How Accurate Is a CGM in People Without Diabetes?
The sensor is reading something real, but it is not reading the same thing as your blood.
How closely a consumer continuous glucose monitor tracks reference blood glucose in people without diabetes, and the conditions under which the gap widens. It does not cover accuracy in people with diabetes, where the devices were originally validated, and it does not cover whether any particular reading means anything about long term metabolic health.
The honest answer is that it depends on what you are asking the number to do, and the studies disagree about how large the gap is. In healthy adults the sensor tends to read above capillary blood, both fasting and after meals, and that offset gets amplified whenever it is counted against a fixed line. One crossover trial put the gap at about 0.9 mmol/L and found it overstated time spent above 7.8 mmol/L by roughly fourfold. Other work reports much tighter agreement, and what all of it agrees on is that the gap is not the same for every person and widens around exercise.
The sensor says 7.9 and the finger prick says 7.0
You bought a sensor to see what your own glucose does, and now you are watching numbers that feel high. A meal that should be unremarkable pushes you into a range the app colours differently. You check against a finger prick and the two do not match. So which one is right, and how far off is the one on your arm?
The gap matters more for people without diabetes than for people with it. These sensors were built and validated for managing a disease, where the job is catching large swings and a clinician reads the output. Reading an absolute number inside a healthy range, against a line drawn at a specific value, is a different job from the one the device was designed around. A harder one, arguably.
3 studies
- Across seven carbohydrate challenges, sensor estimates of fasting and postprandial glucose ran about 0.9 mmol/L higher than capillary measurements. Unadjusted, the sensor overestimated time above 7.8 mmol/L by roughly fourfold, falling to about twofold after correcting for baseline differences, and the gap differed between participants. The authors concluded that caution applies when inferring glycemic responses to foods from a sensor.
- In adults without diabetes around a strength training session, agreement with capillary blood was acceptable at group level with mean biases between 7.95 and 17.83 mg/dL, but only 50.5 to 64.3 percent of paired readings fell in zone A of the error grid and the discrepancy was largest after exercise. The authors reported that accuracy varied substantially between individuals and did not meet clinical criteria for diabetic monitoring.
- Measured alongside a new non-invasive device during an oral glucose tolerance test, two commercial needle-based sensors reached mean absolute relative differences of 10.42 and 12.93 percent against reference in healthy participants, a considerably tighter band than the nutrition studies report.
The disagreement between studies is the finding
Put the numbers side by side and they do not resolve into one answer. A flash sensor study in healthy adults reported a mean absolute relative difference of 25.4 percent against the laboratory method, with readings mostly lower than reference, and still concluded that the device tracked the shape of the post meal rise well enough for nutrition work. Its own title says the method can accurately reflect postprandial change. Both are true. They stop contradicting each other the moment you separate tracking a curve from reading a value off it.
That separation runs through all of this. A sensor can be reliable about the direction and timing of a rise while being unreliable about where the peak sits. The crossover trial shows it cleanly. A bias of about 0.9 mmol/L is modest on its own, but set against a fixed threshold it became a fourfold overstatement of time above range. A threshold turns a small offset into a large counting error, which is a different kind of mistake from simply being a bit off.
But exercise widens the gap, and that much the studies do agree on. One group found error rose from a baseline of 13.61 percent to 17.46 percent during aerobic activity, and that feeding the sensor an activity signal from a wearable pulled it back toward baseline. The strength training work found its largest discrepancy in the period after the session, which is the window where someone reading a sensor around training would be looking hardest. Altitude is a separate condition with its own evidence.
The studies here are small. The crossover trial ran 15 participants, the strength training study 9, the flash sensor study 10, and the validation study that produced the tighter 10 to 13 percent figures had 6. None of them tested an over the counter sensor sold for general wellness use, and none followed people across weeks of ordinary eating. What has not been measured is how far a sensor drifts over a full wear period in someone with no diabetes going about a normal life.
What the number can and cannot carry
Nothing here establishes that consumer sensors are inaccurate, and nothing here establishes that they are accurate enough to read absolutely. The evidence splits by what is being asked of it. Where studies looked at whether a sensor captures the shape and timing of a glucose response, they generally found it does. Comparisons of the sensor value against a reference value went the other way, with a gap large enough to matter and inconsistent enough between people that no single correction would fix it.
The practical consequence is about thresholds. Any app feature that counts minutes above a line, scores a meal, or ranks foods against each other inherits whatever offset the sensor carries for that person, and the crossover trial showed that offset is individual rather than fixed. What a sensor shows in people without diabetes is a question about pattern, that is the question the evidence supports better.
Every study here compared the sensor against capillary or laboratory blood sampled at intervals, which captures agreement at those moments rather than across a continuous trace. None of them followed a sensor through a full wear period. How the gap behaves on day ten of a patch, nobody has measured. Not in anything published here, anyway.
Common questions
How accurate is a CGM if you do not have diabetes?
The studies do not agree on a single figure. A randomised crossover trial in healthy adults found the sensor read about 0.9 mmol/L above capillary blood fasting and after meals. A validation study measuring two commercial sensors during a glucose tolerance test reported mean absolute relative differences of 10.42 and 12.93 percent. A flash sensor study reported 25.4 percent. The range across those is wide enough that no one number describes it.
Why does the sensor read higher than a finger prick?
The crossover trial found a consistent upward offset of about 0.9 mmol/L against capillary sampling, and found that the size of that offset differed between participants. Adjusting for each person baseline difference reduced the error but did not remove it.
Does the app time above range figure mean anything?
Treat it carefully. In the crossover trial, an unadjusted sensor overstated time above 7.8 mmol/L by roughly fourfold, and about twofold even after baseline correction. A small offset becomes a large counting error once it is applied against a fixed threshold.
Do the readings get worse around exercise?
That is the one condition the studies agree on. One group reported sensor error rising from 13.61 percent at baseline to 17.46 percent during aerobic activity. A study around strength training found its largest discrepancy in the post exercise phase and an increase in error grid treatment errors during and after the session.
Can a sensor still be useful if the absolute numbers are off?
The evidence is more supportive of pattern than of value. The flash sensor study found that the maximum rise from baseline captured by the sensor did not differ significantly from the laboratory method, even though the absolute readings did. Tracking the shape of a response and reading a number off the curve are different uses with different levels of support.
Sources
- Continuous glucose monitor overestimates glycemia, with the magnitude of bias varying by postprandial test and individual - a randomized crossover trial.
- Accuracy of a continuous glucose monitoring system applied before, during, and after an intense leg-squat session with low- and high-carbohydrate availability in young adults without diabetes.
- Clinical Validation on Healthy Humans of a Portable Non-Invasive Continuous Glucose Monitor Based on Transdermal Band-Pass Raman Spectroscopy.
- Flash Glucose Monitoring Can Accurately Reflect Postprandial Glucose Changes in Healthy Adults in Nutrition Studies.
- Enhanced Accuracy of Continuous Glucose Monitoring during Exercise through Physical Activity Tracking Integration.