Closed-Loop Algorithms with Multi-Analyte Data

Automated insulin delivery (AID) systems on the market steer insulin from one measured signal: glucose, read by a continuous glucose monitor (CGM). A 2025 review of artificial pancreas research describes current AID systems as relying "on a single sensor source, CGM, to adjust insulin" [1]. "Multi-analyte" closed-loop control means giving the algorithm more to work with, such as ketones, lactate, insulin itself or heart rate from a wristband. Small trials suggest the idea works. As of October 2026 the first glucose-ketone sensor has FDA authorization [2], but its pairing with pump algorithms is still announced rather than shipped [3].

How today's algorithms decide, using glucose alone

A closed-loop algorithm is software that reads each new CGM value, predicts where glucose is heading, and tells the pump to deliver more or less insulin. Tandem, for example, says its Control-IQ+ technology "uses CGM sensor values to predict glucose levels 30 minutes ahead and automatically adjust insulin" [4]. Our explainer on PID vs. MPC algorithms covers the two main controller designs.

Everything else the algorithm knows is typed in by the user or estimated. Insulin on board, the insulin from earlier doses that is still working, is calculated from generic insulin action curves rather than measured. A 2023 perspective in Nature Reviews Endocrinology notes that closed-loop systems "rely solely" on glucose, not on insulin levels or markers such as cortisol, beta-hydroxybutyrate (the main ketone) or adrenaline [5].

Here's what surprised us. The FDA regulation for interoperable automated glycemic controllers (iAGCs), the device type for AID algorithms that communicate with connected pumps and sensors, already defines the device as calculating doses "based on inputs such as glucose and other relevant physiological parameters" [6]. The rule leaves room for extra inputs. The products haven't used it yet.

Why glucose alone leaves gaps

Glucose-only control already works well. In a 6-month randomized trial of 168 people with type 1 diabetes, time in the target range of 70 to 180 mg/dL rose from 61% to 71% with the Control-IQ closed-loop system on a t:slim X2 pump, while a sensor-augmented pump group stayed at 59% [7]. As of October 2026, Tandem's t:slim X2 page describes the pump as running Control-IQ+ [4].

The weak spots are physical. A CGM measures glucose in interstitial fluid, the fluid between cells, so its reading trails blood glucose, especially when glucose changes fast. Insulin delivered under the skin also takes time to act. An EASD/ADA consensus report lists both as core limits of AID, which is why first-generation systems need users to announce meals and upcoming physical activity [8]. The 2025 review calls these delays the reason today's systems are "hybrid" rather than fully automated [1].

Exercise shows the gap most clearly. In a 2025 crossover trial, 38 people using Omnipod 5 completed a 70-minute treadmill session under three conditions. Switching on the system's Activity feature beforehand cut insulin delivery and softened the glucose drop. The share who developed hypoglycemia during exercise was 42% in standard automated mode and 29% or 24% with the feature, a difference that wasn't statistically significant [9]. Here the person is the activity sensor: the algorithm only knows about exercise because someone told it.

Ketones: the input closest to market

Ketones are compounds that build up when the body doesn't have enough insulin, and the 2025 review describes them as a signal of AID "system performance" [1]. High levels can lead to diabetic ketoacidosis (DKA), a life-threatening complication [10].

A 2024 review of continuous ketone monitoring notes that AID algorithms "use glucose levels as their primary input and do not consider ketone levels." After high-intensity exercise, ketones can rise while glucose falls. A glucose-only algorithm answers the falling glucose by delivering less insulin, which the authors say risks pushing ketones higher, and they argue a ketone-aware system should signal the user instead [10]. How to respond to a ketone reading is a decision to make with your care team.

On August 25, 2026, the FDA granted Abbott's De Novo request for the Libre Duo 10 Day Continuous Dual Glucose Ketone Monitoring System, creating a new device type [2]. Abbott's announcement said it planned a US launch later in 2026 and expected the sensor to work with Beta Bionics' iLet and Sequel Med Tech's twiist AID systems by the end of 2026, with Insulet, Tandem and MiniMed integrations expected in 2027 [3]. It says the AID systems will "connect with the sensors," not whether any algorithm will use the ketone reading in its dosing [3]. Our article on multi-analyte sensing in CGMs covers the sensor side.

Heart rate, movement and lactate

In a 2014 crossover trial, 12 people with type 1 diabetes did 30 minutes of mild exercise on a closed-loop system, once with heart rate feeding the algorithm and once without. With heart rate added, glucose fell significantly less during exercise. Hypoglycemic events during exercise (none versus two) and time in range (81% versus 75%) also favored the heart-rate version, though neither difference was statistically significant [11].

A team at the Illinois Institute of Technology built a multivariable AID system that adds physiological signals from a wristband and uses artificial intelligence to detect physical activity and estimate its intensity, with no announcement from the user. In pilot clinical studies it showed more time in range and less hypoglycemia after unannounced activity and meals [12].

Lactate is the chemical version of the same idea. Its level rises with physical activity, and the 2025 review suggests a continuous lactate reading could help an artificial pancreas "identify exercise without user input," listing it under future multiplexed sensors [1]. Our sensor fusion in AID article covers how several signals get combined.

Measuring insulin itself

The most ambitious input is insulin. Centralized laboratory tests can't deliver timely results, so the 2023 perspective calls for frequent, real-time insulin measurement and reviews early work on disposable strips and wearable insulin sensors [5]. A live reading would replace estimated insulin on board with a measured value and help account for how differently people respond to the same dose [5]. The 2025 review files it with ketone and lactate sensing as a future input [1].

What still stands in the way

Each new input is another sensor that can be wrong and another wireless link to secure. The consensus report stresses that AID components can fail [8], and the insulin-sensing perspective calls data privacy and security "a major concern" [5]. The FDA controller rule requires specifications for connected devices "eligible to provide input" to the algorithm, with glucose sensor performance as its example [6]. A ketone or heart-rate input would need its own performance case before an algorithm could act on it.

The near-term question is concrete. When the first Libre Duo pairings with iLet and twiist arrive, will the ketone reading only feed an alert, or will it enter the dosing math? Abbott's announcement doesn't say [3].