij volwassen patiënten met diabetes type 1, 22 September 2026.
• Inreda Diabetic. Reactie op het standpunt van Zorginstituut Nederland.
• Jansen TJP, et al. Three-year real-world performance of bihormonal fully closed-loop therapy in type 1 diabetes. Diabetes Technol Ther 2026. doi:10.1177/15209156261492793
• Zeng B, et al. Dual-hormone artificial pancreas for glucose control in type 1 diabetes: a meta-analysis. Diabetes Obes Metab 2022;24:1967 to 1975. PubMed