Artificial Pancreas 2.0: What Automatic Insulin Delivery Systems Can't Do Yet — and How to Fix It

Diabetes Technology & Therapeutics published a review by an international group of engineers and clinicians about the gaps that prevent automated insulin delivery systems (AID) from becoming a truly “fully closed loop.” The authors honestly state that current devices reduce HbA1c, improve quality of life, and manage sugar more safely - but they work best at night, and during the day they require the user to declare meals and physical activity to avoid hyper- and hypoglycemia. In addition, many systems are not yet designed for pregnant women and the elderly. The review shows the results of new algorithms that recognize food and exercise automatically, and early data on the use of AID in “complex” groups. Key conclusion: the next round of evolution is artificial intelligence and adaptive control, including for multihormonal configurations (insulin ± glucagon).

Background of the study

Automated insulin delivery systems (AIDs) are a combination of a continuous glucose monitor (CGM), an insulin pump, and a control algorithm that adjusts insulin delivery in real time. In recent years, “hybrid” circuits have significantly reduced HbA1c, increased Time in Range, and reduced nocturnal hypoglycemia in people with type 1 diabetes. But “full autopilot” is not yet available: during the day, when glucose is constantly affected by food, stress, and movement, most systems still require manual carbohydrate input and an activity warning - otherwise the algorithm cannot compensate for rapid sugar spikes.

Clinical practice has shown other gaps. Algorithms work best during sleep, when the metabolism is more stable, but postprandial peaks, exercise, and bolus delays remain Achilles' heel. Some systems are not yet designed for pregnant women (different glycemic targets, high cost of errors) and the elderly (polymorbidity, increased risk of hypo), where adapted safety modes and interfaces that reduce cognitive load are needed.

Technically, the next frontier is to reduce the “human factor.” To this end, algorithms are being developed for automatic recognition of food intake and physical activity based on CGM patterns and wearable sensors; multihormonal circuits (insulin ± glucagon) are being tested as “insurance” against hypo; adaptive/AI models are being implemented that adjust to the user’s individual rhythms and the context of the day. In parallel, the industry needs interoperability and cybersecurity standards so that systems are updated “over the air,” and data is securely exchanged between devices and clinics.

Finally, it is not only sugar control that is important, but also life convenience: less anxiety and manual actions, stable sleep, accessibility of the technology for people with different levels of digital skills and income. Therefore, the “artificial pancreas 2.0” is not just a “faster” algorithm, but an ecosystem that works equally reliably day and night, requires a minimum of interventions and covers wide groups of patients.

Why is this important?

Automated circuits are one of the major breakthroughs in diabetology in recent decades, and their contribution is officially reflected in modern diabetes management standards. But “full autonomy” is still unattainable: the user still enters carbohydrates “manually,” and with an active lifestyle, algorithms are often late. The review systematizes where to move so that AIDs become more accessible and smarter - and for those who are pregnant, over 65, play sports, or simply cannot count carbohydrates every few hours.

What AID Can Do Now - and Where Progress Is Stalling

Today's hybrid "pancreases" are great at maintaining Time in Range (TIR) and reducing Time Below Range (TBR), especially during sleep. But during daytime "challenges" - food, stress, training - weak points emerge:

• Food/exercise announcements are required. Without them, the circuit does not have time to "catch" the postprandial surge or prevent hypo after activity.
• Limited "civilian" suitability. A number of systems are not intended for pregnant women and the elderly, where the goals and risks are different.
• Daytime instability. The devices are most effective at night; glucose levels vary more during the day.
• "Human Factor" - Carbohydrate counting and manual steps are tedious, making adherence difficult - this is emphasized by clinical reviews and practice.

What the authors of the review suggest

The researchers point out areas where encouraging results have emerged in recent years - and where efforts are needed:

• Automatic food and activity recognition. Algorithms that can, without user input, assess the fact and scale of food intake/exercise and dose insulin accordingly.
• Multihormonal circuits. Adding glucagon as a "safety pedal" against hypo is a separate branch of development.
• New target groups. Trials in the elderly and during pregnancy with adaptation of goals and protective barriers.
• AI and adaptive control: Personalized models that “learn” from everyday data remove some of the manual work and simplify access to the technology.

Where to look for developers and regulators

To bring AID to a “full loop” for everyone, in addition to algorithms, we will also have to solve “systemic” problems:

• Interoperability and updatable. Data exchange standards and secure remote software updates.
• "Real life" benefit metrics. In addition to HbA1c - TIR/TBR, alert burden, night sleep, user cognitive load.
• Access and fairness: Simplify the interface and make systems cheaper so that AIDs can be accessed by those who do not use them today.
• Cybersecurity and privacy. Especially in the context of increasingly smart and networked devices.

What this means for people with diabetes - now

Even without being “fully autonomous,” modern AIDs already provide benefits in sugar and safety — this is confirmed by randomized and observational studies. If you use a contour today, the main “life hack” is high engagement (timely announcements of food/loads, sensor charge/connectivity, correct setting of goals). And for those who are just considering an AID, the review gives a clear vector: in the coming generations, devices will require less manual actions and cope better with the day, and not just with the night.

Where are the boundaries and what's next?

This is a review - it does not replace clinical trials, but it sets the agenda: intellectualization of contours and expansion of indications. Home trials of systems that independently dose around food and load are already underway; multi-hormonal solutions are being developed in parallel. The next step is multi-center studies in the elderly, pregnant women, people with an "unpredictable" schedule, as well as work on accessibility and implementation.

A short cheat sheet: what prevents a “full loop” and what will bring it closer

It interferes with:

• the need for manual entry of carbohydrates and activity declarations;
• decreased stability during the day (food, sports, stress);
• lack of modes for pregnancy and the elderly in some systems.

Approximate:

• auto-detection of food/load and adaptive algorithms;
• multihormonal circuits (insulin ± glucagon);
• unified data standards, security, accessibility.

Conclusion

The review clearly formulates the goal of “version 2.0” for the artificial pancreas: to reduce the user’s role to a minimum, make the circuits work equally reliably day and night, and open access to those who are currently left behind – including pregnant women and the elderly. The path to this lies through AI algorithms, adaptive control, and multi-hormonal schemes – and there are already initial results that this is real. Now it’s up to clinical trials and engineers to turn these ideas into reliable devices “for everyone and every day.”

Research source: Jacobs PG et al. Research Gaps, Challenges, and Opportunities in Automated Insulin Delivery Systems. Diabetes Technology & Therapeutics 27(S3):S60-S71. https://doi.org/10.1089/dia.2025.0129