Glucose and Insulin: A Feedback Loop

Ninety milligrams per decilitre#

You have about five litres of blood, and dissolved in it, right now, roughly four grams of glucose. That is less than a teaspoon of sugar, spread through your entire circulation.

Now consider what happens across an ordinary day. You eat a bowl of rice containing sixty grams of carbohydrate — fifteen times the amount currently in your blood — and it is absorbed over ninety minutes. Later you skip dinner and go twenty hours without eating anything at all, while your brain continues to burn glucose at about 120 grams a day, indifferent to your schedule.

Through both of those, your blood glucose stays between roughly 70 and 140 mg/dL. It moves by perhaps 50%. The input to the system moved by a factor of fifteen in an hour and then went to zero for a day, and the output barely twitched.

That is not what an uncontrolled reservoir does. Pour fifteen times its contents into a bucket and the bucket overflows. The only way to get that kind of rejection of a large, fast disturbance is to build a control system — something that measures the variable, compares it to a target, and acts to close the gap. Physiology got there a long time before engineering did, and the glucose–insulin axis is one of the cleanest examples in the body.

A loop, not a level#

The vocabulary of control theory maps onto this system almost embarrassingly well.

The sensor is the beta cell of the pancreatic islet. It is not merely near glucose; it metabolises it. Glucose entering the beta cell is phosphorylated by glucokinase — an enzyme whose affinity is deliberately tuned so that its rate tracks glucose concentration across the physiological range rather than saturating — and the resulting ATP closes an ATP-sensitive potassium channel, depolarising the cell and letting calcium in. Calcium triggers exocytosis of insulin granules. The cell's electrical activity is, quite literally, a readout of ambient glucose.

The controller is that same beta cell's secretory decision: how much insulin to release for a given error signal. The effector is insulin acting on tissue — recruiting GLUT4 transporters to the membranes of muscle and fat cells so glucose can enter, and simultaneously telling the liver to stop producing glucose and start storing it as glycogen. The plant, in control language, is the glucose pool itself.

Crucially the loop is negative: high glucose produces a signal whose effect is to lower glucose. Positive feedback would be a runaway. Negative feedback is a thermostat.

And like a thermostat, it needs to push in both directions. The alpha cell, sitting next door in the same islet, does the opposite job: when glucose falls, it secretes glucagon, which drives hepatic glycogenolysis and gluconeogenesis and pushes glucose back up. Insulin and glucagon are a push–pull pair, and the pancreatic islet is a single small organ containing both halves of a bidirectional controller with the sensor built in.

This is why fasting is uneventful. Between meals your liver is quietly releasing glucose at roughly the rate your brain consumes it, and the balance is held by the ratio of the two hormones rather than by either alone. The counter-regulatory side has redundancy that the insulin side does not — cortisol, adrenaline, and growth hormone all raise glucose too — which is a reasonable design choice given that a low blood sugar kills you in minutes and a high one takes years.

Watching a meal go by#

The widget below runs the loop. The green band is the healthy range; pink is glucose, gold is insulin, and each violet marker is a meal arriving.

Press play with the defaults. Watch the order of events: glucose rises first, insulin follows it up with a short lag, and glucose turns over and comes down while insulin is still high. Insulin is not preventing the rise — it is chasing it. The peak lands a little above 130 mg/dL and the whole excursion is over inside three hours, which is what a healthy loop looks like.

Now drag insulin sensitivity down to about 0.3. Two things change, and only one is obvious. The glucose peak climbs and the curve takes noticeably longer to come home — that is the visible effect. The subtler one is the gold curve: insulin secretion roughly doubles. Nobody told the beta cells that sensitivity had dropped. They simply see a glucose error that is not resolving and keep secreting, and the extra output largely compensates. This is the state a great many people are in for years before anything shows up on a glucose test: normal glucose, purchased with abnormally high insulin.

Then press Silence beta cells. Now sensitivity is fine — the tissues would respond perfectly well — but there is no signal. Glucose climbs past 180 and is still elevated hours later, drifting down only through the slow, insulin-independent disposal that happens regardless. Nudge the meal slider up to 100 g and watch it go higher still. Notice that this failure looks quite different from the resistance case: not a sluggish loop but an open one.

Finally, use Eat now to stack a second meal onto a curve that has not finished resolving, and compare a healthy loop with a resistant one. The healthy loop absorbs the second hit almost as if the first had not happened. The resistant loop does not get back to baseline in between, and each meal starts from a higher floor.

The equations behind the loop#

Strip the physiology down to two coupled differential equations — essentially Bergman's minimal model — and the behaviour above falls out of the algebra.

Let $G$ be glucose concentration, $I$ insulin concentration, $G_b$ the fasting set point, and $R_a(t)$ the rate of glucose appearance from the gut. Then:

$$\frac{dG}{dt} = -p_1\,(G - G_b) \;-\; S_I\,I\,(G - G_b) \;+\; R_a(t)$$ $$\frac{dI}{dt} = \beta\,\max(0,\, G - G_t) \;-\; n\,I$$

Read the glucose equation as three competing terms. The first is glucose effectiveness $p_1$: disposal that happens without any insulin at all, largely by the brain and by mass action. The second is insulin-dependent disposal, proportional both to how much insulin is present and to $S_I$, the insulin sensitivity — how much disposal each unit of insulin buys. The third is the disturbance.

The insulin equation says the beta cell integrates the error above a threshold $G_t$ with responsiveness $\beta$, and insulin is cleared at rate $n$.

Loop gain, and what it buys#

Suppose a sustained disturbance $D$ — a slow glucose infusion, say. At steady state $dG/dt = 0$ and $dI/dt = 0$, so $I = (\beta/n)\,(G - G_t)$, and substituting gives a glucose offset of approximately

$$\Delta G \;\approx\; \frac{D}{p_1 + K}, \qquad K \;=\; \frac{S_I\,\beta}{n}\,\bar{G}$$

where $K$ is the loop gain — the product of how well the controller responds to error and how well the effector converts that response into action. Three consequences follow directly:

1. Higher gain, smaller error. Doubling $K$ roughly halves the residual offset. This is why a healthy loop can be hit with sixty grams of glucose and barely move.
2. The error never reaches zero. Pure proportional feedback always leaves a residual, because the correction is generated by the error. Some error is the price of the control signal. (Integral action would remove it; the beta cell's slower second-phase secretion behaves a little like integral action, which is one reason the loop does better than the simple algebra predicts.)
3. Gain is a product, so its factors trade off. $K$ depends on $S_I \times \beta$. Halve sensitivity, double secretion, and $K$ is unchanged.

That third point is not a mathematical curiosity — it is measured. Plot beta-cell responsiveness against insulin sensitivity across a population and the points fall along a hyperbola $S_I \times \beta \approx \text{constant}$, known as the disposition index. People sit at different places on that curve with the same glucose control. Diabetes is what happens when you fall off the curve, because the compensating factor cannot rise any further.

The time constant of the return is also readable from the equation: the excursion decays with roughly

$$\tau \;\approx\; \frac{1}{p_1 + K}$$

so higher gain does not merely reduce the peak, it shortens the whole event. With $K = 0$ — no insulin — only $p_1$ remains, and $\tau$ becomes hours instead of tens of minutes. That is exactly the difference between the two failure modes in the first widget.

Gain, delay, and why loops oscillate#

If more gain is better, why does the body not simply run enormous gain and hold glucose to three decimal places? Because the loop has delay in it. Glucose must be sensed, insulin must be secreted, it must circulate, bind receptors, and traffic transporters to a membrane. By the time the correction lands, the error it was computed from is stale.

The second widget strips the physiology away and leaves only that structure.

The top half is the loop as a block diagram — comparator, controller, effector, plant, and the sensor feeding the measurement back — with the signal circulating. The bottom half shows the regulated variable after a single disturbance.

Set the gain to 0 first. There is no controller at all: the variable rises with the disturbance and then leaks slowly back, never actually returning within the window. That is an open loop, and it is the shape of the beta-cell-silenced curve in the first widget.

Bring the gain up through 1, 3, 5. The peak drops sharply and the return gets faster, and around the middle of the range the response is crisp — up, over, back, done. This is the regime a healthy islet lives in.

Now push the gain past 8. Something new appears: the variable does not stop at the setpoint, it shoots straight through it and undershoots, then overcorrects the other way, and the trace rings. The controller is still acting on a measurement taken a moment ago, so at high gain it is always applying yesterday's correction to today's error. Enough of that and the loop chases its own tail.

This is the fundamental trade-off in every feedback system, biological or not: gain buys accuracy and speed, delay caps how much gain you can use. It is the same constraint that makes a shower with a long pipe impossible to set, and it is why the counter-regulatory arm of glucose control exists at all — a system that can only push one way must be run at cautious gain, whereas a push–pull system can afford to be aggressive because it can catch its own overshoot.

Two ways to break the same loop#

The elegance of the control-theory framing is that it makes the two major forms of diabetes into two different structural failures rather than two versions of "high blood sugar".

Type 1 diabetes is destruction of the controller. An autoimmune process kills the beta cells, so the sensor and the signal generator are gone together. $\beta \to 0$, and therefore $K \to 0$. The tissues are perfectly capable of responding to insulin — there simply is not any. The loop is open, and it must be closed from outside; the entire enterprise of insulin therapy, continuous glucose monitoring, and closed-loop "artificial pancreas" systems is an engineering effort to rebuild a controller that the body no longer has. Those devices face exactly the problems the second widget shows: subcutaneous insulin acts with a delay of an hour or more, which hard-caps the usable gain and is why they are so much harder to build than the block diagram suggests.

Type 2 diabetes breaks the effector instead. Tissues become resistant — receptor signalling downstream of insulin degrades, GLUT4 recruitment weakens, the liver stops listening to the instruction to shut off glucose production. $S_I$ falls. For a long time nothing is visible, because $\beta$ rises to compensate and the disposition index holds $K$ roughly constant. That compensated phase can last a decade, and it is characterised by high insulin with normal glucose. What eventually produces the diagnosis is not further resistance but beta-cell failure: the compensating factor stops rising, $K$ finally falls, and glucose escapes. Two different variables, in the same product, failing in sequence.

The distinction has real consequences beyond taxonomy. In type 1, the missing quantity is the signal. In type 2, the signal is often abundant and the problem is that it is not being heard, which is why the two conditions behave so differently even when a glucose measurement makes them look identical. It also explains why the earliest detectable abnormality in type 2 is usually not fasting glucose but the shape of the postprandial excursion — a loop with reduced gain and slowed dynamics shows its weakness under load before it shows it at rest, which is precisely what an oral glucose tolerance test is doing when it hands someone 75 grams of glucose and watches the curve.

This article describes the physiology and its control-theoretic structure. It is not medical guidance, and nothing here should be used to make decisions about diagnosing, monitoring, or treating any condition.