Agreement of Over-the-Counter Continuous Glucose Monitors With Capillary-Meter Glucose in Young Adults Without Diabetes

Abstract

Background:

Over-the-counter continuous glucose monitors (OTC CGMs) are increasingly used by individuals without diabetes for lifestyle modification, weight management, athletics, and metabolic screening, and have potential as an affordable tool for glucose data collection in research. However, accuracy and agreement data in their intended populations are limited. No independent study has yet evaluated OTC CGM accuracy or agreement.

Method:

We evaluated OTC CGMs worn by participants (n = 39) without diabetes across sessions combining rest or exercise with placebo or sugar-sweetened drink consumption that mimic scenarios in which CGM values may be used to make health-related decisions. We assessed agreement with capillary-meter glucose, interparticipant differences in CGM performance, and delay of CGM recordings.

Results:

Overall mean absolute relative difference (MARD) was 12.9%, mean bias −4.8 mg/dL, and within-15%/15-mg/dL agreement 66.5%. Agreement was significantly worse under sugar than placebo consumption (MARD 15.3% vs 10.6%, P < .01). Exercise had no consistent effect. Disagreement increased with faster rates of change and, less consistently, higher glucose. A uniform 5-minute delay improved MARD to 11.2% (P < .001), and cross-correlation estimated a 6.2-minute population lag. Interparticipant agreement differed significantly, unattributable to measured biological variables. Participant/device identity explained direction of error more than magnitude. High-MARD participants did not have consistent high MARD across conditions. 13.8% of CGM-change-indicator directional arrows were classified as moderate to extreme risk.

Conclusions:

Over-the-counter continuous glucose monitors seemingly agreed with overall capillary glucose trends but were less reliable for exact-value agreement. They may aid lifestyle monitoring but warrant caution for diagnostic thresholds, real-time change-indicator use, and research.

We analyze data from a clinical trial: ClinicalTrials.gov, https://clinicaltrials.gov/study/NCT07255183?id=NCT07255183, NCT07255183.

Clinical Impact:

Over-the-counter continuous glucose monitors are reaching patients through primary care recommendations and direct consumer purchase, often before any structured education on interpretation. This first independent evaluation in adults without diabetes shows that OTC CGMs agree reasonably well with capillary trends but are less reliable for exact values, particularly after carbohydrate intake, at higher glucose levels, and during rapid change. Clinicians can reasonably guide patients toward using OTC CGMs to observe postprandial trends and behavioral patterns that fingerstick logs cannot practically capture, while cautioning against threshold-based decisions such as hypoglycemia detection or self-screening for diabetes, which warrant confirmatory testing.

Introduction

Historically, non-over-the-counter (nOTC) continuous glucose monitors (CGMs) have been used to improve glycemic control in individuals with diabetes. ¹ Glucose dynamics in individuals without diabetes may also contain important health information. Glucose nadirs before eating predict food intake in both healthy and obese adults. ² Hypoglycemia, if unaccounted for, around exercise can reduce performance.^3,4 Studies show inter- and intraparticipant variation in blood glucose response to the same exercise or meal.^{4 -7} Therefore, general glycemic-health guidelines may be insufficient, and many individuals without diabetes may need to self-monitor glucose to understand their unique glucose dynamics. CGMs conveniently provide comprehensive long-term glucose data.

Recently, certain CGM sensors, represented by Abbott’s Lingo and Dexcom’s Stelo, have been authorized for over-the-counter (OTC) use for individuals without diabetes.^8,9 As of June 2025, Dexcom reported over 15 000 Stelo users without diabetes and over 12 000 with prediabetes, while Amazon data suggests that the Lingo may be more popular, ranking it the number-one best seller above the Stelo in Blood Glucose Monitors as of June 2026.^10,11

Potential uses of CGMs for individuals without diabetes include preventing and identifying metabolic disease, informing diet/exercise, providing affordable data for research, managing weight, optimizing athletic performance, and monitoring for hypoglycemia in elite athletics.^{12 -14} Different applications place different demands on performance: Glycemic trend pattern recognition may tolerate more disagreement than decisions using diagnostic thresholds. Independent information about OTC CGM’s strengths and limitations in populations without diabetes is needed.

No identified study has evaluated OTC CGM accuracy or agreement. Abbott Lingo, for example, obtained Food and Drug Administration (FDA) clearance through substantial equivalence with the Abbott Libre 2, making clearance data and Libre 2 evaluations proximal performance markers, though they may not capture manufacturing differences. However, its FDA clearance data were recorded from people with diabetes. ⁸ Seven identified studies evaluated the Libre 2 in users without diabetes.^{15 -21} Only three studies, along with ours, have populations of sedentary-to-moderately-active healthy individuals, comparatively representative of free-living Lingo users, 63% of whom live with excess weight. ²² The Stelo has an analogous situation. Remaining gaps include exercise effects on agreement, average CGM delay, heteroscedasticity, and glucose-change indicator evaluation. These measures inform OTC CGM reliability for free-living users across applications. Related nOTC studies suggest that disagreement may increase at higher glucose levels and during rapid change in glucose, while reported sensor delays vary between studies and across conditions.^{23 -26}

This study analyzes capillary-meter and OTC CGM data from participants without diabetes, collected across 4 randomized-order laboratory sessions each. Sessions include participants either having a placebo (stevia) beverage and rest, a placebo beverage and exercise, a sugar-sweetened beverage and rest, or a sugar-sweetened beverage and exercise (hereinafter PR, PE, SR, and SE, respectively). These conditions capture a subset of consumption–exercise combinations under which OTC CGM users are likely to act on their data. The 75 g sugar beverage was comparable to the amount of glucose in an oral glucose tolerance test (OGTT), which screens for prediabetes and diabetes and is frequent in research, ²⁷ but was not used diagnostically. Exercise corresponded to moderate-intensity continuous exercise, typical of regimens pursued for weight loss or general health.^28,29 Beyond evaluating OTC CGM reliability and agreement with glucose meters, this study establishes proof-of-concept analytical practices and baseline results to inform OTC-specific evaluation and its differences from evaluation of nOTC CGMs.

We posed 3 research questions: First, how well do CGM values agree with capillary-meter values overall, and where do broader disagreements arise? We hypothesized greater and more variable disagreements at higher glucose levels and in exercise and sugar conditions. Second, can time-delay correction improve agreement, and does the optimal delay vary across participants or conditions? We hypothesized that shifting OTC CGM values by a small delay would significantly improve agreement. Third, are there interparticipant differences in OTC CGM agreement, and can they be explained? We hypothesized high variability in CGM–meter agreement unattributable to participant physiological differences.

Methods

Study Design and Participants

This study was a prospective unblinded ancillary analysis of the clinical trial NCT07255183, ³⁰ a double-blind, placebo-controlled, within-participant crossover in which each participant completed all four conditions in randomized order. It was approved by the University of South Florida Institutional Review Board (STUDY009043). Participants provided written informed consent, including consent for publication. Analyses were preregistered on the Open Science Framework (10.17605/OSF.IO/4EQJZ). ³¹ The placebo drink contained erythritol, a non-glycemic-acting sweetener that does not raise blood glucose, making it an effective placebo for this article’s target. ³² Full details of the design, registration status of individual analyses, CGM export and display comparison, and methods are provided in the Supplement.

Test Methods and Data Collection

Glucose was recorded every 5 minutes by the Lingo Glucose Biosensor (Abbott Diabetes Care Inc., Alameda, CA) and exported at the end of each participant’s 2-to-3-week participation. Staff also manually recorded CGM values every 5 minutes of the session. These 2 traces were only used for analyses requiring delayed values or rate of change. CGM values outside the device range (>200 or <55 mg/dL) were retained at the nearer bound. Capillary glucose was measured by fingerstick with the Contour Next (Ascensia Diabetes Care US Inc., Mishawaka, Indiana), the top-ranked meter in the Diabetes Technology Society Blood Glucose Monitor System Surveillance Program. ³³ All metrics describe CGM–capillary agreement since the meter itself has error. Staff separately recorded CGM values paired with each fingerstick reading at the 5-, 30-, 50-, 65-, 100-, and 130-minute marks of every session, forming the primary dataset.

Statistical Analyses

Overview

Analyses used Python (version 3.12; Python Software Foundation). All tests were two-sided with significance at 0.05. Bootstrap confidence intervals were in percentage points unless otherwise specified. Primary inference was at the participant level. Participant ID was considered a function of both the participant and device, as most participants only used one CGM. Summary metrics (mean absolute relative difference [MARD], median absolute relative difference [ARD], within-15%/15-mg/dL and within-40%/40-mg/dL agreement rates, and mean bias) were pooled unless otherwise specified. Signed differences were calculated as CGM minus capillary glucose. Mean glucose was defined as the mean of CGM and capillary values. Pearson r was reported as a secondary descriptor of co-movement. For each mixed-effects model (with participant ID as a random intercept), we checked residual normality (Shapiro-Wilk) and homoscedasticity (Breusch-Pagan); where these were violated, we refit the fixed effects by ordinary least squares (OLS) with conventional and participant-clustered standard errors and reported any coefficient whose significance changed between them and the mixed models.

Research Question 1

We computed agreement metrics overall and within stratifications (condition, consumption, exercise, and glucose range). We used the Consensus and Diabetes Technology Society (DTS) error grids, from which recordings are reported within risk ranges to convey the magnitude of disagreement. As an exploratory best-case, we also matched each fingerstick to the CGM value closest in magnitude within ±5 minutes to illustrate upper-bound CGM–capillary trend agreement. Participant-level MARD was compared between drink and activity categories by paired Wilcoxon signed-rank tests (Holm-adjusted) with participant-resampled bootstrap intervals, and linear mixed-effects models (LMMs) tested their independent and interactive effects on log-ARD. Mean glucose and rate of change were also related to absolute differences through Spearman correlation and log-ARD through an LMM. Heteroscedasticity was assessed with a Bland-Altman plot, locally estimated scatterplot smoothing (LOESS) trend, and Breusch-Pagan test. Mean glucose above 160 mg/dL was too sparse to support the plotted trend and was excluded from it and tested as a sensitivity. CGM-to-capillary standard deviation (SD), range, and coefficient of variation (CV) ratios were computed per participant and tested against 1 (one-sample Wilcoxon signed-rank, Holm-adjusted).

Research Question 2

We calculated agreement metrics after delaying CGM values by −5, 0, 5, 10, and 15 minutes, comparing each with the unshifted baseline overall and within each condition, both with participant-level paired Wilcoxon signed-rank tests (Holm-adjusted). We separately estimated the CGM–capillary time lag by cross-correlation. Each session depended on only 6 capillary readings, so the population lag, reported with a confidence interval in minutes, was a noisy average of unstable per-session estimates.

Research Question 3

We calculated per-participant agreement metrics and their spread across participants. We tested for differences in the distribution of ARDs across participants (Kruskal-Wallis). Each participant’s MARD was correlated with collected demographic variables (Spearman, Holm-adjusted). To quantify how much error was attributable to participant/device identity, we used mixed-effects variance partitioning, deriving intraclass correlation coefficients (ICCs) for relative-error magnitude (log-transformed ARD) and signed error. Each mixed model was compared against its OLS parallel by Akaike information criterion (AIC) and an approximate likelihood-ratio test. To assess whether high-MARD participants were consistently high-MARD across conditions, we fit a mixed-effects model of condition-level MARD and compared the between-participant SD of mean condition-level MARD with the average within-participant SD across conditions. We tested whether demographics predicted per-participant median CGM lag (Spearman, Holm-adjusted) and reported the ICC of session-level lag.

Results

Overview and Participants

Forty participants completed the study. One was excluded since their glucose exports could not be recovered, leaving 39 participants. Of the planned 936 capillary observations, 918 (98.1%) were present; of these, 905 (98.6%) paired with exact, 9 (1.0%) with display-censored (8 >200, 1 <55 mg/dL), and 4 (0.4%) with missing staff-recorded CGM values, yielding 914 (99.6%) capillary–CGM pairs. Of the planned 4056 CGM values in each session’s 5-minute trace, 3976 (98.0%) were exact, 27 (0.7%) display-censored, and 53 (1.3%) missing, forming an overall set of 4003 (98.7%). Of the planned 118 778 recordings in the full lifetime exports, 116 445 (98.0%) were exact, 1025 (0.9%) display-censored, and 1308 (1.1%) missing, forming an overall set of 117 470 (98.9%).

Participants were aged 18 to 28 years (19 female, 20 male). They were normal weight to overweight by body mass index (BMI); some had normal-weight obesity, and they were not particularly athletic by maximal oxygen uptake (VO₂ max).^34,35 Detailed demographic statistics are shown in Table 1.

Table 1.

Demographic Variables Collected for Each Participant.

Characteristic	Overall sample
Age range, n (%)
Underclassman range (18-20 years)	20 (51.3%)
Upperclassman range (21-23 years)	18 (46.2%)
Graduate student range (24-28 years)	1 (2.6%)
Sex, n (%)
Female	19 (48.7%)
Male	20 (51.3%)
BMI, kg/m2	22.82 ± 2.53
BMI category, n (%)
Normal weight, 18.5 to <25 kg/m2	32 (82.1%)
Overweight, 25 to <30 kg/m2	7 (17.9%)
Percent body fat, %	22.90 [16.55-27.35]
Body fat category, n (%)
Athlete range	8 (20.5%)
Fitness range	6 (15.4%)
Average range	18 (46.2%)
High/obese range	7 (17.9%)
Normal-weight BMI with high/obese  body fat, n (%)	4 (10.3%)
VO2 max, mL/kg/min	30.18 ± 6.60
VO2 max, % predicted	69.16 ± 14.27
VO2 max, relative to age/sex-predicted value, n (%)
<80% predicted	28 (71.8%)
80% to <100% predicted	11 (28.2%)
Max workload, W	188.92 ± 51.08
Ventilatory threshold workload, W	108.69 ± 38.30
Fasting glucose, mg/dL	91.87 ± 7.60
Fasting glucose category, n (%)
Normal fasting glucose, <100 mg/dL	34 (87.2%)
Prediabetes range, 100 to 125 mg/dL	5 (12.8%)
Height, in	67.32 ± 4.30
Weight, lbs	148.09 ± 28.96

Normally distributed variables reported with Mean ± SD and others reported as Median (IQR/interquartile range). Predicted VO₂ was calculated through the methods described by van der Steeg and Takken. ³⁵

Residual normality was rejected for every mixed-effects model (all P < .001). Residual heteroscedasticity was present for most models (listed comprehensively in the Supplement). One coefficient changed significance between mixed and OLS models, reported below.

Research Question 1

Overall Agreement

The overall MARD was 12.9%, median ARD 10.6%, mean bias −4.8 mg/dL, and within-15%/15-mg/dL and within-40%/40-mg/dL agreement rates 66.5% and 98.0%. Pearson correlation was 0.80. On the Consensus and DTS error grids, 99.8% and 99.7% of points respectively fell under zones A+B. The DTS error grid suggested CGM-displayed directional arrows had higher disagreement in estimating glucose directionality, with 13.8% of values having moderate to extreme risk (Figure 1).

Figure 1.

DTS and Consensus error grid results for our dataset. The DTS error grid also displays a grid for CGM directional glucose arrow (indicator) agreement. CGM-displayed trends had 2 extreme inaccuracies.

The best-case selection had a lower MARD (8.3%), higher Pearson correlation (0.90), and high within-15%/15-mg/dL agreement (83.9%). The best-case selection method is visualized for a specific session in Figure 2.

Figure 2.

Best-case selection method plotted. This method is an optimistic demonstration of CGM–meter trend agreement, as it illustrates how much error is reduced after matching a CGM value in a bounded time frame around each capillary-meter value. This method is not a verified measure of agreement and instead an exploratory, custom method.

Agreement by Condition

Differences in agreement metrics were larger between sugar and placebo drink than between exercise and rest. Mean bias was negative across all conditions. Metrics are plotted in Figure 3.

Figure 3.

Pooled metrics plotted for each condition. The data suggested lower agreement and more variability in absolute differences in sugar conditions (SR, SE).

Mean participant-level MARD was 15.3% under sugar versus 10.6% under placebo, a significant difference (paired Wilcoxon P < .01; bootstrap 95% CI [3.15, 6.20]). In contrast, participant-level MARD was 12.8% at rest versus 13.1% with exercise, not significant (paired Wilcoxon P = .83; 95% CI [−1.25, 1.83]).

In the LMM using all 4 conditions, the PE-PR difference was significant (P < .001) with a negative coefficient (−0.39) as well as the SE-PR difference (P < .05) with a positive coefficient (0.23). SR did not differ significantly (P = .26). In the LMM with drink and exercise categories, only drink was significant (P < .001) with a positive coefficient (0.37). In the LMM with drink, exercise, and their interaction, the exercise and interaction terms were significant (exercise P < .001, coefficient = −0.39; interaction P < .001, coefficient = 0.50; sugar P = .26).

Agreement by Glucose Level and Rate of Change

Splitting recordings into quintiles showed growing MARD at higher glucose, especially past ~100 mg/dL (Table 2).

Table 2.

Metrics From Glucose Data Split by Quintiles From Fingerstick Glucose Values.

Quintile	N	MARD
[43, 84]	190	10.9%
(84, 91]	186	11.2%
(91, 99]	187	11.5%
(99, 125]	171	15.4%
(125, 253]	180	16.0%

Agreement greatly reduced after the 3rd quintile (at values above 99 mg/dL). This cutoff approximates the accepted upper boundary of normal fasting glucose.

Spearman correlations of absolute residuals with mean glucose and rate of change were 0.30 and 0.31, respectively (P < .01). In the log-scale LMM with both predictors, rate of change was significant (coefficient = 0.20, P < .01) but mean glucose was not (P = .45). In the condition-adjusted model, the standardized rate of change and mean glucose were significant (P < .001, coefficient = 0.22; P < .05, coefficient = −0.10). Mean glucose was marginally insignificant when refit by OLS and clustered on participant (P = .06).

Bland-Altman analysis revealed heteroscedasticity, with smaller, more uniform differences before ~100 mg/dL, confirmed by a Breusch-Pagan test (Lagrange multiplier = 82.13, P < .001); it remained significant when values above 160 mg/dL were excluded (Lagrange multiplier = 80.87, P < .001). The LOESS trend of absolute residuals rose with glucose level (Figure 4).

Figure 4.

Bland-Altman plot with raw residuals (top) and scatter plot of absolute residuals (bottom) from all glucose values less than or equal to 160 mg/dL. Each point is colored by the rate of change of glucose calculated from CGM values. LOESS trends are plotted.

Variation Metrics

The CGM-to-capillary ratios of SD, range, and CV were 1.18, 1.17, and 1.23, respectively. All were significant above 1 (Wilcoxon P < .01).

Research Question 2

A 5-minute delay significantly improved MARD (11.2% vs 12.9%; Holm-adjusted Wilcoxon P < .001); no other delay was significant and better than baseline. By condition, the 5-minute delay significantly improved agreement in SE and SR; although some other delays differed significantly from baseline, none beat the 5-minute delay (Figure 5). In addition, the cross-correlation population lag was 6.2 minutes (peak r = 0.57; 95% CI [4.1-10.1]).

Figure 5.

Time delay analysis both on the entire dataset and on each condition. Statistical significance from the baseline (two-sided Wilcoxon signed-rank test, Holm-adjusted) is indicated by an asterisk over points. Overall, MARD (top-left), MARD by condition (top-right), overall within-15%/15-mg/dL agreement (bottom-left), and within-15%/15-mg/dL agreement by condition (bottom-right) are plotted. The only statistically significant improvement in MARDs when splitting by condition and on the entire dataset was the 5-minute delay. Displayed metrics are pooled.

Research Question 3

Across participants, the range (CV) of MARD and within-15%/15-mg/dL agreement were 15.1% (24.2%) and 68.1% (22.4%), respectively. A Kruskal-Wallis test (H = 88.54, P < .01) indicated participants did not share the same ARD distribution. The CGM underestimated for 30 participants and overestimated for 9 (Figure 6).

Figure 6.

Per-participant data sorted under each subplot. MARD (top-left), mean bias (top-right), within-15%/15-mg/dL agreement (bottom-left), and standard deviation ratios (bottom-right) are plotted. Columns are colored by gender. The “mean” lines are participant-level means instead of pooled values.

No demographic correlated significantly with participant MARD (Holm-adjusted P > .10). In variance partitioning, participant/device identity accounted for only 4.9% of residual relative-error magnitude but improved model fit (ΔAIC = 12.9, approximate likelihood-ratio P < .01). It explained 19.1% of residual signed-error variance. Participant-by-condition analysis did not support consistent MARD within participants across conditions: Participant ID accounted for 14.0% of condition-level MARD variance and did not meaningfully improve fit (ΔAIC = 1.7, approximate likelihood-ratio P = .06). The within-participant SD of condition MARD (5.2%) exceeded the between-participant SD (3.1%).

Per-session lag had a median of 6.2, mean of 9.0, and SD of 8.7 minutes. Per-participant medians ranged from 0 to 17.6 minutes. The ICC of session-level lag was low (<.05), indicating most variation of lag was intraparticipant session-to-session. No demographic predicted the per-participant lag.

Discussion

Principal Findings

Over-the-counter continuous glucose monitor agreement with the capillary meter was moderate for exact glucose values. The overall MARD (12.9%) and PR MARD (11.5%) both were relatively high; however, since the comparator itself has error, it is not directly attributable to OTC CGMs. Disagreement grew with the rate of glucose change and, less consistently, with glucose level, concentrating past ~100 mg/dL. The heteroscedastic error pattern matters since it concentrates disagreement in the elevated ranges that threshold-based decisions rely on. MARD was significantly higher during sugar consumption but did not change consistently during moderate-intensity continuous exercise. The device seemingly agreed more with capillary trends. The best-case illustration MARD was 8.3%, demonstrating CGM values in a bounded timeframe around capillary readings had more agreement with capillary values than concurrent CGM values, but is likely an optimistic measure. Correlation was 0.80. However, CGMs added noise to glucose dynamics (all variation ratios > 1). A uniform 5-minute delay was optimal but still relatively high (11.2%). The estimated lag (6.2 minutes) fell within the range reported for nOTC sensors. Interparticipant agreement varied widely. Participant/device identity contributed more to direction than magnitude of error. Error was unexplained by measured demographics, leaving open whether unmeasured physiology, device-quality variation, or another factor was responsible. Within-participant MARD was not consistent across conditions. CGM-lag varied greatly within participants, tempering the value of always treating CGM glucose as delayed by 5 minutes.

Lower agreement may matter less because incorrect glucose measurements usually pose less threat to people without diabetes. However, such users often lack clinical guidance in interpreting CGM data, and many potential OTC CGM uses compare glucose against thresholds, raising the accuracy burden.

These findings suggest OTC CGMs are currently best suited for tasks requiring overall glucose trends and larger-scale directionality; specifically, trends examined from glucose values, not real-time directional arrows, which warrant caution. Such devices may therefore help people making lifestyle changes around exercise and diet who look for patterns rather than solely one-time events or real-time change indicators. We cannot substantially comment on high-intensity exercise, but the link between rapid glucose changes and lower agreement suggests performance may be worse. We also cannot substantially comment on threshold-based tasks like OGTTs, but the higher disagreement of postconsumption values suggests caution. Use in research requires further study, given the multiple contributors to CGM–meter disagreement, but interparticipant agreement differences may be concerning.

Limitations and Future Studies

Capillary glucose was measured with a fingerstick meter, which itself carries error. The few paired readings per session limited the precision of agreement metrics and finer-grained trend capture. Generalizability was limited by the relatively small single-site convenience sample, all young (18-28) and relatively healthy, with few (5) in a prediabetic fasting glucose range. Findings do not extend to older, heavier, dysglycemic, or clinically at-risk populations. Because participants removed sensors at the end of participation, we could not investigate end-of-lifetime agreement decline. Conditions were limited, with sugar conditions not representing typical daily intake and exercise conditions not representing intense exercise. Finally, most participants wore a single CGM, limiting isolation of device-quality variation from participant effects. Future studies could add a laboratory or YSI reference; recruit larger, more diverse samples; collect more comparator readings across the sensor lifetime; test additional foods and exercise intensities; and use multiple sensors per participant to isolate the source of interparticipant differences.

Conclusion

To our knowledge, the present study is the first independent evaluation of OTC CGM use in various conditions of exercise, rest, and glucose consumption in a sedentary-to-moderately-active population. It identified factors of OTC CGM use important for end users (both wearers and researchers) and sets the foundation for future directions in OTC CGM research.