Part 2: Different Models For Calculating IOB

Part 1 explained how rapid-acting insulin behaves biologically.

Part 2 turns to how diabetes devices attempt to model that insulin mathematically.

These are not the same thing.

Your body runs on pharmacodynamics. Your diabetes device uses mathematics and decay algorithms.

The key point is simple: physiological insulin action is variable, but device IOB models are simplified.

What devices are actually calculating

Devices define IOB using:

• Active Insulin Time (AIT/DIA)
• A mathematical decay curve
• Total insulin delivered within the time window

Once the selected time window ends, the displayed IOB becomes 0.0, regardless of whether physiological insulin activity is still present.

This graphic illustrates how physiological insulin exposure from a 0.2 U/kg (usual meal dose e.g. 50kg person 10 untis) is massively under when AIT/DIA is set to 2, 4, or 6 hours.

This graphic illustrates how glucose from carbohydrate appears in the bloodstream over time. Carbohydrate absorption is highly variable and depends on the type of food eaten, the composition of the meal, gastric emptying rate, and individual physiology.

Simple carbohydrates and high glycaemic foods can lead to rapid glucose appearance, while meals containing fibre, fat, or protein often slow digestion and produce a more gradual rise in glucose. Some meals, particularly those high in fat, can delay glucose appearance for several hours.

Beyond the meal itself, many physiological factors influence how quickly glucose enters the bloodstream. These include gastric emptying speed, stomach acid levels, gut hormones, physical activity, gut microbiota, and individual digestive physiology. Conditions such as gastroparesis can further alter or delay glucose appearance.

For this reason, carbohydrate absorption is not a fixed curve. The patterns shown in this graphic represent typical examples ranging from rapid to slow glucose appearance, illustrating why post-meal glucose responses can vary significantly even when carbohydrate amounts in these meals are similar.

In other words, we are trying to match variable insulin with variable glucose appearance using a simplified model.

Is it any wonder managing Type 1 diabetes can sometimes feel like a proper graft?

The meal insulin vs correction insulin problem

Here’s the structural issue:

Most systems do not distinguish between:

• Insulin given for carbohydrate
• Insulin given for correction

They are lumped together as “IOB” whic is problamtic because meal insulin is often still covering carbohydrate that has not yet finished digesting. It’s only the correction insulin targeting excess glucose in circulation.

Physiologically, these behave differently.

Mathematically, most devices treat them the same, leading to:

High IOB blocks a correction (so you override)

Example: you take 14 units total.

🟢 10 units for food
🔴 4 units for correction

Your AIT/DIA is set to 4 hours and the bolus calculator treats everything as one pool of IOB — a single blue line.

Two hours later, when glucose is still entering circulation as your food has not been fully digested ,your glucose is 13.9 mmol/L (250 mg/dL) and rising.

You expect a correction. The device says no.

It sees ~4 units of total IOB and blocks insulin — even though only ~1 unit may be true correction insulin, and the rest is still covering food that ahs not been absorbed yet. When carbohydrate has been underestimated (which happens 30–50% of the time), the algorithm assumes it is preventing stacking by treating bolus insulin as one pool of IOB.

The result isn’t “wrong” — it’s conservative modelling that can feel restrictive when glucose is high and rising after carbohydrate underestimation.

Here’s the key detail most people miss:

The calculator does not know why insulin was given. It only knows how much insulin was delivered and how much remains within the model.

Most of the insulin still circulating is actually meal insulin working against carbohydrate absorption — not “free correction insulin”, which is the one that actually matters in this particular situation.

But because everything is merged into one total, the system behaves conservatively.

So what happens in real life?

You override.
You ghost-carb.
You inject by pen.

Glucose comes down. You avoid hours of hyperglycaemia.

Later, someone says, “Just trust the technology.”

You think, “Why don’t you Foxtrot Oscar?”

But this isn’t about trust — it’s about modelling.

• The algorithm sees total risk.
• Humans think dynamically: food vs correction vs activity

Naturally, the solution is to shorten the AIT/DIA so that larger correction doses are permitted when carbohydrate intake has been underestimated.

Why shortening AIT “works” for tighter glucose management

Consider someone who consistently runs high at 2–3 hours post-meal, due to carbohydrate underestimation (i.e., he or she eats more carbs than assumed).

If AIT is set at 5 hours, the system may show substantial IOB and suppress a correction, because it assumes that insulin is still active.

But if much of that insulin is still dealing with unabsorbed carbohydrate, a correction may in fact be physiologically appropriate.

So what happens?

They shorten AIT to 2 hours.

This reduces displayed IOB sooner — allowing more aggressive corrections.

Using the same example, a 2-hour AIT allows the full 4 units of correction.

On paper, that looks logical — but physiologically, there is still correct insulin working in the background (about 2 units). The result? You may overshoot by roughly 2 units

Now imagine a different scenario.

There is no food involved, just a very high glucose — say 20 mmol/L (≈360 mg/dL) — and you’re correcting every two hours.

Because the algorithm treats bolus insulin as effectively gone at 2 hours. it keeps offering full corrections each time. But in reality, some insulin from the previous dose is still active.

This is where insulin stacking risk rises quickly.

Not because the person is doing anything wrong — but because a short AIT compresses the model’s memory of insulin action. Each new correction is calculated as if the previous one has finished, even when physiology hasn’t caught up yet.

What makes this clinically important is not that the model is “wrong”.

It’s that physiological insulin exposure is often underestimated by the display.

When the screen shows 0.0 IOB, many people reasonably assume they are bolus insulin-free. That assumption can create a false sense of security — particularly before exercise (which supercharges insulin), or when giving repeated corrections every two hours.

In reality, circulating insulin may still be meaningfully active.

The result is predictable:

• Unexpected exercise hypoglycaemia
• Late crashes after stacked corrections
• Loss of trust in the bolus calculator

This is not about good or bad settings.

It’s about what you’re optimising for.

Not all IOB models are identical

Up to this point, we’ve described “device IOB” as if it were one thing.

In reality, there are several modelling approaches — though most share the same structural limitation.

Before going further, we need to add nuance.

Correction-only IOB models

Unfortunately, most systems lump meal insulin and correction insulin together. And we get it, as it might sounds easier for the user to interpret, right? But it is not entirely correct, as explained above.

However, a minority of bolus calculators allow correction insulin only to count as IOB for future correction calculations within the bolus calculator.

This partially solves the “correction suppression” problem.

• Meal insulin is assumed to cover carbohydrate.
• Correction insulin is tracked to prevent stacking.

With a longer DIA (4–6 hours), this can allow more physiologically appropriate corrections while still limiting repeated correction stacking.

Thinking back to the earlier example, if a correction-only IOB model (pink) were in place, the calculator would recognise that most of the remaining meal insulin is covering undigested carbohydrate. Instead of treating all insulin as one lumped pool and blocking the correction completely, it would focus only on the correction component, thereby allowing a new bolus correction and avoid unwanted hyperglycaemia.

In this scenario, the system could recommend a more appropriate top-up of around 3 units.

If active insulin time were shortened to 2 hours, the full 4-unit correction would likely be delivered.

If both the meal bolus and correction insulin were fully counted as IOB, the system would offer 0 units.

This sounds like the ideal model at first glance. However, there are two important trade-offs.

First, it assumes glucose should not rise after eating because meal insulin is not factored into the IOB calculation. As a result:

• Any post-meal rise is interpreted as glucose that requires correction.
• The system does not recognise that most mixed meals produce a physiological rise of approximately 2–4 mmol/L (40–80 mg/dL) in the first hour.
• In many cases, glucose then returns towards the target over the following two to three hours if the bolus was appropriate.

If corrections are given in the first couple of hours after eating:

• They are often unnecessary.
• They are responding to a normal meal rise rather than true insulin insufficiency.
• This increases the risk of late hypoglycaemia.

This approach may work reasonably well for someone eating a very low-carbohydrate diet, where meal rises are minimal. It is far less appropriate for someone consuming 40–50% of total energy from carbohydrate, where post-prandial excursions are expected in physiology.

Second, if only correction insulin is displayed as IOB:

• Circulating insulin is significantly under-represented.
• Exercise hypoglycaemia risk may appear lower than it truly is.
• Users may falsely assume they have “little insulin on board” when substantial meal insulin is still active.

Some bolus calculators that exclude meal insulin from correction logic still display both meal and correction insulin on the screen. This transparency helps mitigate the risk by allowing the user to see the total active insulin. It is essential to check how a system presents this information. This distinction becomes critical when comparing devices and will be explored in Part 3.

In summary:

• The model improves correction logic.
• It may reduce insulin stacking from aggressive correction calculations.
• But it risks over-correction early after meals.
• It can underestimate true circulating insulin.
• It may increase hypoglycaemia risk with high-glycaemic meals or when pre-bolusing is inconsistent.

It solves one problem, but it creates others.

Correction suggestions without IOB deduction models

Some bolus calculators allow users to choose not to deduct IOB from correction recommendations. In these systems:

• The full calculated correction dose is suggested.
• IOB is still displayed on the main screen.
• IOB is calculated using the chosen DIA/AIT, typically 4–6 hours.

This approach can appear to provide a useful balance:

• Full correction responsiveness when glucose is high.
• Longer DIA/AIT settings that better reflect physiological insulin exposure and help identify exercise-related hypoglycaemia risk.

However, the same structural limitations remain.

The model still does not account for:

• Expected glucose behaviour after meals, where rising glucose may still be driven by ongoing carbohydrate absorption.
• Multiple corrections given within a short time period, where physiological insulin exposure may accumulate even if the calculator allows repeated full corrections.

Because the device is no longer limiting corrections based on IOB, the responsibility shifts more heavily to the user.

This type of system therefore, works best when the user is highly aware of two key risks that can increase hypoglycaemia:

• Post-meal glucose dynamics that may temporarily justify higher glucose levels.
• Stacked corrections given close together while insulin from previous doses is still active.

Without that awareness, the risk of delayed hypoglycaemia can increase despite apparently reasonable calculator recommendations.

Comparing Three Methods of Deducting IOB From Correction Suggestions

This graphic summarises how different IOB calculation models change the amount of correction insulin offered, even when the underlying physiological situation is identical.

In this example:

• 10 units were given for food
• 4 units were given for correction
• Active Insulin Time (AIT) is set to 4 hours with a linear decay model. . Hence, at 2 hours, approximately 50% of insulin remains active

The chart shows:

• Blue bars: Insulin on Board (IOB) seen by the calculator
• Orange bars: Correction insulin allowed by the system

Depending on whether the calculator counts:

• no IOB = 4 units
• correction insulin only = 2 units
• meal and correction insulin together = 0 units

The underlying physiology has not changed — only the mathematical model used by the device.

The trade-off nobody explains

You can optimise AIT for:

• Correction performance: Shorter AIT (2-3 hrs) improves correction responsiveness – More aggressive correction insulin but lower exercise hypo prediction risk.
• Physiological accuracy: Longer AIT (4-6hrs) better reflects physiological exposure – Less aggressive correction of insulin but better exercise hypo prediction risk.

But you cannot perfectly optimise it for both at the same time if meal and correction insulin are lumped together.

This is the structural limitation of current IOB models.

It is not user error. It is a model architecture.

Linear vs curvilinear decay models

Most commercial systems use one of two mathematical decay shapes:

1. Linear degradation — insulin declines evenly over time.
2. Curvilinear degradation — insulin falls more steeply early on, with a longer tail.

Insulin action is front-loaded in the curvilinear model.

• The greatest glucose-lowering effect occurs in the first two hours after dosing.
• After this early peak phase, the rate of effect progressively slows.
• This pattern is represented by the concave curve in the figure, showing a rapid initial decline followed by a more gradual tail.

However — and this is important — the figure above shows theat chosen duration (AIT/DIA) matters more than the exact curve shape.

In practical terms, getting the duration broadly right is the major issue. Whether the curve is linear or slightly curved iis minor in comparison.

Advanced models: offset, peak, and meal-rise settings

Some systems move beyond a simple “one decay curve” model by introducing the idea of the expected glucose rise after eating.

Three settings allow personalisation of expected glucose behaviour after eating.

• A meal-rise expectation — an estimate of the peak glucose typical after meals for that individual (1.7-5.6 mmol/L or 30-100 mg/dL)
• An offset time — At what point does the glucose start to decrease from the peak meal-rise (45-120 minutes).
• An acting time — an estimate of how long the insulin lasts (3-8 hours), and creates the drop off from the end of the off-set time to the acting time.

Why this solves the correction problem better

In a simple lumped model, all remaining insulin counts equally against a correction. That means meal insulin still covering undigested carbohydrate can block a needed correction.

With an offset + meal-rise + acting time model:

• The system allows for personalised glucose behaviour after meals
• It allows corrections to the expected glucose behaviour
• It only takes away the correction insulin as IOB to prevent insulin stacking

This graphic shows three different expected post-meal glucose exposures in one person:

• Blue: low expected glucose exposure
• Orange: moderate expected glucose exposure
• Green: high expected glucose exposure

The expected glucose exposure or behaviour will depend on:

• Pre-bolus timing
• Carbohydrate amount
• Carbohydrate type and absorption speed

These patterns can usually be identified clearly using CGM trends, allowing for personalised adjustments based on observed glucose responses rather than assumptions.

This reduces the need to artificially shorten AIT and generally personalises post-meal correction behaviour far better than basic linear models.

That is a genuine strength.

Importantly, the displayed IOB in this model still reflects total insulin delivered — both meal and correction insulin combined, using the acting time selected and linear or curvlinear degradation model.

That means if the Duration of Insulin Action is set realistically (for example, at least 4 hours), the display does a reasonable job of representing total circulating insulin exposure and exercise hypoglycaemia risk.

The limitations are threefold;

• First, the model assumes glucose behaves the same after each similar meal. In reality, two meals can behave very differently depending on:
  • Carbohydrate amount
  • Fat content
  • Protein content
  • Total meal size
  • Glycaemic load and absorption speed
  • Pre-bolus timing
  • Injection site and recent activity

• Second, the size of the bolus in units per kilogram fundamentally alters physiological insulin exposure. A 0.05 U/kg dose does not behave like a 0.20 U/kg dose — in duration, tail, or hypoglycaemia risk. The model does not quantify how the insulin dose per kilogram modifies:
  • Late hypo risk
  • Exercise-associated hypoglycaemia
  • Stacking vulnerability
  • Duration of biological insulin action

• Third, lack of total IOB visibility
  • Most devices using this approach do not display the total IOB from both meal and correction insulin on the main screen.
  • The full circulating insulin exposure is therefore not visible.
  • This makes it difficult for users to understand true insulin exposure, which is the key driver of hypoglycaemia risk during activity or exercise.

Despite this, compared with linear AIT or lumped decay models, it is still the most coherent and clinically usable bolus framework available at present.

AID/HCL System

It is important to recognise that automated insulin delivery (AID) and hybrid closed loop (HCL) systems do not all handle IOB in the same way.

For user-initiated boluses via their bolus calculators:

• Many systems use a relatively simple lumped decay model.
• All bolus insulin is treated as a single pool.
• That pool decays over the selected active insulin time (AIT/DIA).

However, the internal algorithm logic is often more complex.

For automated adjustments:

• Most systems distinguish between user-entered correction boluses and algorithm-driven micro-corrections.
• Different logic may be applied to each.
• Some apply different decay assumptions for user insulin versus automated insulin.

This creates several practical challenges:

• The IOB displayed to the user may not match what the algorithm is using internally.
• How the algorithm IOB is calculated and adjusted is often proprietary and not well defined. Several variables are at play.:
  • The expected basal rate may be based on the programmed basal profile or an adapted estimate derived from the total daily insulin dose.
  • When insulin delivery falls below this expected level, this may or may not be deducted from the IOB pool and may display negative IOB.
    • Although it can look unusual, negative IOB effectively indicates that circulating insulin exposure is lower than anticipated, which can help refine the understanding of hypoglycaemia risk.
    • Whether the value comes from the user-facing model, an internal algorithm calculation, or a combination?
    • Whether it is derived using the correction factor in the settings or the algorithm-generated correction factor from the recent total daily insulin
    • What duration of insulin action is being assumed internal IOB vs the one set by the user
    • What decay model (linear or curvilinear) is being applied, and is it the same for internal vs user IOB pools
  • It may be unclear whether IOB includes meal boluses, correction boluses, automated corrections, or basal modulation.

The key implication:

• What is displayed on screen may not fully represent what the algorithm is accounting for in the background.

Despite these structural differences, the central tension remains:

• Insulin modelling must balance protection against over-correction.
• At the same time, it must reflect the front-loaded physiological action of rapid-acting insulin.

That trade-off between safety and physiological realism underpins most commercial AID architectures.

We will examine how commercial AID and HCL systems model IOB in Part 3, breaking down the underlying logic, assumptions, and practical implications for users.

For a comprehensive review of this, read this excellent netIOB article from Prof Mike Riddell.

DIY systems

DIY platforms for both AID/HCL and bolus calculators allow even greater flexibility. Several additonal variables may be at play:

• Custom insulin action curves
• Carbohydrate absorption modelling
• Dynamic adjustments based on sensitivity shifts

This guide focuses on commercially available systems and does not attempt to review the DIY landscape.

Detailed Summary of IOB Models and Example Settings

This table brings together the main IOB modelling approaches used in pumps, bolus calculators, and AID systems, and shows how different configuration choices influence day-to-day behaviour.

It is not essential to read this table in full immediately, as a simplified summary appears at the start of Part 3.

However, for readers who want to explore the detail, the full comparison is presented here.

At a high level, the table shows three key things.

• The type of IOB model used
  • Linear or curvilinear insulin decay
  • Meal and correction insulin are pooled or separated
  • Expected glucose behaviour models
  • Automated insulin delivery systems with a separate algorithm for IOB
• How system settings change behaviour
  • Each model is shown with aggressive, medium, and protective configurations.
  • These are typically driven by settings such as Active Insulin Time (AIT) or Duration of Insulin Action (DIA) and, in some systems, meal-response parameters such as offset time or meal rise.
• The practical consequences of those settings
  • How strongly are corrections limited after insulin has been given
  • The risk of insulin stacking and correction-induced hypoglycaemia
  • The likelihood of persistent hyperglycaemia if carbohydrate intake is underestimated
  • How well the system reflects true circulating insulin exposure and exercise-related hypo risk

Coming next

With these modelling approaches in mind, we can now examine how real devices implement them.

We’ll walk through how common pumps, AID/HCL systems, and bolus calculator apps implement IOB, what is configurable, what is fixed, and how to think about the trade-offs each design creates when choosing your settings.

The IOB Guide for T1D

• Hub: The Insulin On Board Guide for T1D
• Part 1 – The Insulin On Board–Physiology Mismatch
• Part 2 – Different Models For Calculating Insulin On Board
• Part 3 – Choosing a Device-Specific Insulin On Board Settings: What Are You Optimising For?
• Part 4 – The Future of calculating Insulin On Board: combining correction behaviour and exercise hypoglycaemia risk
• Part 5 – GNL Exercise Insulin on Board Calculator for T1D
• Part 6 – Reccomended Reading and Resources