Impact of AID (Automated Insulin Delivery) Algorithms on CGM Requirements

Impact of AID Algorithms on CGM Requirements

An automated insulin delivery (AID) system acts on every glucose reading it receives, so the sensor feeding it has to clear a higher bar than one that only displays numbers. In the US, that bar is the FDA's integrated continuous glucose monitoring (iCGM) category. Its rules set accuracy targets for low, in-range and high glucose, forbid the most dangerous wrong readings, limit trend errors, and require proof that readings keep arriving without clinically significant gaps [1]. As of October 2026, the rule hasn't changed since it was codified on February 18, 2022 [1].

Why a dosing algorithm raises the stakes

The FDA created the iCGM category in March 2018 by granting Dexcom's De Novo request for the Dexcom G6. Its decision summary names the risks: treatment based on falsely high or low values, no readings because of signal drop-out or lost communication, and an iCGM whose performance can't support the AID system it feeds [2]. The G6 is now a legacy product; Dexcom stopped manufacturing it after July 1, 2026, to focus on newer sensors such as the Dexcom G7 15 Day [3].

The dosing software has a matching rule. An interoperable automated glycemic controller (iAGC), the FDA's category for AID algorithms, may take glucose only from cleared iCGMs unless the FDA accepts another sensor type [4]. Our AID interoperability article covers which sensors pair with which pumps.

Accuracy, one glucose range at a time

The iCGM rule doesn't use a single accuracy score. It splits adult study readings into three ranges with a target for each [1]:

• Below 70 mg/dL, readings are judged in absolute terms: more than 85% must fall within 15 mg/dL of the lab value.
• From 70 to 180 mg/dL, more than 70% must fall within 15% of the lab value.
• Above 180 mg/dL, more than 80% must fall within 15%.

Within a wider 40 mg/dL or 40% band, the targets rise to above 98% for low readings and above 99% for the other two, and across the whole range more than 87% of readings must land within 20% [1].

Here's what surprised us: none of those percentages is a plain average. Each must be met by the lower one-sided 95 percent confidence bound [1], the low end of the range where the true rate probably sits, so a borderline result in a small study can fail even when its raw percentage clears the line. Performance also has to hold across the whole wear period, and children need similar accuracy or a justification the FDA accepts [1].

The outlier rule allows no exceptions

Two lines in the rule permit zero misses. When the iCGM reads below 70 mg/dL, no matching lab value in the study may be above 180 mg/dL. When it reads above 180, no lab value may be below 70 [1]. A reading that far off would push an algorithm the wrong way, the first risk the G6 decision summary names [2].

Trend accuracy matters because algorithms predict

AID software looks ahead. The FDA's 2019 memo on Tandem's original Control-IQ describes it predicting glucose 30 minutes out from prior iCGM readings and other inputs, then adjusting or suspending basal insulin every 5 minutes [5]. Its successor, Control-IQ+, now runs the t:slim X2 and makes the same 30-minute prediction [6].

So the iCGM rule polices direction too. No more than 1% of readings may show glucose rising faster than 1 mg/dL per minute while lab values show it falling faster than 2 mg/dL per minute, and the same limit applies in reverse [1]. Labeling must report how accurate the rate-of-change data were [1].

Gaps, transmission and what algorithms do without data

An iCGM must show that, over its claimed life, it doesn't allow clinically significant gaps that would stop connected devices from working. It also needs secure, reliable transmission at "clinically meaningful time intervals," a lockout once the wear period ends, and labeling that reports how often gaps happened and how long they lasted [1]. The rule doesn't put a number on "clinically significant."

The G6, which sent a reading every 5 minutes, shows what that reporting looks like. Across 374 sensors, 98.6% of expected readings arrived, and the daily capture rate slipped from a peak of 98.7% on day 3 to 96.4% on day 10 [2].

The iAGC rule covers the other end. Dosing software must keep therapy safe when communication is interrupted, lost or re-established, and keep working safely when incoming data fall outside specified bounds [4]. The FDA summaries we read describe how some systems do this:

• In its 2022 clearance summary, Omnipod 5 switches to a state called Automated: Limited during sensor warm-up and after 20 minutes or more without sensor data, giving no more than the preprogrammed basal insulin until a valid reading returns [7].
• The iLet's 2023 summary describes BG-run mode, which keeps delivering basal insulin based on past needs and treats fingerstick entries as sensor data, for the shortest time possible [8].
• MiniMed's SmartGuard needs valid sensor and blood glucose values to start automation, and the FDA's 2025 comparison table says both SmartGuard and Control-IQ "can transition to a set limited basal delivery rate should conditions arise" [9].

That SmartGuard summary also accepts "compatible interoperable Medtronic continuous glucose monitors" alongside iCGMs [9], the kind of exception the iAGC rule allows [4]. None of these summaries gives every trigger and timer, so each system's user guide is the place to check, and how to cover a gap in sensor data is a decision to make with your care team.

Lag, filtering and the cost of a smooth line

AID sensors measure glucose in interstitial fluid, the fluid around cells under the skin. A Mayo Clinic tracer study in eight healthy, fasting adults put the delay from blood to that fluid at about 5 to 6 minutes, and its authors judged it no obstacle for overnight or fasting closed-loop use [10]. Meals and exercise were outside the study.

Raw sensor signals are noisy, so they are filtered before an algorithm sees them. A 2016 review describes real-time smoothing built on a Kalman filter (a method that estimates the true signal step by step from noisy measurements) and counts the blood-to-interstitial delay among the key parts of sensor error [11]. Bequette's 2010 review treats filtering, lag compensation and dropout detection as linked problems [12]; our signal processing article goes further into the math.

Compression lows

Pressure on the sensor site, such as lying on it during sleep, can produce falsely low readings that researchers call pressure-induced sensor attenuations (PISAs) [13]. In more than 1,125 nights from an at-home trial of a predictive low-glucose suspend system, about 3% of CGM readings were marked as PISA events, and the authors note these artifacts can trigger unwanted overnight pump shut-offs [13]. Their detection algorithm caught 88.34% of events, and tuning brought false detections as low as 1.70% [13]. That was a research algorithm; the FDA summaries we read don't say which commercial systems screen for compression lows.

Factory calibration

Fingerstick calibration can carry its own errors into an algorithm's input. The FDA wrote that the G6's factory calibration "ensures continued accuracy without influence from potentially erroneous SMBG calibrations," SMBG meaning self-monitoring of blood glucose with a meter [2].

What the rule leaves open

The iCGM rule says what a sensor maker must prove, but it leaves "clinically significant gap" undefined and lets each algorithm decide what to do when data stop. The figures that separate sensors for AID use, how often data went missing and how well trend direction held up, sit in each sensor's labeling because the rule requires them there [1].