JMIR Publications

ABSTRACT

Background:

The CGMacros continuous glucose monitoring (CGM) study reported Dexcom reading higher than FreeStyle Libre in every group and a device-by-metabolic-state interaction for time in range (TIR). The interaction's cause and consequences for correction and classification are unclear.

Objective:

To compare paired Dexcom and Libre recordings with each other and blood references, and assess whether corrections and classifiers preserve numerical, zone, and label endpoints.

Methods:

CGMacros provided simultaneous Dexcom G6 Pro and FreeStyle Libre Pro recordings from 45 adults in 3 glycated hemoglobin-defined stages (healthy, prediabetes, type 2 diabetes [T2D]). References were capillary glucose (CGMacros, ShanghaiT2DM) or laboratory glucose within ±5 minutes (AI-READI). Population and personal scaling, ridge regression, and a temporal convolutional network translated readings between sensors. Nested participant-grouped elastic-net classifiers used mean glucose or 22 CGM features, with sensor identity withheld or supplied. Post hoc analyses covered a leave-one-out Dexcom shift identical across stages, a cross-strategy contrast, label correctness, and CGM added to demographics.

Results:

Libre-minus-Dexcom mean glucose was -32.37 mg/dL (95% CI -37.07 to -27.50). Blood glucose lay between the devices: against CGMacros capillary glucose, Dexcom read +16.51 and Libre -14.33 mg/dL; ShanghaiT2DM (Libre H, -16.31) and AI-READI (Dexcom G6, +25.04) supported these directions. Libre recorded less TIR in healthy participants (-19.94 percentage points) and more in T2D (+14.08; stage interaction P<.001). The uniform shift reproduced -27.50 of the observed -34.02 percentage-point healthy-minus-T2D TIR contrast, and no residual stage interaction was detected (P=.72; residual -6.52, 95% CI -17.61 to 4.32). Toward Dexcom, personalized temporal ridge had the lowest mean absolute error (10.73 mg/dL), but personal scaling had higher balanced zone accuracy (77.9% vs 67.3%) with lower positive predictive value. In exploratory pooled evaluation, a manufacturer indicator increased mean-only macro-averaged area under the receiver operating characteristic curve (AUROC) from 0.739 to 0.783 (paired gain 0.044; 95% CI 0.006-0.086), leaving within-sensor AUROC unchanged. A sensor-blind 22-feature model did not clearly differ from the indicator model. Both reduced sensor-dependent labels from 30 of 45 to 13 and 12. Only 11 of 19 and 11 of 21 participants whose labels became sensor-independent were correct on both sensors. Adding 22 CGM features to age, body mass index, and gender increased AUROC from 0.658 to 0.788 (gain 0.130; 95% CI 0.019-0.239).

Conclusions:

In exploratory analyses, the best correction depended on the endpoint: lower numerical error did not guarantee zone detection toward Dexcom. Blood glucose lay between the devices, so the correction target is a choice. One additive between-device difference was compatible with the opposite-signed stage TIR differences without requiring a stage-specific device difference. More consistent stage labels were not necessarily correct. CGM added stage discrimination beyond demographics. Pooled CGM studies should name the target device and report numerical, zone, and label agreement separately.