Currently accepted at: JMIR Diabetes
Date Submitted: Oct 5, 2026
Open Peer Review Period: Oct 6, 2026 - Dec 1, 2026
Date Accepted: Nov 20, 2026
(currently open for review)
This paper has been accepted and is currently in production.
It will appear shortly on 10.2196/113658
The final accepted version (not copyedited yet) is in this tab.
Correcting Differences Between Dexcom G6 Pro and FreeStyle Libre Pro Glucose Readings: Retrospective Paired-Device Study
ABSTRACT
Background:
Continuous glucose monitors (CGMs) from different manufacturers can give different readings for the same person. Studies combining datasets from different sensors can correct readings toward one sensor (the target) to make them comparable. However, it remains unclear how much reducing glucose prediction error improves agreement in terms of clinically meaningful endpoints such as time in range (TIR) and glucose zone assignments. Sensor differences may also affect the consistency and accuracy of glycemic stage predictions.
Objective:
To (1) compare Dexcom and FreeStyle Libre with each other and with capillary and laboratory blood glucose references, (2) evaluate how well corrected readings match the target's glucose values, TIR, and glucose zone assignments, and (3) assess whether richer CGM features improve cross-sensor diabetes stage classification without device identity.
Methods:
In CGMacros, 45 adults, staged by glycated hemoglobin (15 healthy, 16 prediabetes, 14 type 2 diabetes [T2D]), simultaneously wore Dexcom G6 Pro and FreeStyle Libre Pro sensors. We corrected readings using scaling, ridge regression, or a temporal convolutional network, fitted on other participants (population) or each person's first days (personalized) and tested on later days. We compared elastic net stage classifiers using 22 CGM features versus mean glucose, with and without device identity. Models were trained on both sensors and tested on unseen participants.
Results:
Libre averaged 32.37 mg/dL lower than Dexcom (95% CI 27.50-37.07). In matched samples, Dexcom averaged above and Libre below finger stick glucose (+16.51 and −14.33 mg/dL). External samples agreed in direction: Dexcom G6 above laboratory glucose (AI-READI); FreeStyle Libre H below capillary glucose (ShanghaiT2DM). Libre showed 19.94 percentage points less TIR than Dexcom in healthy participants but 14.08 points more in T2D. On Libre, TIR was therefore higher in T2D than in healthy participants (82.33% vs 76.51%). Post hoc, the mean sensor difference, applied equally to all Dexcom readings, could explain most of this pattern. Correcting toward Dexcom, personalized temporal ridge gave the lowest mean absolute error (10.73 mg/dL) and cut daily TIR error from 15.13 to 3.46 points, whereas personalized scaling detected zones better (balanced accuracy 77.9% vs 67.3%), at lower positive predictive value. Without device identity, a mean glucose model staged 30 of 45 participants differently across sensors. Adding device identity reduced this to 13. Replacing mean glucose with 22 CGM features, without device identity, reduced it to 12 from 30. Agreement did not imply correct staging. Neither model clearly discriminated better. Post hoc, these features added information beyond age, body mass index, and gender (area under the curve gain 0.130; 95% CI 0.019-0.239, conditional on saved predictions)
Conclusions:
In exploratory analyses, these sensors were not interchangeable, and no correction was best for every endpoint. Because Dexcom averaged above and Libre below blood glucose in matched samples, correcting one toward the other improves comparability, not necessarily accuracy. Studies combining CGM data should name the target sensor and match the correction to the endpoint, including glucose values, glycemic zones, or glycemic stage.
Citation
Request queued. Please wait while the file is being generated. It may take some time.
Copyright
© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.