Why MARD Is the Wrong Headline Metric for CGM Performance

Mean Absolute Relative Difference (MARD) has become the default shorthand for evaluating continuous glucose monitor (CGM) performance—not because it is clinically meaningful, but because it is convenient. It compresses vast amounts of heterogeneous data into a single, easy-to-compare number that looks authoritative in a table or press release. Unfortunately, that simplicity comes at a cost. By averaging error across the full glucose range, MARD systematically obscures performance where accuracy matters most: hypoglycemia, rapid glucose transitions, and decision-critical moments. A CGM can post an impressive overall MARD while still underperforming at low glucose levels, during exercise, or in the post-prandial window—exactly the scenarios where patients, caregivers, and automated insulin delivery systems are most exposed to risk. In this way, MARD rewards statistical smoothing rather than physiological relevance.

The deeper problem is not that MARD exists, but that it is routinely overinterpreted. Differences of a few tenths of a percent—often well within the noise of study design, reference method bias, or glucose distribution—are frequently framed as meaningful performance advances, despite having no demonstrated impact on outcomes such as hypoglycemia avoidance, dosing safety, or time-in-range. Because MARD is highly sensitive to protocol choices, exclusion criteria, and cohort composition, it can be “optimized” without materially improving the user experience. When elevated to a headline metric, MARD distracts from analyses that actually inform clinical and algorithmic safety: error distributions, Bland–Altman plots, low-glucose bias, and rate-of-change accuracy. Precision without context is not insight, and in the case of CGM performance, MARD too often provides the former while masking the latter.

The stakes are even higher as iCGMs become the primary sensing layer for automated insulin delivery. Algorithms do not experience averages; they react to individual data points, trends, and inflection errors in real time. A sensor that looks “excellent” by MARD can still introduce systematic low-glucose bias, lag during rapid change, or asymmetric error that meaningfully alters insulin dosing decisions. As iCGM data increasingly drives therapy rather than merely informing it, performance reporting must evolve beyond a single, comforting number toward metrics that reflect real-world risk. In an automated future, MARD is not just insufficient—it is potentially misleading.