What a continuous glucose monitor tells you when you do not have diabetes

The case for wearing a continuous glucose monitor without diabetes is intuitive. You eat something, you watch the line move, you learn which foods spike you. The device turns an invisible process into feedback, and feedback is how people change behaviour.

That story contains three separate claims, and they have to be checked separately: that the device measures what you think it measures, that its numbers track something about your health, and that seeing them changes what you do. The literature treats these very differently, and the first is where the trouble starts.

The device and the finger-prick disagree

A 2025 randomised crossover trial put a FreeStyle Libre 2 against capillary finger-prick sampling in 15 healthy adults — nine women, six men, mean age 34, mean BMI 24.1 [1]. Each completed seven test meals, both methods sampled every 15 minutes for two hours, with capillary as the criterion measure.

Fasting glucose read 0.9 ± 0.6 mmol/L higher on the CGM and postprandial glucose 0.9 ± 0.5 mmol/L higher, both at p < 0.001. Those offsets compound into the summary metrics people actually look at: time above the 7.8 mmol/L threshold was overestimated by more than 3.8-fold, and correcting for the fasting bias still left roughly a doubling.

Enough to change the answer

The same smoothie was high-GI by sensor and medium-GI by finger-prick. The rank order of incremental area under the curve changed for five of the seven products — sort those foods best to worst by sensor and you get a different order than the reference method gives. The authors are unambiguous:

This is one device model in 15 people, with a single reference test per food where GI protocols call for duplicates. Not the last word. But the size and direction of the bias are hard to wave away, and the trial was run in exactly the population these devices are sold to for wellness.

What the metrics track

Suppose the measurement were clean. Do the numbers mean anything? The largest relevant dataset is a 2026 analysis of 3,634 people without diabetes or prediabetes from the PREDICT cohorts, wearing CGMs for two to four free-living days [2]. Median coefficient of variation ran 14.5–15.3% across the three cohorts.

The dietary associations exist and are small. Glycaemic variability correlated with carbohydrate as a share of energy at rs 0.16. A multivariate model of diet, lifestyle and demographics explained an adjusted 6.3% of the variance in variability (95% CI 0.05 to 0.15), and age was its only significant individual predictor, at a partial R² of 0.02.

Time in the stringent 3.9–5.6 mmol/L range did discriminate predicted cardiovascular risk, at AUC 0.75 (0.59 to 0.92). Variability did not, and no CGM metric significantly discriminated HOMA-IR or liver fat probability. Variability — the metric behind every instruction to flatten your curve — was not associated with HbA1c, mean glucose, or OGTT glucose at all.

Does seeing the number change anything?

A 2024 systematic review pooled 25 randomised trials of CGM as a behaviour-change tool across 2,996 participants [3]. HbA1c fell by 0.28 percentage points (95% CI −0.42 to −0.15), with heterogeneity at I² = 88.5%. Weight fell 0.7 kg (−1.4 to 0.0, p = 0.066) and BMI 0.4 kg/m² (−0.9 to 0.0, p = 0.080) — neither conventionally significant.

Three of the 25 trials enrolled people without diabetes; seventeen were in type 2 diabetes. The authors pre-specified a subgroup analysis by diabetes status and could not report it, for want of enough studies per subgroup. Activity and diet were measured too inconsistently to pool — five studies reported activity using nine different variables, none appearing twice. No GRADE ratings were given, and 11 of 23 trials were rated low risk of bias.

What the evidence does not establish

That CGM metrics predict anything in a healthy person over time. The PREDICT analysis is cross-sectional, so no association in it carries a direction of causation, and the authors name the open question themselves: whether CGM metrics predict progression from normal glucose tolerance to prediabetes or type 2 diabetes. Nobody here was followed to a diagnosis or an event.

Nor that the pooled HbA1c benefit reaches people with normal glucose. That estimate comes overwhelmingly from diabetes trials, where HbA1c is both elevated and the treatment target. In someone normoglycaemic the outcome is already near its floor. And the PREDICT sample was 83% female, its largest cohort 93% white, on two to four days of wear, with some participants able to see their own output as it was recorded.

Why this one is harder than it looks for n-of-1

We have argued before that a biased instrument survives a personal trial, because a constant offset sits on both sides of a within-person comparison and cancels. That needs the offset to be constant. Here it is not. Fasting bias differed significantly between participants, with a between-participant standard deviation of 0.6 ± 0.1 mmol/L — above the 0.5 the authors set as clinically meaningful — and was unexplained by fasting glucose, BMI, age or glucose tolerance, so there is nothing to correct it against.

Worse for the food-ranking case, the magnitude of bias varied by test meal. An error that changes size with what you ate does not cancel when you compare one meal against another — it is the comparison. That is why the rank order moved for five of seven products, and it is precisely what someone is doing when they test two breakfasts against each other on a sensor.

The honest reading is narrow. Within one person on one sensor, a large and repeated change in glucose is probably real. A difference between two similar foods, read off the absolute numbers, sits inside the instrument's error. Nothing here is medical advice, and if a clinician put a CGM on you, this post is not about you.