Non-Adjunctive Meal Detection Algorithms

Introduction to Non-Adjunctive Meal Detection Algorithms

A meal detection algorithm watches the glucose trace from a continuous glucose monitor (CGM) and decides, without being told, that someone has eaten. The goal is an automated insulin delivery (AID) system that can respond to a meal the user never entered, which researchers call the first step toward fully closing the loop [1]. As of October 2026, that is still research. A January 2026 review found that commercial AID systems remain dependent on manual meal announcement, and that no manufacturer had regulatory clearance to run without it [2].

A note on the word "non-adjunctive"

The title borrows a term that means something else in CGM regulation. A non-adjunctive CGM is one whose readings can be used for treatment decisions without a confirming fingerstick. FDA first approved that use for the Dexcom G5 Mobile in December 2016, and its database now lists that approval as withdrawn on March 20, 2026 [3]. The same claim appears in the 2025 clearance for the Dexcom G7 15 Day, one of Dexcom's current sensors as of October 2026 [4].

Meal detection is a separate idea. It needs a CGM reliable enough to act on, but what it removes is the meal announcement, not the fingerstick.

Why unannounced meals are hard

Meals are where automated systems struggle most, and the reasons stack. Rapid-acting insulin is slow to start working. CGM readings also trail blood glucose by about 5 to 15 minutes, because the sensor measures glucose in interstitial fluid, the fluid between cells under the skin [1]. An algorithm that waits for the CGM to show a rise is already behind. That is why hybrid closed-loop systems still ask for a premeal bolus, the dose given when a meal is entered; it remains the most effective way to blunt the after-meal rise [1].

How the algorithms work

Most approaches fall into two families. The first uses a physiological model. Ramkissoon and colleagues ran a mathematical model of glucose and insulin through an Unscented Kalman Filter, a statistical tool that estimates hidden quantities from noisy measurements, to track a "disturbance" term that climbs when glucose from food enters the blood [1]. When that estimate and the CGM signal moved together strongly enough over a sliding window, the algorithm flagged a meal [1].

The second family learns from data. Daniels and colleagues used long short-term memory (LSTM) networks, a type of neural network built for time series, to estimate the recent glucose trajectory with a 95% prediction band. When real CGM readings stayed outside that band, the algorithm treated the gap as an unannounced meal [5]. In simulation it caught meals and snacks with 93% precision and 76% recall, and the average flag came 38 minutes after eating began [5].

Some groups add signals beyond glucose, an approach we cover in Sensor Fusion in Automated Insulin Delivery (AID). A 2022 review found heart rate and accelerometer data from wearables among the most common extra inputs, used mostly to detect exercise but also for meal detection and to offset sensing delays [6]. One feasibility trial paired a MiniMed 780G pump with a smartwatch app that detects eating and drinking hand gestures and converts them into carbohydrate amounts for an automatic bolus [7].

The core trade-off: speed versus false alarms

Here's what surprised us in the tuning results. One algorithm can be set to catch nearly every meal or to almost never raise a false alarm, but not both at once. In Ramkissoon's 14-day simulation, the most sensitive setting caught 99% of meals at an average of 28 minutes, with about one false detection per day [1]. The setting tuned against false alarms had essentially none. It also caught only 47% of meals and took 46 minutes on average [1]. A middle setting caught 93% at 37 minutes, with 0.2 false detections per day [1].

False alarms matter because of what follows them. Insulin delivered for a meal that never happened can push glucose too low, which led one group to weight detection by each person's usual hourly meal schedule to cut false detections [8]. Late or missed detections cost the other way, and glucose runs high.

What clinical studies show

In a randomized crossover trial at Oregon Health & Science University, a system with a neural-network meal detector found meals with 83.3% sensitivity and a mean detection time of 25.9 minutes; 16.6% of its detections were false [9]. Compared with a standard model predictive control algorithm, it cut time above 180 mg/dL by 10.8% in the four hours after unannounced meals [9].

FCL@Home, run at the University of Virginia and two other U.S. centers, tested the AIDANET system with no meal announcements in 34 participants aged 14 and older. Mean sensor glucose fell from 178 mg/dL under usual care to 164 mg/dL, the gains came mainly overnight, and time spent low was non-inferior [10]. The CLOSE IT trial randomized 73 adults on an open-source AID system to keep announcing meals or stop for 12 weeks. Time in range (70 to 180 mg/dL) ended at 66% without announcements and 69% with them, and the adjusted difference was not statistically significant [11].

Algorithms not designed for unannounced meals show the gap. When 14 adults used the MiniMed 780G for 90 days at home without announcing meals of up to 80 grams of carbohydrate, time in range was 67.5%, compared with 77.7% during 90 days with announcements, though time spent low went down [12].

How close commercial systems get

Two FDA-authorized systems come closest, and both still ask for some meal input. FDA's summary for the MiniMed 780G, approved in April 2023, says it can deliver an auto correction bolus without user input based on sensor glucose [13]. Medtronic Diabetes engineers, writing in 2024, describe its algorithm as including meal detection technology [14]. The Beta Bionics iLet Bionic Pancreas, cleared in May 2023, replaces carb counting with a meal announcement in which users rate a meal as small, medium or large, and its software commands correction doses on its own [15].

Open-source systems are further along in practice. The 2026 review reports that some users of the OpenAPS algorithm run without meal announcements, and warns that this takes expert configuration and that rapid changes in insulin sensitivity, such as during exercise, remain a challenge [2]. Whether any of these systems suits a particular person is a decision for their care team.

What "fully closed loop" should mean

Researchers don't agree on the term yet. Some use "fully closed loop" for any system that needs no meal announcement; others reserve it for systems that need no user input at all [2]. In October 2026, an international expert group proposed a shared definition graded on time in tight range, time below range and validated person-reported outcomes. Every qualifying system has to hit its targets without mandatory meal or activity announcements. The group places current commercial systems in the hybrid category because they need meal and exercise announcements to work at their best [16].