📄 bioengineering

Controlled In Vitro Characterization of the Dynamic Response of Continuous Glucose Monitoring Systems: Adaptation of a Programmable Flow Platform and Decomposition of Dynamic Error

This study adapts a programmable flow-based in vitro platform to decouple the accuracy of generated glucose profiles from the dynamic response of continuous glucose monitoring (CGM) systems, proposing a novel set of metrics that decompose dynamic error into amplitude, rate, shape, and hysteresis components to reveal performance limitations obscured by traditional summary statistics like MARD.

Original authors: Khoroshun, E. V., Kozlov, V. A., Ivanov, I. V., Momynaliev, K.

Published 2026-08-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Khoroshun, E. V., Kozlov, V. A., Ivanov, I. V., Momynaliev, K.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to teach a robot to drive a car through a storm. You don't just want the robot to know where the car is right now; you need it to know exactly how fast the car is speeding up, how sharply it's turning, and whether it's about to slide into a ditch. In the world of diabetes care, the "robot" is a Continuous Glucose Monitoring (CGM) system—a tiny sensor worn on the skin that tracks blood sugar levels. For years, doctors and engineers have mostly asked one simple question: "Is the number the sensor shows close to the real number?" They used a single score, called MARD, to answer this. But just like a robot driver that knows the speedometer says 60 mph but doesn't realize the car is actually accelerating into a wall, a sensor can be "close" on average but terrible at predicting rapid changes. This is a big deal because modern diabetes treatment often uses these sensors to automatically deliver insulin. If the sensor is slow to react or gets the shape of the sugar curve wrong, the robot might give the wrong medicine at the wrong time.

This paper is like a high-tech driving test for these glucose sensors, but instead of a real car on a rainy road, the researchers built a "glucose simulator" in a lab. They wanted to see if they could break down exactly why a sensor makes a mistake. Instead of just saying "the sensor is 10% off," they wanted to know: Did it get the height of the sugar spike wrong? Did it get the speed of the rise wrong? Did it get confused when the sugar was going down versus going up? They built a machine that could mix sugar water to create perfect, predictable waves of rising and falling sugar levels. Then, they compared what the machine tried to make, what the machine actually made (because machines aren't perfect), and what the sensor reported.

The researchers found that you can't just trust the "recipe" you give the machine; you have to measure the actual soup it cooks up first. When they tested their "glucose simulator," they discovered that even though they programmed the machine to create a perfect wave, the actual sugar levels in the tube were slightly different—sometimes a bit higher, sometimes a bit lower, and sometimes the speed of the change wasn't exactly what they ordered. For example, when they tried to create a sudden jump from 5.5 to 12.0 mmol/L, the machine sometimes delivered a jump that was 6.931 mmol/L instead of the planned 6.5. If they had compared the sensor directly to the "recipe," they would have blamed the sensor for being wrong, when actually the machine had made a tiny mistake.

Once they fixed this by measuring the actual sugar levels in the tube, they looked at two different sensors, which we'll call Sensor A and Sensor B. They found that these sensors behaved very differently, and a single "average score" would have hidden the truth. Sensor A was like a cautious driver who always underestimated the speed; it saw the sugar rising but reported a much smaller jump (an amplitude transfer coefficient of about 0.65). It was slow to react, compressing the big changes into small ones. Sensor B, on the other hand, was like an over-enthusiastic driver who thought every bump was a mountain; it reported changes that were actually bigger than what was happening (an amplitude transfer coefficient of about 1.2).

The most interesting part was how the sensors handled the "direction" of the change. The researchers measured something called "hysteresis," which is like a laggy echo. Imagine walking up a hill and then walking back down; if your shadow didn't move with you, but stayed stuck on the way up, that's hysteresis. Sensor B had a huge hysteresis loop—its readings were very different depending on whether the sugar was going up or down, even if the sugar level was the same. Sensor A was much more consistent in this regard. The study showed that Sensor A had a terrible "average error" score (MARD of 33.70%), while Sensor B had a great score (MARD of 11.49%). But when they looked at the shape of the curve, Sensor A actually matched the pattern of the sugar changes better than Sensor B, even though its numbers were lower.

The main takeaway is that to truly understand how a glucose sensor works, you can't just look at one number. You have to separate the errors into different buckets: how much the sensor compresses the height of the wave, how fast it reacts to changes, and whether it gets confused about the direction. The authors suggest that for the future, we need to stop treating the "programmed" sugar level as the absolute truth and start measuring the "delivered" sugar level first. Only then can we tell if a sensor is truly good at keeping up with the fast and furious changes in our bodies, or if it's just a slow, confused driver that happens to have a good average score.

Technical Summary: Controlled In Vitro Characterization of the Dynamic Response of Continuous Glucose Monitoring Systems

Problem Statement
Continuous Glucose Monitoring (CGM) systems are increasingly critical for real-time decision-making, including automated insulin delivery (AID). While static accuracy metrics like Mean Absolute Relative Difference (MARD) are widely used, they fail to capture the system's ability to reproduce the direction, rate, amplitude, and shape of glucose concentration changes. A single MARD value cannot distinguish between various error types, such as constant sensor offset, amplitude compression, response rate delays, temporal misalignment, or hysteresis. Furthermore, existing laboratory test benches often compare CGM outputs directly against a programmed trajectory. This approach risks misattributing deviations caused by the test platform itself (e.g., flow dosing errors, mixing inefficiencies, or dead volume) to the sensor under test. There is a need for a controlled in vitro method that isolates the intrinsic dynamic response of the sensor from the performance of the glucose profile generation system.

Methodology
The study adapted a programmable flow-based in vitro platform to generate specific glucose (GLU) profiles and introduced a hierarchical analytical approach: "programmed trajectory — measured glucose — CGM output."

Platform Setup: GLU profiles were generated by programmably mixing two solutions (20.0 mmol/L glucose and zero-glucose) in a phosphate-buffered saline matrix at a constant total flow rate of 2.0 mL/min. The system utilized a macrofluidic manifold with an S-shaped channel to ensure spatial homogeneity, verified through randomized sampling at six control positions.
Reference Measurement: Crucially, the actual GLU concentration delivered to the sensor was independently measured using a SUPER GL2 glucose analyzer. This measured value, rather than the programmed setpoint, served as the reference for evaluating CGM performance.
Experimental Profiles: The study employed a diverse set of profiles to stress-test the system:
- Static levels (3.0, 5.5, 12.0, 20.0 mmol/L).
- Repeated dynamic cycles (5.5→12.0→5.5 mmol/L).
- Hypoglycemic profiles (6.0→3.0→6.0 mmol/L).
- Rate-controlled profiles (5.0→15.0→5.0 mmol/L) with transition times of 50, 100, and 200 minutes.
- A complex multiphase profile (4→18→3→12→5.5 mmol/L).
Metrics: Beyond standard metrics (Bias, MAE, RMSE, MARD), the study proposed a decomposition of dynamic error into specific components:
- Amplitude Transfer Coefficient ( $K_A$ ): Ratio of sensor amplitude change to measured GLU amplitude change.
- Rate Transfer Coefficients ( $K_{up}$ , $K_{down}$ ): Ratios of sensor rise/fall rates to measured GLU rates.
- Normalized Shape RMSE: Error in profile shape after removing offset and amplitude differences.
- Hysteresis Loop Area: Quantification of the difference in sensor readings at the same GLU level during rising vs. falling phases.
- Residual Shift: The difference in sensor error between the initial and final plateaus of a cycle.

Key Results

Platform Performance: The platform demonstrated spatial homogeneity and good repeatability (CV 0.61–2.00%). However, the actual measured GLU consistently deviated from the programmed setpoints. For instance, in the complex profile, the programmed 18 mmol/L plateau was not reached, with a mean measured value of 16.20 mmol/L. In hypoglycemic runs, the duration below 3.0 mmol/L varied significantly (15.2–72.6 min) despite similar time-to-threshold crossings.
Sensor Characterization: Two proof-of-concept sensors (CGM-A and CGM-B) exhibited distinct dynamic behaviors that standard metrics failed to fully differentiate:
- CGM-A showed negative bias and significant amplitude compression ( $K_A \approx 0.65$ ) and rate attenuation ( $K_{up} \approx 0.63$ ). Despite a high MARD (33.70%), it preserved the normalized shape of the profile well (low normalized RMSE).
- CGM-B showed positive bias and amplitude amplification ( $K_A \approx 1.13–1.23$ ) with rate amplification ( $K_{up} \approx 1.27–1.56$ ). It had a lower MARD (11.49%) but exhibited significantly larger hysteresis loop areas (5–10 times larger than CGM-A) and greater dynamic distortion.
Decomposition Utility: The study demonstrated that a sensor with lower absolute error (CGM-B) could possess worse dynamic fidelity (higher hysteresis, shape distortion) than a sensor with higher absolute error (CGM-A). The proposed metrics successfully separated these error sources.

Significance and Claims
The authors claim that the primary contribution of this work is the establishment of a rigorous analytical hierarchy where the measured glucose profile, not the programmed trajectory, serves as the reference standard. This prevents the conflation of platform generation errors with sensor response errors.

The study asserts that decomposing dynamic error into amplitude, rate, shape, and hysteresis components reveals critical differences between CGM systems that are invisible to MARD or correlation coefficients alone. Specifically, the authors note that:

MARD is insufficient: High correlation or low MARD does not guarantee accurate trend reproduction or absence of hysteresis.
Platform verification is mandatory: The "programmed trajectory" should be treated as a control setpoint, while the "measured glucose" is the true reference.
Clinical Relevance: As AID systems rely on trend data for dosing, the ability to quantify amplitude compression, rate lag, and hysteresis is essential for understanding potential dosing errors during rapid glycemic changes.

The authors conclude that while these metrics are not yet normative acceptance criteria, they provide a necessary technical framework for the next generation of CGM evaluation, moving from single-integral metrics toward multidimensional performance profiles. The study positions itself as a proof-of-concept for a standardized bench-testing approach that can support development, comparative technical evaluation, and root-cause analysis of dynamic errors.

More like this