| Jan 19, 2026 |
Dual closed-loop insulin system adds chemical safeguard to protect against dangerous overdosesA wearable device pairs glucose-responsive insulin with algorithmic pump control, creating dual safety loops that reduced hypoglycemia from 4.01% to 0.52% in diabetic rats. |
| (Nanowerk Spotlight) An insulin overdose can kill in hours. For people with diabetes who depend on automated pump systems, this risk shadows every sensor glitch and algorithmic miscalculation. The devices promise freedom from constant manual management, but when glucose readings are wrong, consequences can be fatal. |
| Diabetes management has transformed since patients first gained the ability to monitor their own blood sugar at home. Portable glucose meters replaced laboratory-dependent testing. Insulin pumps emerged to deliver continuous subcutaneous doses, eliminating multiple daily injections. Continuous glucose monitors followed, tracking levels around the clock through sensors worn on the body. |
| These technologies merged into closed-loop systems, sometimes called artificial pancreas devices. A sensor feeds real-time glucose data to an algorithm, which calculates insulin doses and commands a pump to deliver them automatically. Several such systems have reached patients and outperformed traditional manual regimens. |
| But they share a critical flaw. |
| Every decision depends on sensor accuracy. A drifting calibration or transient malfunction can trigger insulin delivery based on faulty data. Excess insulin forces blood sugar dangerously low, causing hypoglycemia, a condition that brings seizures, unconsciousness, or death. |
| Researchers have pursued an alternative: engineering insulin molecules that respond directly to glucose concentrations. These glucose-responsive formulations release insulin only when surrounding glucose rises, creating a chemical safety net independent of electronic sensors. Three main approaches exist, employing glucose-binding proteins, glucose oxidase enzymes, or boronic acid compounds. Yet glucose-responsive insulins typically reach patients through passive patches, pills, or manual injections, methods lacking the precision of electronic pumps. |
| Neither strategy alone solves the safety problem completely. |
| A team spanning the University of Hong Kong, Guangzhou Medical University, and Zhejiang University has now unified both approaches in a single wearable device. Publishing in Advanced Materials ("A Wearable, Dual Closed‐loop Insulin Delivery System for Precision Diabetes Management"), the researchers describe a dual closed-loop system that layers chemical responsiveness onto AI-driven electronic control. |
 |
| The overall concept of the proposed dual closed-loop insulin delivery system (DuoLoop). a) Illustration of the key components of DuoLoop, including the continuous glucose monitors (CGM), algorithm, and insulin pump. b) Cross-sectional illustration depicting the interrelationship among the components in DuoLoop. Schematics of the workflow and mechanism of c) quantized edge-AI implementation of the dual closed-loop system using a PID learning framework. d) Comparative schematic of glucose dynamics between the commercial SinLoop and the DuoLoop. Compared to the traditional single-loop electrical closed-loop system, DuoLoop demonstrates reduced glucose fluctuations and a lower incidence of hyperglycemia and hypoglycemia. Kp : Proportional gain, Ki: Integral gain, Kd: Derivative gain. (Image: Reproduced from DOI:10.1002/adma.202514945, CC BY) (click on image to enlarge) |
| The first loop operates through hardware and deep learning: a continuous glucose monitor tracks blood sugar, a Transformer neural network predicts where levels are heading, and a pump delivers insulin accordingly. The second loop operates through chemistry: the insulin itself releases faster when glucose is high and slower when glucose is low. Should the electronic system err, the insulin's molecular properties provide backup protection. |
| The team built their glucose-responsive insulin by modifying a biodegradable polymer called poly-L-lysine. They attached polyethylene glycol and a fluorinated boronic acid compound, creating a material that carries a positive electrical charge. This charge attracts negatively charged human insulin molecules. When mixed, the polymer and insulin self-assemble into spherical particles roughly 100 nm in diameter, about one-thousandth the width of a human hair and small enough to pass through needles and pump tubing without clogging. |
| The release mechanism exploits straightforward chemistry. Glucose molecules in the body bind to the boronic acid groups on the polymer surface, neutralizing some positive charge and weakening the electrostatic grip on insulin. Insulin escapes into the bloodstream. When glucose levels fall, fewer glucose molecules bind, the positive charge persists, and insulin remains locked in place. The result is automatic dose adjustment at the molecular level, requiring no sensor. |
| Testing in diabetic rats demonstrated the advantage. A single injection of glucose-responsive insulin maintained normal blood sugar for 6 to 10 hours. Conventional insulin at equivalent doses controlled levels for only about 3 hours. The extended action delivers more stable glucose control with fewer injections. |
| The predictive algorithm powering the first loop uses a Transformer architecture, the same deep learning framework underlying many modern AI applications. The model processes 100 minutes of historical glucose readings and predicts levels 30 minutes ahead, giving the system time to act before problems develop rather than merely reacting after they occur. |
| Training proceeded in two stages. The team first used simulated patients generated by validated diabetes modeling software, then refined the model with real data from diabetic rats monitored every five minutes over 10 days. The trained algorithm achieved an R² value of 98.03% on rats included in training and 97.83% on rats excluded from training. R² measures prediction accuracy on a scale where 100% represents perfect agreement between predicted and actual values. These figures indicate the model generalizes reliably to new subjects. |
| Predictions from the neural network feed into a controller governing insulin delivery. This controller follows a proportional-integral-derivative framework, a standard engineering approach that adjusts output based on current error, accumulated past error, and rate of change. Conventional implementations estimate rate of change from past measurements alone, making them purely reactive. |
| The researchers modified this design by incorporating the Transformer's forward-looking predictions. Their controller anticipates where glucose is heading rather than merely responding to where it has been. They compressed the neural network to one quarter of its original size, enabling it to run on a mobile phone processor without internet connectivity. |
| The team assembled a compact prototype: a Bluetooth-enabled pump, an insulin reservoir, and embedded control software. In diabetic rats, following the algorithm's recommendations kept blood sugar within normal bounds. Deliberately reducing doses to 80% of recommendations caused hyperglycemia. Increasing doses to 120% triggered hypoglycemia. These results confirmed the algorithm identified correct values. |
| The critical test compared the complete dual-loop system against a conventional single-loop system using standard insulin. Results diverged sharply. Animals receiving dual-loop treatment spent 98.82% of the observation period with normal glucose levels, versus 92.10% for single-loop treatment. Time in hypoglycemia dropped from 4.01% to 0.52%. Time in hyperglycemia fell from 3.89% to 0.65%. Glucose variability decreased substantially by both standard deviation and coefficient of variation. Tissue examination at injection sites revealed no notable inflammation or scarring, suggesting good tolerability. |
| Limitations remain. All validation occurred in rats, not humans. The animals ate freely, whereas real diabetes management demands attention to diet. Long-term durability of the wearable components has not been assessed. |
| Nonetheless, the work demonstrates that layering chemical safeguards onto AI-powered electronic control can address a fundamental weakness of sensor-dependent devices. Sensors will occasionally fail or drift. A dual-loop architecture cannot prevent measurement errors but can blunt their consequences. If human trials confirm these findings, similar designs might extend to other conditions demanding precise, responsive drug delivery. |
By
Michael
Berger
– Michael is author of four books by the Royal Society of Chemistry:
Nano-Society: Pushing the Boundaries of Technology (2009),
Nanotechnology: The Future is Tiny (2016),
Nanoengineering: The Skills and Tools Making Technology Invisible (2019), and
Waste not! How Nanotechnologies Can Increase Efficiencies Throughout Society (2025)
Copyright ©
Nanowerk LLC
|
|
ORCID information
|
Become a Spotlight guest author! Join our large and growing group of guest contributors. Have you just published a scientific paper or have other exciting developments to share with the nanotechnology community? Here is how to publish on nanowerk.com. |
