AI & Machine Learning in CGM Calibration

Introduction to AI & Machine Learning in CGM Calibration

Calibration is how a continuous glucose monitor (CGM) turns the tiny electrical signal from its sensor into a glucose number. Sensors such as the FreeStyle Libre and the Dexcom G6 were calibrated at the factory instead of with your fingersticks [3][11], and it's tempting to credit artificial intelligence (AI) for the change. Here's what surprised us: Abbott's engineers trace factory calibration mainly to sensor stability and a tightly controlled manufacturing process. Each sensor lot is built for low sensor-to-sensor variation, sample sensors are tested in the lab, and the lot's measured sensitivity is programmed into the sensor as a code [3]. Machine learning (ML), software that learns patterns from data instead of following hand-written rules, is widely studied for forecasting glucose minutes ahead [1]. To check whether a specific device counts as AI-enabled, the FDA keeps a public list of authorized AI-enabled devices; the copy we read was last updated in September 2026 [2].

Key Technical Drivers

Three kinds of math come up again and again in CGM signal processing. Only one of them is machine learning in the usual sense.

• Kalman Filters: A Kalman filter estimates a hidden value by alternating two steps: predict the next state, then correct that prediction with a new, noisy measurement. R. E. Kalman published the method in 1960 [4]. A University of Padova team applied it to CGM data in 2010 and compared it with moving-average filtering, which they described as in use in "likely all commercial CGM devices" at the time [5]. So the Kalman filter is better described as a well-studied research method than as a confirmed industry standard. Our article on signal processing algorithms in CGM covers filtering in more depth.
• Predictive Algorithms: Recurrent neural networks (RNNs), including long short-term memory (LSTM) networks, are ML models built for data that arrives in sequence, like a CGM reading every few minutes. One 2021 study cleaned CGM data with Kalman smoothing, then used a stacked LSTM to forecast glucose 30 and 60 minutes ahead [6]. A 2026 systematic review of 26 studies called deep learning the state of the art for glucose forecasting, while noting that interpretability and physiological fidelity still stand in the way of clinical use [1]. Forecasting doesn't remove the lag between blood and interstitial fluid, the fluid around cells under the skin where most CGMs measure. Measured directly in healthy adults after an overnight fast, that physiological delay was about 5 to 6 minutes [7].
• Impedance Spectroscopy: This technique measures how tissue or an electrode resists alternating current across many frequencies. We couldn't confirm that Abbott or Senseonics uses it to correct sensitivity drift. The FDA summary for the first Eversense system describes a fluorescence-based sensor, one that detects glucose optically [8], and Abbott's engineers describe the FreeStyle Libre sensor as an enzyme-based electrochemical design [3]. Where impedance does appear is in research on measuring glucose itself: a 2023 battery-less implant estimated glucose from subcutaneous impedance in pigs [9].

Market Landscape

Dexcom, Abbott and Medtronic (now MiniMed) each took a different path away from fingerstick calibration:

• Dexcom: The Dexcom G7 includes an "urgent low soon" alert, meant to warn ahead of an urgent low rather than at it [10]. The FDA summary lists the alert but doesn't say how it is computed, and we couldn't verify that it uses machine learning. Its predecessor, the Dexcom G6, was the first CGM authorized in the FDA's integrated CGM (iCGM) category and was factory calibrated [11]. Dexcom said it would no longer manufacture the G6 after July 1, 2026, and recommends moving to the Dexcom G7 or Dexcom G7 15 Day [12].
• Abbott: Abbott leaned on hardware. Its engineers wrote that factory calibration depends on sensor stability and a robust manufacturing process, and that FreeStyle Libre sensors were the first commercially available factory-calibrated CGM sensors [3]. The original FreeStyle Libre 2 and Libre 3 sensors have since been discontinued. Abbott told prescribers they would be available through September 30, 2025, with the FreeStyle Libre 2 Plus and FreeStyle Libre 3 Plus sensors taking their place [13].
• Medtronic: Medtronic's standalone Guardian Connect system, approved by the FDA in 2018, paired the Guardian Sensor 3 with a phone app that handled fingerstick calibration and could alert when glucose was predicted to cross a level the user set [14]. Guardian Connect has since been superseded. Medtronic's diabetes business now operates as MiniMed, and its current sensors include Instinct, Simplera Sync, Guardian 4 and the standalone Simplera CGM [15]. Instinct is made by Abbott for MiniMed systems [20]. With the MiniMed 780G pump, the Guardian 4 sensor needs no calibration, though the pump still uses any blood glucose reading you enter to calibrate it, and the transmitter sends sensor-integrity data to the pump along with glucose values [16]. The standalone Simplera system, approved in 2024, removed calibration entirely [17].

Critical Issues and Future Directions

The hard problems sit where the math meets the body:

• Over-smoothing: Smoothing removes noise but adds delay, so a heavily smoothed trace can trail a fast-moving glucose level. In the Padova study, both filters added delay; at comparable noise reduction, the Kalman filter's delay was about 35% shorter than the moving average's [5]. How much smoothing a given commercial sensor applies isn't described in the sources we read.
• Compression Lows: A compression low is a falsely low reading caused by pressure on the sensor site, for example lying on it while sleeping [18]. Researchers built an algorithm to flag these events and tested it on more than 1,125 nights of in-home data. It caught 88.34% of them, and adjusting its settings cut false detections to 1.70% [18]. Their aim was to avoid unneeded pump shut-offs in a predictive low-glucose suspend system. How to respond to a reading you suspect is a compression low is a decision for your care team.
• Non-Invasive Future: Needle-free sensing in smartwatches and smart rings would have to pull a glucose signal out of very noisy optical data, which puts a heavy load on software. In February 2024 the FDA said it had not authorized any smartwatch or smart ring that measures or estimates blood glucose on its own, and it told people not to use them for that purpose [19]. That safety communication still carried its February 2024 date when we checked in October 2026.