Signal Processing Algorithms in CGM

A CGM sensor never measures glucose directly. It produces a tiny electrical current (or, in the implanted Eversense, a light signal), and software turns that signal into the number on your screen. That software calibrates the raw signal into glucose units and filters out noise. Research systems go further and try to make up for the minutes that sensor readings trail blood glucose. We read FDA decision summaries, a Dexcom patent and engineering papers to see how each step works.

Turning a current into a glucose number

In the Dexcom G6, the enzyme glucose oxidase reacts with glucose in interstitial fluid (the fluid just under the skin), producing a current in proportion to the glucose level. The FDA's summary says only that the transmitter "converts the signal using an algorithm" into mg/dL [1]. Abbott's FreeStyle Libre 3 Plus and Libre 2 Plus sensors work on the same enzyme principle [2].

The conversion rule is called a calibration function. A 2018 University of Padova review explains that most manufacturers have used the simplest one: a straight line set by sensitivity (current per unit of glucose) and baseline (current that flows anyway) [3]. Early systems refitted that line from fingersticks one or two hours after insertion and then every 12 to 24 hours [3].

Here's what surprised us: the line itself moves. Sensitivity drifts after insertion, most of all on the first day, as the sensor's membrane meets the body's immune response, so a straight line only fits for a limited stretch [3]. We go deeper in Sensor Calibration Algorithms and Data Fusion.

Factory calibration and sensor codes

Factory calibration moves that work to the production line. In a 2017 paper, two Abbott engineers laid out the steps: build sensor lots with very little sensor-to-sensor variation, test a sample from each lot in lab glucose solutions, and turn the lot's sensitivity into a code that ships with every sensor [4]. The code only holds if sensitivity stays stable on the shelf and on the body [4]. They were writing about the original FreeStyle Libre [4]; later Libre generations include the Libre 3 Plus and 2 Plus [2].

The G6 used codes. Once a user entered a sensor's code, no fingersticks were needed for its 10-day wear. Skipping the code meant entering two fingerstick values at startup and every 24 hours after [1]. The FDA wrote that factory calibration keeps the G6 accurate "without influence from potentially erroneous SMBG calibrations," meaning faulty fingerstick entries [1]. Abbott's Plus sensors go further: they are factory calibrated and "cannot be calibrated by the user" [2].

Not every CGM has dropped fingersticks. The implanted Eversense 365 reads a fluorescent molecule that glows brighter as it binds glucose, and its FDA summary lists four calibrations after warm-up, one a day through day 13, then one a week for up to a year of wear [5].

Calibration logic is patent territory too. Dexcom's US Patent 7,276,029, filed in 2003, describes calibrating from matched sensor and fingerstick pairs, updating as new reference values arrive, and rejecting reference data judged clinically unacceptable [6]. A patent shows what a company claimed, not what any sensor runs today.

Smoothing out noise, and what it costs

Raw CGM data is jittery, and the noisier it is, the more likely a reading is to cross an alert threshold when glucose itself didn't [7]. Filtering helps. But a smoothing filter averages recent readings, so its output always trails reality a little.

A 2009 review by Medtronic engineers offers a rare look inside commercial filters, for two Medtronic monitors of that era [8]. The CGMS Gold chained a moving average (an average of the last several samples), a "rate-limiting" filter that clipped changes too fast to be physiological, and a final smoothing step [8]. The Guardian RT's filters added a net delay of 8.25 minutes [8]. Neither monitor appears on MiniMed's sensor comparison page as of October 2026; it lists the Instinct, Simplera Sync, Guardian 4 and Simplera CGM [9].

A Kalman filter, which estimates the true signal from noisy data using a model of how glucose changes, can do better. In a 2010 study, one that tuned itself to each person's noise level added about 35% less delay than a fixed moving average at similar smoothness. At the time, the authors wrote, fixed moving averages were "in use in likely all commercial CGM devices" [10].

Compensating for the body's own lag

Even a perfect filter can't remove physiology. Glucose has to move from blood into interstitial fluid, and a tracer study in eight healthy fasting adults measured that delay at about 5 to 6 minutes [11]. The Medtronic review estimated roughly 3 to 12 minutes of physiological lag, plus 1 to 2 minutes for the sensor chemistry [8].

Research groups have tried several fixes. One is to calibrate only when glucose is steady, so blood and tissue values agree [3]. Another is deconvolution, which runs a blood-to-tissue model in reverse to estimate the blood glucose behind a reading [7]. A third looks ahead: in one Padova example, a 30-minute forecast flagged a coming low 15 minutes before the CGM trace crossed the threshold [7]. Predictive alerts, the Medtronic engineers noted, can offset both processing and physiological delays [8].

Prediction has limits. A 2019 review of 55 machine-learning studies found no single model that predicts glucose accurately in every circumstance, and said the effect of CGM time lag was "not well covered" [12]. There's more in AI and ML in Glucose Signal Processing.

What the FDA checks: the output, not the method

None of the FDA summaries we read names a filter or a model; they say the signal is converted "using an algorithm" [1] [5]. The FDA regulates the result. When it authorized the G6 in 2018, it created the integrated CGM (iCGM) category for sensors that feed data to insulin pumps and other connected devices, with special controls that later iCGMs must meet [1]. Abbott's Plus sensors fall under the same regulation, 21 CFR 862.1355 [2]. Several controls test the output directly [1]:

• Across the measuring range, the share of readings within 20% of a lab reference must have a lower 95% confidence bound above 87%.
• No more than 1% of readings may show glucose rising faster than 1 mg/dL per minute while blood glucose is falling faster than 2 mg/dL per minute, and the same limit applies in reverse.
• The sensor must not leave clinically significant data gaps, and readings must reach connected devices at clinically meaningful intervals.
• Labeling must report accuracy by glucose range, rate-of-change accuracy, and how often data gaps occur.

We read the rate-of-change rule as the one closest to filtering and lag, since it checks a trend's direction, not just its distance from the lab value. Our iCGM Regulatory Standards article covers the full list.

Where the algorithms still fall short

Software also decides when a sensor has failed. In the G6's clinical study, 36 of 374 sensors (9.6%) had an "early sensor shut-off," where the algorithm detected a sensor not working as intended and turned it off [1]. The G6 is now a legacy product: Dexcom will no longer manufacture it after July 1, 2026, and recommends that G6 users move to the Dexcom G7 or G7 15 Day [13].

What we couldn't find is much detail on today's commercial processing. The most detailed filter description we found dates from 2009 [8], and current FDA summaries describe what the output must achieve, not how the software gets there [1] [2] [5]. If a reading doesn't match a fingerstick or how you feel, what to do next is a question for your care team.