Confidential — For Dexcom Spain and Medical Affairs Europe — Internal review only
CGM Accuracy in Insulin Decisions
Calibration is the route back
when the sensor drifts
Every insulin decision made from a CGM reading trusts the sensor completely. That is true of an AID algorithm running every five minutes, and it is equally true of a person on MDI adjusting a mealtime dose. This is a clinician’s perspective on why that dependency matters, when it becomes dangerous, and why optional user calibration is a safety feature that factory calibration alone cannot provide.
The Fundamental Principle
AID does not make decisions.
Your sensor does.
Automated insulin delivery represents a genuine step change in T1D management. But beneath the sophistication – the predictive algorithms, the adaptive basal rates, the micro-bolus calculations running every five minutes – there is one irreducible dependency: the glucose reading the system receives.
The algorithm does not think. It does not feel. It cannot ask whether a reading makes sense in context, whether the person has been lying on the sensor, or whether the value has drifted slowly over three hours. It receives a number and acts on that number. Every time, without doubt.
“The algorithm will always make the right decision for the number it receives. If the number is wrong, so is the decision.”
Why the threshold matters, same actual glucose, different outcomes
20% error, AID can compensate
Sensor still reads above suspend threshold when glucose drops. PLGS catches it.
40% error, AID cannot compensate
Actual glucose is 3.2. Sensor reads 4.5, above the suspend threshold. AID never suspends.
The Compounding Error
A small inaccuracy is not a small problem
A mean absolute relative difference (MARD) of 8-10% sounds reassuring – particularly when we accept similar variation from blood glucose meters in clinical practice. But AID systems operate continuously near critical thresholds. An 8% error at 12.0 mmol/L is clinically irrelevant. The same error at 4.5 mmol/L can trigger a correction that causes a hypoglycaemic event. Move the sliders to see how error compounds in real clinical terms.
Sensor Error Impact Explorer
* Extra insulin is illustrative, assuming ISF of 2.2 mmol/L per unit. Individual sensitivity varies significantly.
What the accuracy data actually shows
Even the best factory-calibrated CGM sensor, under controlled trial conditions, does not achieve universal 20/20 agreement. The Dexcom G7 De Novo Summary submitted to the FDA shows approximately 5% of paired readings falling outside the 20/20 zone and 0.5% outside 40/40. In a person checking glucose every five minutes, that is a meaningful number of readings per day.
Dexcom G7, paired reading accuracy distribution
Source: Dexcom G7 Continuous Glucose Monitoring System, De Novo Summary (DEN220056), FDA, 2022. Data from pivotal trial across adult and paediatric populations under ambulatory conditions.
Why this matters, frequency versus recoverability
Every factory-calibrated sensor has a tail. At five-minute reading intervals, 0.5% outside 40/40 works out to roughly 1 to 2 readings per day in the trial data. Most self-correct within a reading or two, pressure artefacts, rapid-change lag, transient noise. What warrants a user-led calibration is the sustained drift episode where the sensor stays in the tail. In lived experience that is around three times a month. Rare. But when it happens, calibration is the only user-side mechanism that can intercept it before an algorithm or a person acts on it. This applies equally to AID and to anyone on MDI making a manual dose decision from a CGM reading. See the GNL CGM Accuracy guide →
When Accuracy Fails
Five scenarios where sensors drift –
and AID cannot tell
Factory calibration provides excellent baseline accuracy across the majority of sensor wear. But five predictable scenarios challenge that accuracy in ways the algorithm has no way to detect or compensate for. In each case, the system keeps running – it just runs on the wrong information. Tap each card to expand.
The Airbag Principle
You install it before you need it.
That is the point.
An airbag does not make you a better driver. It does not prevent accidents. It does not improve the car’s performance under normal conditions. You cannot feel it, hear it, or see it – until the moment you need it.
Calibration works on exactly the same logic. Most of the time, your sensor is accurate within clinically acceptable limits and calibration changes nothing meaningful. But calibration is not for most of the time. It is for the scenario you did not plan for – the compression artifact at 3 a.m., the sensor drift on day nine, the post-exercise reading that does not match how the person feels.
Same night. Same system. Different outcome.
“The algorithm cannot doubt its sensor. Calibration gives the person in the loop the ability to do what the algorithm cannot: ask whether the reading makes sense.”
Practical Guide
How to calibrate well
Calibration is most effective when it provides a clean blood glucose reference under stable physiological conditions. The GNL practice threshold is two consecutive readings more than 20% from a fingerstick, single discrepancies are often noise, but two in a row signals genuine drift. Following this rule, most people calibrate only a few times a month. Without the ability to calibrate at all, there is no route back when a reading lands outside 40/40, the 0.5% of readings where neither an algorithm nor a manual insulin decision can compensate, however well designed.
Summary
The bottom line
Automated insulin delivery, MDI in type 1 diabetes, and increasingly MDI in type 2 diabetes all share the same dependency: the glucose reading the person or the algorithm receives. The effectiveness of every one of those approaches is bounded not by algorithm sophistication or pump hardware, but by the accuracy of the information on which the decision is made. An algorithm will always make the best possible decision for the number it sees. A person on MDI will make the best decision they can for the number they see. Our job as clinicians and educators is to ensure that number is as reliable as possible.
Calibration is not about distrust of modern sensor technology. It is about understanding where any technology’s limits lie, and keeping a simple user-side pathway available for the moments when those limits are tested. Every factory-calibrated sensor has a tail. Only some offer the user a way back when it happens.
“Calibration does not improve average accuracy. It gives the person in the loop a route back from the tail.”