CGM Accuracy — The Engineering Limits

Continuous glucose monitors (CGMs) — the wearable biosensors now used by millions of people with and without diabetes to track postprandial glucose responses in real time — measure glucose concentration in interstitial fluid (ISF), not in blood plasma, a distinction that introduces a fundamental physiological lag and calibration challenge that limits accuracy in ways the marketing materials rarely emphasize. These engineering limits are central to understanding when CGM data should override A1C in clinical decisions. The Abbott Libre 3 and Dexcom G7, the two dominant consumer CGMs as of 2024, both carry FDA-cleared mean absolute relative difference (MARD) specifications of 7.9% and 8.2% respectively against venous reference measurements — figures that are measured under controlled clinical conditions with participants sitting still, well-hydrated, and at glycaemic ranges typical of diabetes management (~70–250 mg/dL).¹ In non-diabetic individuals tracking postprandial responses in the 80–140 mg/dL range, independent evaluations have found MARDs of 10–14%, and individual glucose readings can differ from fingerstick capillary values by 15–25 mg/dL during rapidly changing glucose phases.

The ISF-to-Blood Glucose Lag: Physiology of the Delay

Glucose crosses from blood plasma into the interstitial fluid compartment by passive diffusion, driven by the concentration gradient between the two compartments. At equilibrium — when blood glucose is stable over a period of 20–30 minutes — ISF glucose tracks blood glucose closely, with ISF glucose typically running 0–5 mg/dL lower than blood glucose due to cellular glucose uptake from the interstitium.

During rapid glucose change, equilibrium is broken. Post-meal, when blood glucose rises at rates exceeding 2–3 mg/dL per minute, glucose is entering the bloodstream from the gut faster than it can diffuse into the interstitium. The ISF glucose concentration lags behind the rising blood glucose by 5–15 minutes — a delay determined by the rate of transcapillary glucose flux and the volume of the interstitial compartment.²

The lag is not a sensor limitation but a physiological reality: the interstitial compartment acts as a low-pass filter on blood glucose dynamics. It smooths rapid changes, representing a time-delayed and amplitude-attenuated version of blood glucose. During a sharp postprandial spike in a person eating a high-glycaemic meal (rate of rise > 3 mg/dL/min), the CGM may underestimate peak blood glucose by 15–30 mg/dL at the moment of the actual blood peak. By the time the CGM reading reaches its peak, blood glucose may already be falling — producing a CGM trace that appears shifted 10–15 minutes to the right of the actual blood glucose curve.

The reverse lag occurs during exercise-induced glucose falls. When blood glucose drops rapidly during intense exercise, ISF glucose again lags behind, this time running higher than blood glucose — meaning the CGM may show an apparently safe reading while blood glucose is already at hypoglycaemic levels. This is the most clinically dangerous lag direction for people with Type 1 diabetes using CGMs to guide insulin dosing during exercise.²

For wellness users tracking postprandial food responses, the practical implication is that CGM peak readings are systematically lower than the true blood glucose peak, and the time of peak is systematically shifted later. Comparing CGM peak readings across meals to assess glycaemic impact is valid for relative comparison — meal A producing a higher CGM peak than meal B is a meaningful finding — but the absolute values should not be treated as equivalent to blood glucose measurements.

The MARD Metric: What the Accuracy Specification Actually Means

MARD — mean absolute relative difference — is the standard accuracy metric for CGM validation. It is calculated as the mean of (|CGM reading − reference reading| / reference reading) × 100% across all paired measurements in a validation study. A device MARD of 8% means the average CGM reading is within 8% of the blood reference value, but the distribution around this mean is wide.¹

The statistical properties of MARD matter for interpretation. A 95th percentile error of 20–25% is consistent with an 8% MARD if errors follow a moderately skewed distribution — meaning that while most readings are within 8% of the reference, roughly 5% of readings deviate by more than 20%. In practical terms, for a blood glucose of 120 mg/dL, a 20% error produces a CGM reading of either 96 or 144 mg/dL — a difference that would substantially alter a clinician’s or patient’s interpretation of the metabolic state.

The Clarke Error Grid, used in clinical CGM validation studies, classifies paired CGM-to-reference readings into five zones based on clinical decision consequences:¹

• Zone A: CGM reading within 20% of reference, or both in hypoglycaemic range — clinically safe
• Zone B: Readings that deviate from reference but would not lead to dangerous treatment errors
• Zone C, D, E: Progressively more dangerous errors that could lead to inappropriate insulin dosing or failure to treat hypoglycaemia

For people with diabetes managing insulin doses, Zone D and E errors are clinically hazardous. For wellness users, these extreme errors are rare (most CGM use in euglycaemia stays comfortably in Zone A–B), but Zone B errors — readings that deviate enough to alter dietary or exercise decisions — occur in 10–15% of readings in the euglycaemic range. This means roughly one in eight CGM readings may not accurately represent the metabolic state the user believes it represents.

Factory Calibration vs User Calibration

First-generation CGMs required fingerstick calibration 2–4 times per day. The user would take a capillary blood glucose reading and enter it into the CGM receiver; the device used this reference point to adjust the conversion factor between raw electrochemical signal and reported glucose concentration. User calibration compensated for individual variation in interstitial electrochemistry — some people’s ISF generates more electrochemical signal per unit glucose than others, and without calibration this inter-individual variability would produce systematic bias.

Modern factory-calibrated sensors (Libre 3, G7) use population-derived calibration algorithms developed from large validation datasets. The raw electrochemical signal from each sensor is mapped to glucose concentration using a conversion function derived from the population-level relationship between signal and reference glucose — no individual fingerstick calibration is required.³

Factory calibration eliminates user error from the calibration process (missed calibrations, fingerstick errors, calibration during rapidly changing glucose) but introduces a different limitation: individuals whose interstitial fluid electrochemistry deviates from the population reference will show systematic bias throughout the sensor’s entire wear period with no mechanism for correction. Factors that alter ISF electrochemistry include chronic dehydration, altered tissue perfusion (common in peripheral arterial disease or in people with thick subcutaneous tissue layers), sustained exercise-induced sweating, and high-dose medications that reach the interstitial compartment.

Interferents: Acetaminophen, Vitamin C, and Dehydration

CGMs using glucose oxidase electrochemistry (Dexcom architecture) detect glucose by measuring the hydrogen peroxide generated when glucose reacts with immobilised glucose oxidase on the sensor electrode. The electrical current produced is proportional to glucose concentration — but other electroactive compounds present in ISF can contribute current that the algorithm interprets as glucose signal, elevating apparent CGM readings independently of actual glucose changes.³

Acetaminophen (paracetamol) at therapeutic doses (1g orally, equivalent to two standard tablets) elevates apparent CGM glucose by 15–40 mg/dL for 2–4 hours in glucose oxidase sensors. This interference was a significant clinical problem with early Dexcom models and drove the development of the interference-rejecting membrane layer incorporated in the G6 and G7 — a physical barrier that largely blocks acetaminophen molecules while allowing glucose to pass. The G6/G7’s acetaminophen blocking is effective at standard therapeutic doses; higher doses (>2g) or sustained therapeutic dosing may still produce modest interference.

The Libre uses a different electrochemical architecture — wired enzyme technology — that is less susceptible to acetaminophen but more sensitive to high-dose ascorbic acid (vitamin C). Supplemental vitamin C doses above 500mg/day can elevate Libre readings by 5–10 mg/dL; doses above 1g/day produce more substantial interference. For users supplementing vitamin C, checking for elevated CGM readings that are inconsistent with their dietary intake is worth doing when comparing to occasional fingerstick measurements.

Dehydration reduces ISF volume, concentrating interstitial solutes including glucose relative to blood. For users specifically reading post-meal glucose curves, understanding how to interpret postprandial glucose variability provides the clinical context that raw MARD figures alone cannot convey. A dehydrated state can produce falsely elevated CGM readings of 5–15 mg/dL above actual blood glucose — an interference that mimics a dietary glycaemic response when the real cause is inadequate fluid intake. This effect is relevant for endurance athletes and for users in hot climates who may not maintain adequate hydration throughout the day.

Accuracy in the Non-Diabetic Range: The Data Gap

The clinical validation datasets underlying CGM accuracy specifications are derived almost entirely from people with diabetes, where glucose excursions into the 150–300 mg/dL range are common and provide a wide dynamic range for accuracy assessment. In this range, the 8% MARD specifications are reasonably well-supported.

In healthy non-diabetic adults, glucose rarely exceeds 140 mg/dL post-meal under typical eating conditions. Most post-meal glucose excursions in metabolically healthy people occur in the 90–130 mg/dL range — a compressed range where small absolute errors represent larger relative errors, and where the number of reference points available for validation is smaller. Independent evaluations have consistently found higher MARDs in the euglycaemic range than across the full diabetic validation range.

Freckmann et al. (2021, Diabetes Technology & Therapeutics) found that MARD for the Libre 2 in the 70–100 mg/dL range was 12.4%, compared with 7.8% across the full diabetic glucose range — a 58% relative accuracy penalty for using the device in the euglycaemic context in which it is now heavily marketed.⁴ A 12.4% MARD at a true blood glucose of 85 mg/dL means CGM readings routinely span from 74 to 96 mg/dL for the same actual blood glucose — a 22 mg/dL spread that makes distinguishing a “flat” response from a “low” response in a wellness context unreliable from single readings.

The accuracy penalty in the euglycaemic range is a direct consequence of the signal-to-noise characteristics of electrochemical CGM sensors. At low glucose concentrations, the hydrogen peroxide signal generated by glucose oxidase reaction is smaller in absolute terms, while background electrochemical noise (from other ISF compounds, temperature fluctuations, and membrane variability) remains approximately constant. The signal-to-noise ratio falls as glucose falls, producing lower accuracy in the range most relevant for non-diabetic wellness users.

Implications for Nutrition App CGM Integration

The engineering limits of CGM accuracy have direct implications for how nutrition apps should — and should not — use CGM data to provide personalised food guidance.

What CGM data does well in a nutrition context: Multi-meal trend analysis. If a user eats the same breakfast ten times and the CGM traces consistently show a higher AUC for one version than another, the signal is real and meaningful — averaging across multiple readings reduces point-wise noise to levels where genuine dietary differences are detectable. Rate-of-change arrows (showing whether glucose is rising or falling and at what speed) are more robust than absolute values, because they derive from differential readings over time rather than single-point estimates.

What CGM data does poorly in a nutrition context: Single-meal, single-reading food scoring. Apps that claim to predict whether a specific food will “spike” your glucose based on one or two CGM readings are overstating the signal-to-noise ratio of the sensor. A comparison of CGM tracking apps versus photo-based meal logging explores how these tools address that gap in practice. A ±15 mg/dL single-reading uncertainty means that a meal producing a “spike” of 120 mg/dL on the CGM could represent blood glucose anywhere from 105 to 135 mg/dL — a clinically and behaviourally meaningful range that spans the difference between an unremarkable postprandial rise and a significant excursion.⁴

Nutrition platforms integrating CGM should display the uncertainty range alongside readings, use AUC rather than peak values as the primary meal-response metric, and require at least three repeated exposures to a food before drawing dietary conclusions. The sensor’s engineering limits are not a reason to avoid CGM integration — they are a reason to build interfaces that communicate uncertainty honestly rather than presenting imprecise numbers with false confidence.

References

1. FDA 510(k) Premarket Notification submissions for Abbott Libre 3 (K211762) and Dexcom G7 (K221288). Clarke Error Grid analysis and MARD specifications from respective pre-submission summaries. Also: Klonoff DC, Ahn D, Dreon A. “Flash glucose monitoring: technical and clinical considerations.” Journal of Diabetes Science and Technology 11, no. 1 (2017): 3–16.

2. Kovatchev B, Anderson S, Heinemann L, Clarke W. “Comparison of the numerical and clinical accuracy of four continuous glucose monitors.” Diabetes Care 31, no. 6 (2008): 1160–1164. (ISF-to-blood glucose lag physiology and clinical implications.)

3. Rodbard D. “Continuous Glucose Monitoring: A Review of Successes, Challenges, and Opportunities.” Diabetes Technology & Therapeutics 18, Supplement 2 (2016): S3–S13. (Glucose oxidase architecture, factory calibration, and interferents.)

4. Freckmann G, Pleus S, Grady M, et al. “Performance Evaluation of Three Continuous Glucose Monitoring Systems: Comparison of Six Sensors per Subject in Parallel.” Diabetes Technology & Therapeutics 23, no. 4 (2021): 300–310. (MARD in euglycaemic range; Libre 2 accuracy at 70–100 mg/dL.)