Accuracy Discrepancies: MARD Analysis by Sensor Generation

Accuracy Discrepancies: MARD Analysis by Sensor Generation

MARD, the most quoted accuracy number for continuous glucose monitors (CGMs), has improved a lot across sensor generations. The first sensors studied in children missed by a median of 12 to 19 percent [1]. FDA summaries for sensors sold as of October 2026 mostly report figures between about 7 and 10 percent [2][3][4][5]. The discrepancy in the title is the catch: the same sensor can post a very different MARD in another study, so a one-point gap between brands measured in separate trials tells you little [6].

What MARD measures

MARD stands for mean absolute relative difference. Researchers pair each sensor reading with a reference glucose value taken at the same moment and express the gap as a percentage of the reference. Then they drop the plus or minus sign and average every pair [7]. Lower is better. In the pivotal studies we read, the reference was a YSI 2300 laboratory analyzer run on venous blood [8][2].

It's easy to calculate, which is why it shows up everywhere. A 2020 review concluded that MARD should not be the only measure used to characterize a CGM [7]. A 2017 analysis by several of the same authors found that a study's MARD reflects the study's design as well as the sensor, so published values carry real uncertainty [6].

The FDA doesn't set a MARD cutoff either. Its special controls for integrated CGMs (iCGMs), sensors designed to work with connected devices such as automated insulin pumps, use percent-within targets. Across the measuring range, for example, the lower 95 percent confidence bound for the share of readings within 20 percent of the reference must exceed 87 percent [9]. Our article on iCGM regulatory standards covers the other criteria.

The first generation: errors near 20 percent

The oldest figures we found come from DirecNet, a pediatric research network. In an inpatient study, Medtronic's original CGMS sensor had a median relative absolute difference of 19 percent, against 11 percent for a modified version [1]. That's a median, not a mean, so it isn't strictly a MARD. When 200 children were asked to wear the modified sensor at home, the median error was 12 percent against a home meter, and it doubled in low glucose compared with high glucose (20 versus 10 percent) [1].

The 2010s: single digits arrive

Medtronic's Guardian Sensor 3, called a fourth-generation sensor in its pivotal paper, had a MARD of 9.4 percent on the abdomen with the Guardian Connect app and 8.7 percent on the arm in an 88-person study. Participants still calibrated it with fingersticks 3 to 4 times a day [10]. Guardian Connect has since been superseded: as of October 2026, MiniMed's US sensor comparison lists the Instinct, Simplera Sync, Guardian 4 and Simplera sensors instead [11].

Factory calibration, where the manufacturer calibrates each sensor so users don't have to, arrived in the same years. Abbott's first FreeStyle Libre, a factory-calibrated 14-day sensor, had an overall MARD of 11.4 percent in its US study [12]. Dexcom's G6, whose De Novo review created the FDA's iCGM classification [13], reached 9.0 percent in a reprocessed factory-calibration analysis of 62 participants [14]. Dexcom is now phasing out the G6 [15].

Abbott's FreeStyle Libre 2 was cleared as an iCGM in June 2020 [16], with a MARD of 9.2 percent in adults and 9.7 percent in children [17]. Abbott kept the original Libre 2 and Libre 3 sensors available only through September 30, 2025, then moved users to the Plus versions [18].

Sensors sold as of October 2026

Dexcom G7 and G7 15 Day

Dexcom's US site presents the G7 and G7 15 Day side by side [19]. In the adult G7 study of 316 participants, MARD was 8.2 percent on the arm and 9.1 percent on the abdomen [8]; the FDA cleared the G7 in December 2022 [20]. Here's what surprised us about the G7 15 Day, cleared for adults in April 2025: its FDA summary gives no single overall MARD. It reports one per wear period instead. The figure was 8.6 percent on days 1 to 3, 7.3 to 7.4 percent through mid-wear and 8.9 percent at the end, with 93.7 percent of readings within 20 percent of the reference [2].

FreeStyle Libre 2 Plus and Libre 3 Plus

FDA's 2024 clearance for the Plus sensors points back to a 2023 clearance for their clinical data [21]. That study tested a modified sensor tail built to reduce vitamin C interference and allow use with automated insulin delivery. In its 150 adults, 92.5 percent of readings fell within 20 percent of the reference, and MARD was 10.0 percent on days 1 to 3 before settling between 7.2 and 7.8 percent [3].

MiniMed sensors

MiniMed's standalone Simplera CGM, approved in July 2024, had an overall MARD of 10.2 percent in 116 adults; day 1 was the weakest, at 13.7 percent [4]. Simplera Sync, the version that connects to the MiniMed 780G pump, posted 10.8 percent in children 7 to 17 in the study behind its April 2025 approval [22]. The Instinct sensor, sold with the 780G, is made by Abbott and lasts up to 15 days [23]. We didn't find a public FDA accuracy table for Instinct, so we don't quote a number.

Eversense 365

Senseonics' Eversense 365 is implanted under the skin and rated for a year. In its 110-adult ENHANCE study, 92.0 percent of readings across the year were within 20 percent of a laboratory reference, and MARD in roughly monthly blocks ranged from 7.8 to 9.9 percent [5].

Why the same sensor scores differently

Head-to-head studies, where people wear several sensors at once, show how much the study shapes the result. At Germany's Institut für Diabetes-Technologie, 24 adults with type 1 diabetes wore a FreeStyle Libre 3, a Dexcom G7 and a Simplera together. Against a YSI analyzer the MARDs were 11.6, 12.0 and 11.6 percent; against fingerstick meter readings, 9.7, 10.1 and 16.6 percent. All three did worse than in earlier studies, and the authors called for common study design guidelines [24].

A 2024 study in 55 adults split them much further. The G7, at 8.2 to 9.1 percent in its pivotal study [8], posted 13.6 percent here, against 8.9 percent for the original Libre 3 [25], a sensor since replaced by the Libre 3 Plus [18].

These gaps matter outside the lab. When the same institute compared glucose metrics from the three sensors worn together, some participants' differences would have led to different therapy recommendations [26]. How to weigh any sensor's readings is a decision to make with your care team.

What one average leaves out

A single MARD blends every reading together and hides where the errors cluster.

• Low glucose is harder. For the G7 15 Day, MARD was 14.9 percent when the sensor read below 54 mg/dL and 6.6 percent above 250 mg/dL [2].
• The first day is harder. Simplera's day 1 and the Plus sensors' days 1 to 3 had the highest MARDs in their wear tables [4][3].
• Direction and speed don't show up. Continuous glucose error grid analysis was developed in 2004 because standard methods miss how well a sensor tracks the direction and rate of glucose change [27]. MARD also ignores how often a sensor reports [6].

The drop from median errors near 20 percent to single-digit averages is real. For any one sensor, the more telling numbers sit a layer down in its FDA summary: the low-glucose rows, the first-day rows, and which reference the study used.