Until the cure: the closed loop, part by part
Pumps and sensors are a bridge until biology takes over. A map of the closed loop, from inputs to the human in it, and where each part is limited today.
Managing type 1 diabetes is a control loop: measure glucose, decide, deliver insulin, measure again. Every improvement in diabetes technology sits on one part of that loop. This page maps the parts, so it’s easier to see where today’s limits are, and why a biological cure would replace the whole thing.
The short version is sensor → algorithm → insulin. The full version adds the body being controlled, the model of that body, delays everywhere, imperfect sensors and pumps, the people in the loop, and reliability.
1. Inputs
Inputs come in two kinds: measurements of what is happening, and intent, meaning what is about to happen. Only intent can beat insulin’s delay, because it arrives before the event.
| Input | Kind | Quality today | Timing |
|---|---|---|---|
| Glucose (CGM) | Measurement | Good, but noisy and lagged | 5–15 min behind blood |
| Insulin delivered | Measurement | Good if delivery works | Real time |
| Meal carbohydrate | Intent / estimate | Hard; human estimates are the biggest error source | Should arrive before the meal |
| Fat and protein | Intent / estimate | Rarely entered | Effect delayed for hours |
| Exercise | Measurement + intent | Detectable from heart rate and movement | During and up to a day after |
| Ketones | Measurement | New: continuous ketone sensing approved in 2026 | Real time |
| Illness, stress, hormones, sleep | Context | Poorly measured | Hours to days |
| Growth, puberty, menstrual cycle | Context | Partly predictable | Days to years |
2. The body
The body is the least-known part of the loop. It’s individual, nonlinear and changes over time: the same dose and the same meal give different results on different days.
- Fast variation: time of day (the dawn effect), recent exercise, stress, illness.
- Slow variation: growth, puberty, weight, season, and in the first months after diagnosis, remaining beta-cell function.
- Nonlinearity: insulin sensitivity depends on the glucose level itself and on recent insulin.
3. The model of the body
The algorithm doesn’t see the body. It sees a model of it, written as settings: basal rates, insulin sensitivity factor, carbohydrate ratio, insulin action time and targets. These settings are usually the weakest link, which is why so much development goes into adapting them automatically (Autosens and Dynamic ISF in the open-source systems, self-learning commercial systems). Tuning still mostly happens by hand, at clinic visits, from reports.
4. The algorithm
Every few minutes the algorithm turns inputs, state and the body model into a decision. It forecasts glucose ahead and doses toward a target, within safety limits (maximum basal, maximum bolus, suspension on predicted lows). The open-source algorithms are now considered good enough that the bigger gains lie around them. The open question is the objective: the right trade-off for a child at school is not the same as for an adult athlete.
5. Outputs
Today the output is mostly rapid-acting insulin under the skin. The output has three dimensions: what substance (rapid or ultra-rapid insulin, pramlintide to slow meal absorption, glucagon, carbohydrate), what route (under the skin, inhaled, into the abdominal cavity), and to whom.
Information is an output too. “Eat 15 g of carbohydrate” or “glucose falling fast” makes a person the actuator. That’s why alarm design and follow apps for caregivers matter as much as dosing logic.
6. Imperfect sensor and delivery
Both ends of the loop distort the signal, so the algorithm never sees the true glucose and never knows exactly what insulin arrived.
- Sensor: measures tissue fluid, not blood. Lag, noise, compression lows, warm-up gaps, drift, early failure.
- Delivery: absorption varies with infusion-site age, location and scar tissue. Occlusions, heat-damaged insulin, air bubbles.
7. Feedback and time
Glucose is not just an input; it’s the result of earlier outputs, fed back. Delays everywhere make this a hard control problem:
| Delay | Typical scale |
|---|---|
| CGM behind blood glucose | About 5–15 min |
| Insulin under the skin to peak effect | About an hour or more |
| Fast carbohydrate to glucose rise | Minutes |
| Fat and protein effect | Several hours |
| Exercise effect on sensitivity | During and up to a day after |
The core mismatch is that food acts faster than insulin. That’s why meal announcements, pre-bolusing, faster insulin and pramlintide all matter.
8. The human in the loop
People are both sensors and actuators: they enter meals, change infusion sites, treat lows, override the system, and sometimes stop trusting it. The costs are real: mental load, alarm fatigue, broken sleep. For small children, all of this falls on parents, school staff and other caregivers, who aren’t with the child all the time. They rely on remote monitoring to be part of the loop at all.
9. Reliability
A slightly worse algorithm that always runs beats a better one that is down 5% of the time. Bluetooth dropouts, lost phones, operating-system updates, cloud outages and sensor gaps are all part of the loop’s real performance.
10. Replacing the loop
At the top level there’s a different strategy: replace the sensor, the algorithm and the pump with biology. Cell replacement with islets made from stem cells aims to restore the body’s own glucose control. The main open problem is protecting the cells from the immune system. See the road to a cure and the tracker.
Every engineering improvement is a bridge until then. A good bridge still matters: it’s what children live on while they wait.
Where the field is (October 2026)
| Part | State of the art | Source |
|---|---|---|
| Meal input: photo AI | Dedicated models can match dietitians; general AI chatbots disagree with each other | Thai food study, Turkish food study |
| Meal input: automatic detection | A research algorithm detected meals within about 26 minutes, with 83% sensitivity | NCT06422325 protocol |
| Exercise input | Smartwatch heart rate and movement drove insulin dosing in a randomised trial | Jacobs et al., Lancet Digital Health |
| Ketone input | Abbott’s combined glucose and ketone sensor: CE mark May 2026, FDA August 2026 | Abbott |
| Settings via AI | Five AI models gave textbook settings rather than ones learned from the data | Diabettech |
| Output: pramlintide | Can remove the need to announce meals, with more stomach side effects (29% vs 8%) | AJMC |
| Fully closed loop | CamAPS Liberty, EU approval January 2026 | IDTechEx overview |
| Open-source algorithms | oref ported to Swift; Dynamic ISF; device support is the main focus | Loop and Learn, Dec 2025 |