How aCGM Sensor Study Became a Lesson in What Accuracy Numbers Really Mean

Continuous glucose monitoring has quietly become one of the most consequential technologies in modern diabetes care. By sampling interstitial fluid every few minutes and streaming glucose values to smartphones and insulin pumps, these sensors allow people with type 1 and type 2 diabetes to see trends, avert dangerous hypoglycemia, and fine-tune therapy in ways that fingerstick testing never could. Yet as sensor chemistry and signal-processing algorithms have improved, an awkward truth has emerged: the methods used to judge how accurate these devices actually are have not kept pace. A new perspective published in Frontiers in Clinical Diabetes and Healthcare argues that the field now needs a standardized validation framework, and it uses a recent real-world study of the Sinocare iCan i3 sensor as its teaching case.

The perspective, authored by Auda Ameur of the University Hospital of Oran in Algeria, does not set out to attack a single device or a single study. Instead, it takes the 2025 real-world evaluation of the Sinocare iCan i3 system, in which 36 participants with type 1 or type 2 diabetes wore the sensor for 15 consecutive days, and dissects the methodological choices that shaped its headline results. The study reported an overall mean absolute relative difference, or MARD, of 17.2 percent, with 69.5 percent of paired CGM-reference measurements meeting an agreement criterion of plus or minus 20 mg/dL or 20 percent, and 77.1 percent of sensors surviving the full wear period. Those numbers are scientifically legitimate, Ameur argues, but they cannot be interpreted properly without understanding how they were produced.

The first and perhaps most fundamental issue is the reference standard. In the Sinocare evaluation, CGM readings were compared against capillary glucose measurements from the FreeStyle Optium blood glucose meter rather than against a laboratory reference method such as the Yellow Springs Instrument analyzer, or YSI, which is considered the gold standard for glucose measurement. This choice reflects everyday clinical practice and enables home-based evaluation, but it introduces a hidden complication: blood glucose meters have their own measurement error, bias, and imprecision. The resulting MARD therefore reflects not just the performance of the CGM sensor but the combined variability of two imperfect instruments, which may push the reported accuracy figure either up or down. Comparing that composite figure directly with MARD values from studies that used YSI laboratory references is, Ameur warns, like comparing race times recorded on different tracks.

The contrast is striking when contemporary validation studies are placed side by side. The Dexcom G7 achieved a MARD of roughly 8.2 percent in a large multicenter study using YSI reference measurements, the FreeStyle Libre 3 reported approximately 7.8 percent under similar laboratory conditions, and the Medtronic Simplera reached about 8.7 percent. The Sinocare iCan i3 figure of 17.2 percent is considerably higher, but Ameur cautions that differences in reference methods, participant selection, glucose distributions, pairing strategies, and study settings can substantially influence reported accuracy estimates. The apparent gap may reflect genuine sensor characteristics, methodological differences, or some combination of both. As analytical performance among current-generation sensors converges, study methodology itself may increasingly determine which numbers appear in the literature.

MARD itself comes under scrutiny. The metric, which averages the absolute percentage difference between sensor and reference values, remains the single most cited number in CGM validation, but Ameur emphasizes that no one summary statistic can capture the multidimensional performance of a modern sensor. MARD is highly dependent on the glucose distribution of the study population and on glycemic variability, meaning two devices with identical MARD values can behave very differently in clinical use. The perspective recommends reporting complementary measures, including median absolute relative difference, mean absolute difference during hypoglycemia, root mean square error, Bland-Altman analysis, and the distribution of errors across glucose ranges. The fact that roughly 30 percent of paired measurements in the Sinocare study fell outside the predefined agreement threshold illustrates why error distributions matter more than a single average.

Timing introduces another subtle but important source of distortion. CGM sensors measure glucose in interstitial fluid, which lags behind blood glucose by several minutes, especially during rapid excursions after meals or during insulin action. The Sinocare study paired CGM and fingerstick readings within a plus or minus 5-minute window. During stable glucose periods this flexibility is harmless, but during fast swings the physiological delay between blood and interstitial glucose can generate systematic disagreement that has nothing to do with sensor quality. Ameur recommends that future studies run sensitivity analyses using tighter matching windows of one, two, and five minutes, and stratify results by the rate of glucose change, so that physiological lag can be separated from genuine measurement error.

The sensor survival figure raises a different statistical concern. Because nearly a quarter of the sensors stopped functioning before the intended 15-day wear ended, the pooled accuracy estimates may be affected by survival bias if prematurely failed sensors differed systematically in performance from those that lasted. Without comparing accuracy between discontinued and completed sensors, the magnitude and direction of such bias cannot be determined. Ameur proposes time-to-event methods, including Kaplan-Meier survival analysis and Cox proportional hazards regression, to characterize sensor durability and identify predictors of premature failure. Related to this is the absence of longitudinal performance assessment: post-insertion tissue responses, sensor stabilization, enzyme degradation, and biofouling can all degrade accuracy over time, yet pooling all readings hides whether a sensor drifts toward the end of its wear period. Mixed-effects regression models, Ameur suggests, could detect these temporal changes while accounting for repeated measurements within participants.

Statistical uncertainty and missing data receive similar treatment. The original study reported point estimates without confidence intervals, leaving readers unable to judge how precise the accuracy figures actually are in a cohort of only 36 participants. Bootstrap resampling, Ameur notes, offers a practical way to quantify uncertainty in small studies with repeated glucose measurements. Real-world CGM studies also inevitably generate missing observations from sensor interruptions and technical failures, and the underlying missingness mechanism, whether data are missing completely at random, missing at random, or missing not at random, determines whether simple exclusion introduces bias. The perspective additionally calls for predefined subgroup analyses by diabetes type, age, body mass index, HbA1c, insulin regimen, and glycemic variability, with particular attention to the hypoglycemic range, where inaccurate readings carry the greatest safety consequences.

Transparency and the distinction between analytical and clinical accuracy round out the appraisal. The original study appropriately disclosed its funding and conflicts of interest, but Ameur observes that the reference meter itself, the FreeStyle Optium, came from a manufacturer with a commercial presence in the field, and recommends independent validation with laboratory references and independent statistical review as further safeguards, while stressing that such relationships do not by themselves indicate bias. He also distinguishes analytical accuracy, which measures agreement with reference values, from clinical performance, which encompasses time in range, hypoglycemia burden, patient-reported outcomes, and treatment decisions. Error grid analyses, including the Clarke, Parkes, and newer Surveillance Error Grids, help bridge that gap but do not by themselves establish real-world effectiveness.

The perspective culminates in a proposed multidomain checklist spanning reference methodology, complementary accuracy metrics, temporal pairing, longitudinal performance, sensor survival, statistical uncertainty, missing data, subgroups, and reporting transparency, with recommendations including protocol preregistration and sharing of anonymized paired datasets. The author acknowledges the framework has limitations, being a narrative appraisal rather than a systematic review and lacking formal expert consensus. Still, the argument echoes successful standardization efforts elsewhere in diagnostic research, such as the STARD and TRIPOD initiatives that transformed reporting quality. As CGM sensors proliferate globally and their analytical differences narrow, the biggest variable in how well a device appears to perform may increasingly be the study that tested it. Harmonizing how those studies are designed, analyzed, and reported, Ameur concludes, may prove as important to the next generation of diabetes technology as any improvement in sensor chemistry itself.

Subject of Research: Methodological standardization of continuous glucose monitoring accuracy validation

Article Title: Toward standardized continuous glucose monitoring validation: lessons from the Sinocare iCan i3 accuracy study

Article References: Ameur, A. (2026). Toward standardized continuous glucose monitoring validation: lessons from the Sinocare iCan i3 accuracy study. Frontiers in Clinical Diabetes and Healthcare, 7, Article 1941126. https://doi.org/10.3389/fcdhc.2026.1941126

Image Credits: AI Generated

DOI: 10.3389/fcdhc.2026.1941126

Keywords: continuous glucose monitoring, MARD, sensor accuracy, diabetes technology, validation methodology, reference standard, YSI analyzer, sensor survival analysis, hypoglycemia, statistical uncertainty, Sinocare iCan i3, clinical accuracy