Skip to content
type1.science

End goal 01 — the artificial pancreas

Road to the artificial pancreas

From today's reactive, half-blind loops to a fully closed system as effective as a healthy pancreas.

This roadmap is the site’s spine: where we are today, the gaps stopping us, the research that would close them, and the questions still open. Every item and trial links back to the gap it advances.

  1. Sensing: lag, accuracy, and missing analytes

    CGMs read glucose in the fluid between cells, not blood, so they run roughly 10–15 minutes behind reality. An automated system is always reacting to the recent past — and it still can't see ketones at all.

    Editorial estimate of progress60%

    Where we are today

    Modern sensors are accurate enough for dosing (MARD around 8–9%) and factory-calibrated, but interstitial lag remains, accuracy degrades exactly where it matters most (the hypo range), and continuous ketone sensing — the key safety net for pump failures and illness — is only now entering trials.

    What would close it

    • Sensing chemistry and placement that shrinks interstitial lag
    • Lag-compensation and trend-projection algorithms that are safe near the hypo boundary
    • Continuous multi-analyte sensing — ketones first, then lactate and beyond
    • Redundant or dual-sensor configurations that catch sensor error before the algorithm acts on it

    Open questions

    • How low can interstitial lag physically go — and is intravascular or alternative-site sensing viable long-term?
    • Can lag compensation be made provably safe when glucose is falling fast?
    • What accuracy in the 54–70 mg/dL (3.0–3.9 mmol/L) band is "good enough" for fully unannounced automation?
    Read the full picture →

    Every automated insulin delivery system is built on one input: the CGM. But a CGM doesn't measure blood glucose — it measures glucose in the interstitial fluid between cells. The physiological blood-to-interstitial lag is only ~5–6 minutes,1 but the delay a system actually sees is larger: the total delay is the sum of that physiological lag, the sensor's own electrochemical reaction time, and the signal smoothing built in to produce stable traces.2 When glucose moves fast (a meal, a sprint, a stress spike), the system is reacting to where you were, not where you are.

    Lag compounds every other weakness in the loop. The algorithm must predict further into the future to compensate, and prediction is exactly what today's context-starved systems are worst at. Shrinking sensor lag directly shrinks how clever the rest of the system has to be.

    The frontier has a second front: what we sense, not just how fast. Diabetic ketoacidosis is a life-threatening complication most common in type 1 diabetes, and an international expert consensus has recommended a continuous ketone monitor — ideally one combining CGM technology with β-hydroxybutyrate sensing in a single device.3 Such a sensor would let closed loops and their users catch failed infusion sites and illness before they become DKA. A single-sensor continuous dual glucose-ketone system is now in development,4 and we rank this class as a first-class sensing upgrade, not an accessory.

    References

    1. Direct microdialysis measurement in healthy adults put the physiological blood-to-interstitial glucose lag at 5–6 minutes. Basu A, et al. "Time Lag of Glucose From Intravascular to Interstitial Compartment in Humans." Diabetes 62(12):4083–7 (2013). https://doi.org/10.2337/db13-1132 ↩

    2. The delay between a true blood-glucose change and the value a monitor displays is the sum of the interstitial-to-plasma lag, the sensor's electrochemical delay, and front-end signal-processing (smoothing) delays. Keenan DB, et al. "Delays in Minimally Invasive Continuous Glucose Monitoring Devices: A Review of Current Technology." J Diabetes Sci Technol 3(5):1207–14 (2009). https://doi.org/10.1177/193229680900300528 ↩

    3. DKA is a life-threatening complication most common in type 1 diabetes; the authors note an international consensus recommending development of a continuous ketone monitor combining CGM with 3-β-hydroxybutyrate sensing in one device. Virdi N, et al. "Prevalence, Cost, and Burden of Diabetic Ketoacidosis." Diabetes Technol Ther 25(S3):S75–S84 (2023). https://doi.org/10.1089/dia.2023.0149 ↩

    4. Abbott Diabetes Care is developing a continuous dual glucose-ketone system that uses a single sensor to measure interstitial glucose and β-hydroxybutyrate. Durnwald C, Polsky S. "Continuous Dual Glucose-Ketone Monitoring." Diabetes Technol Ther 27(S4):S51–S55 (2025). https://doi.org/10.1089/dia.2025.0300 ↩

    Our take

    Sensing is the most-solved of the artificial-pancreas gaps — but "most solved" still means every loop today drives while looking 10 minutes into the past. Ketone sensing is the next must-have, not a luxury.

    Items working on itAbbott Libre DuoAbbott LingoConcept: continuous glucose + ketone sensorDexcom G7Dexcom G7 15-DayDexcom ONE+Dexcom SteloFreeStyle Libre 2 PlusFreeStyle Libre 3 PlusFreeStyle Libre RioMedtronic Simplera (standalone)Non-invasive glucose monitoring (Apple, Samsung, Know Labs, Afon Glucowear)Senseonics Eversense 365
    Trials advancing itArgus 2.0 CGM adoption studyCareSens Air 3 CGM accuracy and precisionCKM for safer SGLT2 inhibitor use in type 1 diabetesContinuous glucose-ketone monitor feasibility during pump suspensionEversense Gemini flash glucose measurement studyEversense non-adjunctive use post-approval studyLabPatch glucose sensing system accuracyROME GS continuous glucose monitoring system studyAiDAPT: Automated Insulin Delivery in Pregnancy with Type 1 Diabetes

  2. Insulin speed: peak effect arrives ~1.5–2 hours late

    Even the fastest injectable insulins begin working in ~15–20 minutes but don't reach peak glucose-lowering effect for ~1.5–2 hours — and keep working for 3–5 hours after. A healthy pancreas responds in minutes and stops on a dime. This is the loop's heaviest physical lag.

    Editorial estimate of progress35%

    Where we are today

    Ultra-rapid analogs (faster aspart, ultra-rapid lispro) shaved minutes off onset, not the fundamental curve: subcutaneous insulin must still dissolve, dissociate, and absorb. Inhaled insulin peaks far sooner but is coarse-dosed and not loop-integrated. Truly fast, truly short insulin remains the single biggest unlock for unannounced-meal automation.

    What would close it

    • Ultra-fast formulations and excipients that speed monomer absorption
    • Alternative routes — inhaled, intradermal, intraperitoneal — with loop-compatible dosing
    • Co-formulations and adjuncts (e.g. amylin analogs) that blunt the post-meal spike insulin is too slow to catch
    • Glucose-responsive ("smart") insulins that activate only when glucose is high

    Open questions

    • Can a subcutaneous insulin ever reach a sub-20-minute peak, or is a different route required?
    • How do we make ultra-fast insulin stable enough for pump reservoirs and warm bodies?
    • What does dosing safety look like when insulin acts in minutes and stacking windows collapse?
    Read the full picture →

    A healthy pancreas releases insulin straight into the portal bloodstream within a minute or two of glucose rising — and stops releasing it just as fast. Injected insulin enters fat under the skin, where the rate of absorption from the subcutaneous space into the capillaries is itself the limiting step.1 The practical result, even with the newest "ultra-rapid" analogs: meaningful action begins within roughly the first half hour, the peak glucose-lowering effect lands around 1.5–2 hours later, and it keeps working for about 3–5 hours.2

    Ultra-rapid formulations have narrowed the gap, not closed it. In a head-to-head euglycemic clamp, faster aspart's onset of glucose-lowering action was only about nine minutes ahead of the next-generation candidate it was measured against — the underlying subcutaneous curve still dominates.3

    For an automated system, that tail is as costly as the delay. The algorithm can't issue a strong correction without committing glucose-lowering pressure hours into a future it can't see — so it corrects timidly, and the burden lands back on the person to pre-bolus and announce meals. Unannounced meals remain the single greatest obstacle to fully closed-loop control precisely because of the slow pharmacodynamics of subcutaneous insulin relative to how fast a meal raises glucose.4

    Every path that compresses this curve — better formulations, different routes into the body, insulin that switches itself on only when glucose is high — attacks the artificial pancreas's heaviest physical lag. We rank insulins with time-to-peak weighted highest for exactly this reason.

    References

    1. "One of the primary limiting factors for exogenous insulin response is the rate of absorption into the capillaries from the subcutaneous space"; rapid analogs' time-action "still falls short of the normal physiological response to meal consumption." Heise T, et al. "What is the value of faster acting prandial insulin? Focus on ultra rapid lispro." Diabetes Obes Metab 24(9):1689–1701 (2022). https://doi.org/10.1111/dom.14773 ↩

    2. For faster aspart (Fiasp), reported onset is ~20–30 minutes, the peak glucose-lowering effect occurs ~1.5–2.2 hours after dosing, and duration is ~5 hours. Reviewed in "Ultra-Rapid-Acting Insulins: How Fast Is Really Needed?" Clinical Diabetes 39(4):415–423 (2021). https://pmc.ncbi.nlm.nih.gov/articles/PMC8603316/ ↩

    3. In a randomized double-blind crossover clamp in men with type 1 diabetes, onset of action for the AT247 candidate was ~9 minutes earlier than faster aspart (and ~23 minutes earlier than insulin aspart) — small absolute shifts on a multi-hour curve. Svehlikova E, et al. Diabetes Care 44(2):448–455 (2021). https://doi.org/10.2337/dc20-1017 ↩

    4. "While many disturbances stand in the way of fully closed-loop (automated) control, unannounced meals remain the greatest challenge … the slow pharmacodynamics of subcutaneous injection insulin relative to meal dynamics." Diamond T, Cameron F, Bequette BW. "A New Meal Absorption Model for Artificial Pancreas Systems." J Diabetes Sci Technol 16(1):40–51 (2021). https://doi.org/10.1177/1932296821990111 ↩

    Our take

    This is where we'd point new money. Faster sensing helps the algorithm see; faster insulin is the only thing that lets it *act*. A true sub-20-minute insulin would make today's "pretty good" algorithms feel superhuman overnight.

    Items working on itConcept: sub-20-minute glucose-responsive mealtime insulinFaster aspart (Fiasp)Glucose-responsive ("smart") insulinInhaled insulin (Afrezza)Insulin aspart (NovoLog / NovoRapid)Insulin glulisine (Apidra)MiniMed Flex (formerly NMX8-AID)Ultra-rapid lispro (Lyumjev)
    Trials advancing itAdjunct therapy to reduce iatrogenic hyperinsulinemia in T1DAI bolus priming added to AIDANETAIDANET at Home: fully closed-loop vs hybrid modesCIR-0602K insulin sensitizer for people using AIDGATEWAY: MiniMed Flex / NMX8-AID with Simplera SyncINHALE-1st: Afrezza for youth with newly diagnosed T1DLong-term tirzepatide safety in T1D with overweight or obesityMedtronic implantable insulin pump system (MIIPS 2020)MiniMed Fit Payload Wear StudyOnce-weekly insulin icodec phase 3 in T1DQualitative meal-size estimation vs carb counting in AIDReal-world insulin icodec use in type 1 diabetesSemaglutide phase 3 in T1D with obesitySemaglutide to reduce carb-counting burden on closed loopSEMPA: semaglutide plus empagliflozin added to AIDTirzepatide phase 3 in T1D with overweight or obesityTirzepatide with Control-IQ to reduce meal announcementsTriple therapy in T1DUltra-rapid lispro timing in MiniMed 780GInreda AP fully closed-loop dual-hormone artificial pancreas (DARE trial)

  3. The context gap: predicting blind

    Blood glucose is influenced by dozens of variables — at least 42 by one widely-cited count: food, exercise, stress, sleep, hormones, illness, heat, altitude and more. Today's systems see a handful, mostly typed in by hand. Predictions are inaccurate because the system is guessing with its eyes closed.

    Editorial estimate of progress15%

    Where we are today

    Commercial loops see CGM trend, insulin on board, and whatever the user manually announces. A few consume step counts or activity flags. Nobody ships automatic meal detection good enough to dose on, and nobody fuses the wearable signals — heart rate, HRV, temperature, movement, sleep — that already exist on millions of wrists.

    What would close it

    • Automatic meal detection and estimation (CGM-pattern, wearable, and camera/vision approaches)
    • Fusing existing wearable biomarkers — heart rate, HRV, skin temperature, accelerometry, sleep staging — into dosing models
    • On-device multimodal models that turn raw life-signals into glucose-relevant context
    • Personal "digital twin" simulation for testing context-driven dosing safely

    Open questions

    • Which of the 40-plus variables carry real predictive power, and how much context is enough for unannounced meals?
    • Can context capture be made effortless AND private — everything processed on-device, nothing uploaded?
    • How does a regulator evaluate a system whose inputs include a camera or an LLM's interpretation of your day?
    Read the full picture →

    Ask anyone with T1D what moves their glucose and you'll get a list the loop never sees: the meeting that ran long, the flu coming on, the hot bath, the third coffee, the anxiety before a flight. One widely-shared framework counts at least 42 distinct factors.1 A commercial AID system sees only a handful of inputs — and the most important one (food) only if it's typed in by hand, in advance: today's hybrid closed-loop systems still require the user to announce meals or count carbohydrates, and removing that input remains an open future challenge.2

    This is why predictions miss. It isn't that the algorithms are dumb; it's that they're blind. The control theory is decades old and solid — given good inputs. The inputs are the problem. Fully closed-loop systems that drop the meal announcement have been built, but in early studies they carried a higher risk of hypoglycemia than algorithms that still used announced meals — the system needs more context, not less.3

    The fix is context: captured automatically, interpreted continuously, and — this is the hard constraint — processed entirely on-device. Wearables already stream heart rate, temperature, and movement, and early multivariable systems have used wristband physiological signals, analyzed by AI, to detect activity and adjust insulin without the user announcing anything.4 On-device models can reason over all of it, 24/7, without a single byte leaving the person. That is the version of this future worth building; we treat any cloud-dependent design as a privacy failure, not a shortcut.

    References

    1. Brown A. "42 Factors That Affect Blood Glucose?! A Surprising Update." diaTribe — a widely-shared enumeration spanning food, activity, and biological/environmental categories. https://diatribe.org/diabetes-management/42-factors-affect-blood-glucose-surprising-update ↩

    2. Templer S. "Closed-Loop Insulin Delivery Systems: Past, Present, and Future Directions." Front Endocrinol (2022) — notes that commercial hybrid systems still rely on user meal announcement/carb counting, and that fully closed-loop operation without user input is an open challenge. https://doi.org/10.3389/fendo.2022.919942 ↩

    3. Cameron FM, et al. "Closed-Loop Control Without Meal Announcement in Type 1 Diabetes." Diabetes Technol Ther (2017) — fully closed-loop control without announced meals carried higher hypoglycemia risk than algorithms using meal announcement. https://doi.org/10.1089/dia.2017.0078 ↩

    4. Askari MR, et al. "Multivariable Automated Insulin Delivery System for Handling Planned and Spontaneous Physical Activities." J Diabetes Sci Technol (2023) — wristband physiological signals analyzed by AI to detect physical activity and modulate insulin dosing without manual announcement. https://doi.org/10.1177/19322968231204884 ↩

    Our take

    The least-solved gap, and the one we think on-device AI changes fastest. Privacy is non-negotiable here: the only acceptable version of a context-aware loop is one where the context never leaves your pocket.

    Items working on itBeta Bionics iLet Bionic PancreasMiniMed Flex (formerly NMX8-AID)SOOIL Dana Diabecare RS
    Trials advancing itAdjunct therapy to reduce iatrogenic hyperinsulinemia in T1DAI bolus priming added to AIDANETAID before and during pregnancy in T1DAID beyond glucose metricsAIDANET at Home: fully closed-loop vs hybrid modesAIMING: AID during labor and delivery in T1DArgus 2.0 CGM adoption studyCareSens Air 3 CGM accuracy and precisionCommercial or open-source closed-loop impact on pregnancyContinuous ketone monitoring for safer sotagliflozin in T1D with CKDDapagliflozin plus pioglitazone in T1DDBLG1 closed-loop PROMs and glycemic controlEarly AID pilot in newly diagnosed T1DEarly AID therapy outcomes and cost-effectivenessEmpagliflozin added to AID systemsEversense non-adjunctive use post-approval studyExtended-wear infusion set options at homeFCL@Home: AIDANET fully closed-loop feasibilityGATEWAY: MiniMed Flex / NMX8-AID with Simplera SyncHybrid closed-loop effectiveness in adultsiPancreas context-aware fully closed-loop AIDLabPatch glucose sensing system accuracyLong-term tirzepatide safety in T1D with overweight or obesityLow-dose glucagon before exercise while using AIDLuna AID system safety and effectivenessMedtronic implantable insulin pump system (MIIPS 2020)MiniMed 780G in T1D with gastroparesisMTX228 adaptive phase 2 study in T1DNext-generation AID algorithm in adults with T1DOpen-source artificial pancreas use in ChinaQualitative meal-size estimation vs carb counting in AIDROME GS continuous glucose monitoring system studySafe ketone thresholds for dapagliflozin in T1DSemaglutide phase 3 in T1D with obesitySEMPA: semaglutide plus empagliflozin added to AIDSimplified meal boluses vs carb counting in adolescents using HCLSotagliflozin plus volagidemab adjunct combinationThree-day and seven-day infusion set trialTirzepatide phase 3 in T1D with overweight or obesityTriple therapy in T1DADAPT: MiniMed 780G advanced hybrid closed-loop vs injections + flash CGMAiDAPT: Automated Insulin Delivery in Pregnancy with Type 1 DiabetesBihormonal iLet Bionic Pancreas (insulin + glucagon) feasibility studyCamAPS FX hybrid closed-loop in very young children (1-7y): the KidsAP02 crossover trialControl-IQ closed-loop in children 6–13 (DCLP5 pediatric pivotal)Control-IQ pivotal trial (iDCL): t:slim X2 closed-loop vs. sensor-augmented pumpCREATE: open-source automated insulin delivery vs sensor-augmented pumpELSA: EarLy Surveillance for Autoimmune diabetes (UK)FLAIR: MiniMed 670G vs. advanced hybrid closed-loop (the future 780G algorithm) in adolescents and young adultsiLet Bionic Pancreas (insulin-only) pivotal trialInreda AP fully closed-loop dual-hormone artificial pancreas (DARE trial)Omnipod 5 (Horizon) pivotal single-arm AID trialOmnipod 5 preschool cohort (ages 2-5.9)TrialNet Pathway to Prevention

  4. Algorithms that learn you

    Today's controllers are competent but generic and mostly static: tuned at setup, adjusted at clinic visits. The settings that drive your loop — basal rates, ratios, sensitivities — drift with your life, and nothing retunes them continuously.

    Editorial estimate of progress40%

    Where we are today

    Commercial algorithms adapt slowly (if at all) and hide their reasoning. The open-source community is far ahead on adaptability — automatic sensitivity detection, unannounced-meal handling, dynamic ratios — because the algorithms are open and tunable. Continuous, personal, explainable optimization remains nobody's shipped product.

    What would close it

    • Adaptive controllers that retune basal/ratio/sensitivity continuously and safely
    • On-device LLM/ML co-pilots reasoning over rich context 24/7 and proposing (or making) settings changes — especially on open DIY systems
    • Formal safety envelopes so a learning controller can never propose a dangerous dose
    • Explainability: showing the person why the system did what it did

    Open questions

    • What is the regulatory pathway for a controller that keeps learning after approval?
    • How do you prove safety of per-person adaptation rather than population-average behavior?
    • Where is the line between the algorithm dosing and the algorithm advising?
    Read the full picture →

    Every loop runs on settings — basal rates, carb ratios, sensitivity factors — that encode you. And you change: training load, hormones, stress, seasons, age. In most systems those numbers are updated a few times a year, by hand, from gut feel. Today's commercial controllers are not fully automated either: they still require the person to announce meals and exercise to avoid dangerous highs and lows.1 The controller is only as good as its stalest setting.

    The open-source systems showed what adaptability buys: automatic sensitivity adjustment, dynamic ratios, meal handling without announcements. In the randomized CREATE trial, an open-source loop raised time-in-range to 71.2% versus 54.5% on sensor-augmented pump therapy — about 3 hours 21 minutes more in range every day.2 The next step is a system that watches weeks of your data and your context and continuously proposes — or makes — the retunes a great endocrinologist would, if they could watch you every hour of every day. The research field now frames this explicitly as a job for artificial intelligence and adaptive control, with online-learning controllers and digital-twin models as the leading approaches.1 3

    That is a job for on-device intelligence: a model that never sleeps, never uploads, and can explain every suggestion. Pair it with an open algorithm whose knobs it's allowed to turn, wrap it in a hard safety envelope, and the "static settings" era ends.

    References

    1. A 2025 expert review notes current AID systems "are not fully automated in that they require the person using the system to announce meals and exercise," perform best overnight, and concludes that next-generation systems "will need to support all people... by leveraging the latest technologies in artificial intelligence and adaptive control." Jacobs PG, et al. "Research Gaps, Challenges, and Opportunities in Automated Insulin Delivery Systems." Journal of Diabetes Science and Technology 19(4):937–949 (2025). https://doi.org/10.1177/19322968251338754 ↩ ↩2

    2. In the randomized CREATE trial (97 children and adults), an open-source loop (the OpenAPS algorithm in AndroidAPS, paired with a DANA-i pump and Dexcom G6) raised time-in-range to 71.2% vs 54.5% with sensor-augmented pump therapy — about 3 h 21 min more per day in range — with no severe hypoglycemia or diabetic ketoacidosis in either group. Burnside MJ, et al. "Open-Source Automated Insulin Delivery in Type 1 Diabetes." New England Journal of Medicine 387:869–881 (2022). https://doi.org/10.1056/NEJMoa2203913 ↩

    3. Review of AI-enabled artificial-pancreas systems that combine mechanistic physiological models with data-driven models to build per-person "digital twins" driving automated insulin delivery, and that highlight online-learning and adaptive fault-tolerant controllers as future directions. Ahmadasas M, et al. "Cyber-Physical-Human Systems in Precision Medicine: Advances in Artificial Pancreas for Treatment of Diabetes." Annual Reviews in Control 60:101033 (2025). https://doi.org/10.1016/j.arcontrol.2025.101033 ↩

    Our take

    The cheapest gap to attack right now: it's software. DIY loops are the natural laboratory — open algorithms plus on-device AI optimizing settings is buildable today, and we think it's the fastest route to closing the prediction half of the lag problem.

    Items working on itBeta Bionics iLet Bionic PancreasLoop (iOS)Medtronic MiniMed 670G (first hybrid closed loop)MiniMed Flex (formerly NMX8-AID)SOOIL Dana Diabecare RS
    Trials advancing itAI bolus priming added to AIDANETAID before and during pregnancy in T1DAID beyond glucose metricsAIDANET at Home: fully closed-loop vs hybrid modesAIMING: AID during labor and delivery in T1DCadisegliatin as an adjunct for hybrid closed-loop usersCIR-0602K insulin sensitizer for people using AIDCommercial or open-source closed-loop impact on pregnancyDBLG1 closed-loop PROMs and glycemic controlEarly AID pilot in newly diagnosed T1DEarly AID therapy outcomes and cost-effectivenessEmpagliflozin added to AID systemsFCL@Home: AIDANET fully closed-loop feasibilityGATEWAY: MiniMed Flex / NMX8-AID with Simplera SyncHybrid closed-loop effectiveness in adultsiLet Experience real-world safety and effectiveness studyiPancreas context-aware fully closed-loop AIDLuna AID system safety and effectivenessMiniMed 780G in T1D with gastroparesisNext-generation AID algorithm in adults with T1DOpen-source artificial pancreas use in ChinaQualitative meal-size estimation vs carb counting in AIDRENEW: Medtrum APGO real-world AID studySemaglutide to reduce carb-counting burden on closed loopSEMPA: semaglutide plus empagliflozin added to AIDSimplified meal boluses vs carb counting in adolescents using HCLSTRIVE 2: Omnipod 6 compared with Omnipod 5Tirzepatide as an adjunct to automated insulin deliveryTirzepatide with Control-IQ to reduce meal announcementsUltra-rapid lispro timing in MiniMed 780GADAPT: MiniMed 780G advanced hybrid closed-loop vs injections + flash CGMAiDAPT: Automated Insulin Delivery in Pregnancy with Type 1 DiabetesBihormonal iLet Bionic Pancreas (insulin + glucagon) feasibility studyCamAPS FX hybrid closed-loop in very young children (1-7y): the KidsAP02 crossover trialControl-IQ closed-loop in children 6–13 (DCLP5 pediatric pivotal)Control-IQ pivotal trial (iDCL): t:slim X2 closed-loop vs. sensor-augmented pumpCREATE: open-source automated insulin delivery vs sensor-augmented pumpFLAIR: MiniMed 670G vs. advanced hybrid closed-loop (the future 780G algorithm) in adolescents and young adultsiLet Bionic Pancreas (insulin-only) pivotal trialInreda AP fully closed-loop dual-hormone artificial pancreas (DARE trial)MiniMed 670G pivotal in-home hybrid closed-loop studyOmnipod 5 (Horizon) pivotal single-arm AID trialOmnipod 5 preschool cohort (ages 2-5.9)

  5. No brakes: single-hormone loops can only stop pushing

    A healthy pancreas has two pedals — insulin to lower glucose, glucagon to raise it. Insulin-only loops can ease off but never actively push back, so they run conservatively high and still can't fully prevent exercise lows.

    Editorial estimate of progress30%

    Where we are today

    One dual-hormone system is CE-marked and in real-world use in Europe; others paused on a hard problem — glucagon that stays stable in a pump for days. Meanwhile insulin-only systems defend against lows by suspending early and targeting higher than a healthy pancreas would.

    What would close it

    • Pump-stable glucagon and glucagon analogs
    • Dual-chamber pump hardware that doesn't double the wear burden
    • Micro-dose glucagon strategies for exercise and overnight protection
    • Amylin co-delivery to slow meal spikes while glucagon guards the floor

    Open questions

    • Is chronic micro-dose glucagon safe over years? (glycogen stores, nausea, long-term signaling)
    • Does a second hormone beat simply having much faster insulin — or do we need both?
    • Can dual-hormone systems ever match single-hormone cost and simplicity?
    Read the full picture →

    Watch an insulin-only loop fight a falling glucose and you see the problem: all it can do is stop. Suspend basal, cut the correction, and wait — while insulin already delivered keeps working for hours. The person eats glucose tabs; the algorithm apologizes.

    A healthy pancreas never plays this one-handed. Glucagon is the counter-force, released within minutes when glucose dips. Putting it (or a stable formulation) into the loop gives the controller an actual brake pedal — which in randomized trials cut time spent low versus insulin-only, including around exercise.1 In real-world use, a CE-marked bihormonal fully closed-loop system held median time below range near 1% while running fully automated, without meal or exercise announcements.2

    The blocker has been chemistry as much as engineering: native glucagon is poorly soluble and degrades in aqueous solution, so a second reservoir needs either frequent refills or a stabilized formulation. Liquid-stable glucagon has been shown to work inside a closed loop, though nausea remains a real side effect for some users.3 Solve the formulation and side-effect burden, shrink the hardware, and bihormonal stops being a niche — it becomes a strong candidate for the endgame for anyone whose deepest fear is the overnight low.

    References

    1. In a randomized outpatient crossover trial, adding glucagon to the loop produced the lowest time in hypoglycemia of all arms during the exercise period and across the study, compared with single-hormone closed-loop and predictive low-glucose suspend. Castle JR, et al. "Randomized Outpatient Trial of Single- and Dual-Hormone Closed-Loop Systems That Adapt to Exercise Using Wearable Sensors." Diabetes Care (2018). DOI ↩

    2. Real-world 1-year results for the CE-marked Inreda AP (insulin + glucagon, fully closed-loop) reported mean time in range of 80.3% and median time below range of 1.36% across participants who completed the year. van Bon AC, Blauw H, et al. "Bihormonal fully closed-loop system for the treatment of type 1 diabetes: a real-world multicentre, prospective, single-arm trial in the Netherlands." Lancet Digital Health (2024). DOI ↩

    3. A dual-hormone system using a novel liquid-stable glucagon formulation demonstrated feasibility in a closed loop and reduced hypoglycemia during and after exercise versus insulin-only; however, four participants experienced glucagon-related nausea and three withdrew. Wilson LM, et al. "Dual-Hormone Closed-Loop System Using a Liquid Stable Glucagon Formulation Versus Insulin-Only Closed-Loop System Compared With a Predictive Low Glucose Suspend System." Diabetes Care (2020). DOI ↩

    Our take

    Underrated. Hypo fear is the reason people and algorithms alike run high — it is the tax on every aggressive dosing decision. A reliable chemical floor under glucose changes the psychology of automation, not just the numbers.

    Items working on itMedtronic Guardian 4Medtronic MiniMed 670G (first hybrid closed loop)
    Trials advancing itAID before and during pregnancy in T1DCadisegliatin phase 3 adjunct to insulinCKM for safer SGLT2 inhibitor use in type 1 diabetesCommercial or open-source closed-loop impact on pregnancyContinuous glucose-ketone monitor feasibility during pump suspensionDapagliflozin plus pioglitazone in T1DiLet Experience real-world safety and effectiveness studyiPancreas context-aware fully closed-loop AIDKetone monitoring to reduce DKA risk during adjunct SGLT2 inhibitionLow-dose glucagon before exercise while using AIDSotagliflozin plus volagidemab adjunct combinationSTRIVE 2: Omnipod 6 compared with Omnipod 5ZT-01 to prevent nocturnal hypoglycemiaADAPT: MiniMed 780G advanced hybrid closed-loop vs injections + flash CGMAiDAPT: Automated Insulin Delivery in Pregnancy with Type 1 DiabetesBihormonal iLet Bionic Pancreas (insulin + glucagon) feasibility studyCamAPS FX hybrid closed-loop in very young children (1-7y): the KidsAP02 crossover trialControl-IQ closed-loop in children 6–13 (DCLP5 pediatric pivotal)Control-IQ pivotal trial (iDCL): t:slim X2 closed-loop vs. sensor-augmented pumpCREATE: open-source automated insulin delivery vs sensor-augmented pumpFLAIR: MiniMed 670G vs. advanced hybrid closed-loop (the future 780G algorithm) in adolescents and young adultsiLet Bionic Pancreas (insulin-only) pivotal trialInreda AP fully closed-loop dual-hormone artificial pancreas (DARE trial)Lantidra (donislecel): purified allogeneic islet cell therapy (CIT-07)MiniMed 670G pivotal in-home hybrid closed-loop studyOmnipod 5 (Horizon) pivotal single-arm AID trialOmnipod 5 preschool cohort (ages 2-5.9)Sernova Cell Pouch: implantable islet-transplant scaffold