AI and ML in Glucose Signal Processing

Introduction to AI and ML in Glucose Signal Processing

A CGM sensor doesn't report glucose directly. It produces a tiny electrical current from an enzyme reaction, and software turns that current into a number; in the FDA's review of the Dexcom G6, the transmitter converts the signal "using an algorithm" into a glucose value in mg/dL [3]. That software has to cope with random noise, the lag between blood glucose and the interstitial fluid around the sensor, and drift in the sensor's sensitivity over its wear [1]. Artificial intelligence (AI) and machine learning (ML, software that learns patterns from data instead of following hand-written rules) are one family of tools researchers have aimed at those problems. Our signal processing algorithms article covers the wider toolkit.

An earlier version of this page said AI and ML had replaced calibration in modern CGMs. They didn't. What removed routine fingerstick calibration was factory calibration, where calibration becomes part of manufacturing and is done under controlled lab conditions. Two Abbott engineers explained in 2017 why that only works when the sensor stays stable over its wear and manufacturing is consistent from lot to lot [2]. The FDA's March 2018 De Novo decision on the Dexcom G6 describes a system that "does not require user calibration" and calls it factory calibrated [3]. Dexcom says it will no longer manufacture the G6 after July 1, 2026 [5]. The Dexcom G7 15 Day, one of Dexcom's current sensors as of October 2026, is factory calibrated as well, with an optional fingerstick calibration [4].

Signal Processing Techniques

Labeling and FDA summaries say what a CGM's algorithm achieves, not how it is built. The G6 review names "an algorithm" and stops there [3], so the methods below come from published research rather than from any manufacturer's description of a commercial sensor.

A Kalman filter estimates the true signal from noisy measurements by weighing each new reading against a model of how glucose usually moves. A University of Padova group used one in the denoising module of a "smart" CGM layer that sits on top of a commercial sensor's output. The filter estimates the signal-to-noise ratio as it runs [1]. The same layer adds an accuracy-enhancement module and a prediction module, and the predictor they chose was deliberately simple: a first-order autoregressive model that forecasts from the most recent readings [1].

Simple models went further than we expected. In a 2007 proof-of-concept study of 28 volunteers with type 1 diabetes, a 30-minute prediction horizon let the researchers flag a coming low 20 to 25 minutes before it crossed the threshold [6]. Commercial CGMs do warn of "impending low" and "impending high" glucose, according to the G6 review [3]. Those studies measured how early a low can be predicted. They didn't show that predictive alerts improve health outcomes, and what to do when an alert fires is a decision for your care team.

Deep learning appears mostly in the research literature. A 2019 review of 55 studies on ML glucose prediction in type 1 diabetes found models built on recurrent neural networks (RNNs, which carry information from earlier readings forward; LSTM, or long short-term memory, networks are a common kind), deep and feed-forward neural networks, support vector machines and Gaussian processes [7]. Its authors concluded that no single model yet predicts accurately across lows, in-range glucose and highs, partly because meals and exercise are hard to capture as data [7].

Artifact Rejection and Personalization

A compression low is a false drop in readings when something presses on the sensor, such as lying on it during sleep [9]. In data from 72 people wearing an older Dexcom sensor, the G4 Platinum, compression artifacts lasted a median of 45 minutes and showed up more often at night [8]. Telling these apart from a real low is the job researchers call artifact rejection.

The published detectors are mostly research tools. One algorithm that used only CGM readings, tested on more than 1,125 nights of data from a predictive low-glucose suspend trial, caught 88% of pressure-induced drops; tuning it brought false detections down to 1.7% [9]. An ML classifier, a random forest trained on simulated sensor data, detected compression artifacts after the fact well enough that its authors called it ready for testing on real data [10]. Neither paper describes a commercial CGM running such a detector, and the G6 review says nothing about AI processing on the transmitter itself [3].

Personalized algorithms adapt a prediction model to one person's history. The 2019 review treats this kind of personal profile as central to glucose prediction, and it also found that little work had tested whether a model trained on one person, or one period, holds up for another [7]. Any change to how you manage glucose based on such a model is a decision for your care team.

Regulatory Considerations

The "black box" worry is that a neural network's reasoning is hard to inspect. In June 2024 the FDA, Health Canada and the UK's MHRA published guiding principles that define transparency for ML-enabled devices as clearly communicating a device's intended use, development, performance and, when available, its logic [11].

AI models also change after release, and the FDA has a process for that. Its final guidance on predetermined change control plans (PCCPs), dated August 2025, has a manufacturer describe planned modifications, how they will be developed and validated, and their expected impact. The FDA reviews that plan with the marketing submission, so each listed change doesn't need a new submission [12]. As of October 2026, then, the open question for AI in CGMs is less whether a regulatory path exists and more what evidence a manufacturer brings to it.

Conclusion

Software does a great deal of work inside a CGM, but factory calibration is what let people stop routine fingerstick calibration [2][3]. Kalman filters, neural-network predictors and ML artifact detectors are well documented in research [1][7][10]. What we still can't tell from public FDA summaries is how much of that research runs inside the sensors people wear, because those documents describe performance rather than methods [3].