CGM for Non-Diabetics: What the Data Can and Cannot Tell You
A continuous glucose monitor is a reasonable instrument for a healthy person to wear, and in the United States you can now buy one without a prescription. The summary of the evidence is that CGM in non-diabetic people is a measurement tool with a well-characterized signal and poorly characterized clinical value. A recent systematic review and meta-analysis of CGM in non-diabetic populations found that most studies are observational, small, and heterogeneous in their endpoints.1 So wear one if you want a time-resolved record of your own glucose dynamics. Do not expect it to tell you whether you are healthy.
This page covers what the sensors measure and how accurate they are. It also covers which device we would pick, what normal looks like in large non-diabetic cohorts, and how to get the raw data into a form you can analyze.
What the sensor measures
Every consumer CGM is a subcutaneous electrochemical sensor. A filament sits in the interstitial fluid of the upper arm or abdomen. Glucose diffuses to an immobilized glucose oxidase layer and the resulting reaction generates a current. The transmitter converts that current to a glucose estimate using a factory calibration.2 The reading is interstitial. It is not venous or capillary.
That distinction produces the two failure modes you will notice first. Interstitial glucose lags blood glucose by roughly 5 to 15 minutes, so a fingerstick taken during a fast rise will read higher than the sensor.3 Manufacturers apply smoothing and lag-compensation filters, which help at steady state and can overshoot during sharp transitions.2
The second failure mode is mechanical. Sleeping on the sensor compresses the tissue around the filament and starves it of fluid, producing a sudden dip that looks like nocturnal hypoglycemia. These “compression lows” are usually 30 to 60 minutes long, V-shaped, and recover as fast as they fall. If you see a 55 mg/dL reading at 3 a.m. with no symptoms and a fast recovery, suspect the mattress before your pancreas.
Accuracy, in numbers you should hold in mind
Consumer CGMs report accuracy as MARD, the mean absolute relative difference against a laboratory reference. Current-generation sensors land in the high single digits to low teens as a percentage. That is good for tracking trends and mediocre for absolute values.
Interpret that practically. At a true glucose of 100 mg/dL, a 10 percent MARD means typical error around 10 mg/dL, with the tails wider than that. A sensor reading 112 and a sensor reading 98 after the same meal, on the same person, are not distinguishable measurements. Accuracy also degrades in the first 12 to 24 hours after insertion, which is the period most people are hovering over the app. Reviews of the technology consistently flag this early-wear instability along with day-to-day sensor-to-sensor variation.4
The implication for study design on yourself: never compare a single meal to a single meal. Compare distributions over many repeats.
Which device we would buy
For someone who wants to analyze their own data, we would use a Dexcom sensor, either Stelo over the counter or G7 if a clinician is already involved. The reason is data access rather than hardware. Dexcom publishes a developer API with OAuth 2.0 and returns 5-minute estimated glucose values as JSON, which means you can pull a full wear period programmatically instead of scraping a PDF.
Abbott’s Libre line, including the over-the-counter Lingo and Rio products, has excellent hardware and a smaller sensor. Export is weaker. You generally get a CSV out of LibreView, which is workable, and the reading cadence on some Libre products is coarser than Dexcom’s fixed 5-minute grid. If you want a clean, uniform time series with no reconstruction work, Dexcom is the easier path.
Avoid non-invasive optical devices that claim cuffless, needle-free glucose. A systematic review of non-invasive CGM for cardiovascular prevention in people without diabetes found the evidence base thin and the validation inconsistent.5 No consumer optical device has demonstrated accuracy comparable to a filament sensor.
What normal looks like
This is where most consumer content fails. People see a post-meal reading of 150 mg/dL and conclude something is wrong. Large cohort data says otherwise.
CGMap characterized CGM data from thousands of non-diabetic adults and published percentile reference ranges for standard CGM metrics.6 The headline finding is the width of the distribution. Healthy people spend the large majority of time in a narrow band, and transient excursions above 140 mg/dL are common and unremarkable in individuals with normal HbA1c. The study also showed that CGM-derived metrics separate people who have normal HbA1c but differ in glycemic variability.6
So build your interpretation on percentiles. Do not build it on thresholds you read in a blog post. The metrics worth computing from your own trace are these:
- Mean glucose and standard deviation over the full wear period.
- Coefficient of variation, which is SD divided by mean. This is the standard variability metric in CGM analysis.7
- Time in range, as the fraction of readings between 70 and 140 mg/dL for a non-diabetic context.
- Nocturnal mean, computed over a fixed window such as 00:00 to 05:00, which is much less noisy than daytime values because you are not eating.
- Postprandial incremental AUC over 2 hours from a logged meal start, baselined to the pre-meal value.
Nocturnal mean and coefficient of variation are the two we trust most. Both average over many readings, which suppresses sensor noise, and neither depends on you logging meals accurately.
Getting the raw data out
Assume you will do the analysis yourself. The workflow below is what we would run.
Pull the series first. For Dexcom, register an application in the developer portal and complete the OAuth flow. Then request /v3/users/self/egvs with startDate and endDate in ISO 8601 local time. The API caps each request window, so page through in chunks of a few days and concatenate. Store the raw JSON before you touch it.
Then normalize. Parse to a pandas DataFrame with a timezone-aware DatetimeIndex, then reindex onto a strict 5-minute grid with df.resample('5min').mean(). Gaps appear from sensor warmup, transmitter dropout, and phone Bluetooth loss. Do not interpolate across gaps longer than 20 minutes, and record the percentage of missing readings per day. Any day below roughly 70 percent coverage should be dropped from summary statistics, because partial days bias means toward whichever part of the day was captured.
Next, filter artifacts. Flag any excursion below 60 mg/dL that lasts under 30 minutes and returns to baseline at a similar slope. Review those by hand against your sleep position before including them. Also drop the first 12 hours of each sensor, which is the period where calibration drift is largest.4
Finally, log meals with timestamps rather than descriptions. A plain CSV with start_time, label, and estimated carbohydrate grams is enough to join against the glucose series. Compute incremental AUC with the trapezoidal rule using numpy.trapz on the 2-hour window after each meal start.
What this will and will not tell you
CGM will show you your own postprandial curves, your overnight baseline, and how sleep and exercise shift both. Those are real measurements of real physiology. Interventional studies in non-diabetic people have used CGM feedback to guide lifestyle changes, and reviews of that literature report short-term behavioral effects with limited evidence for durable cardiometabolic benefit.8
What it will not do is establish a diagnosis. Glucose is one axis of metabolic physiology, and a sensor with roughly 10 percent error is not a screening instrument. If your fasting or nocturnal values are persistently elevated, or you see repeated genuine lows, that belongs in front of a physician with laboratory glucose and HbA1c rather than in a spreadsheet. A CGM is not a substitute for a diagnostic workup.
We would treat a two-week wear as an experiment with a defined question. Pick one variable, repeat each condition at least four times, and compare distributions rather than single curves.
Questions people also ask
How do you get a CGM without diabetes?
In the United States, over-the-counter sensors are sold directly to consumers, including Dexcom Stelo and Abbott’s Lingo and Rio. You buy them online or through a pharmacy with no prescription. Prescription models such as the Dexcom G7 and FreeStyle Libre 3 Plus still require a clinician.
What is the best glucose monitor for a non-diabetic?
For anyone planning to analyze the data themselves, we would choose a Dexcom sensor because of the developer API and the fixed 5-minute reading grid. Abbott’s sensors are smaller and comparably accurate, but export is limited to CSV from LibreView.
Are CGM readings accurate enough to act on?
They are accurate enough to compare distributions across days and conditions. They are not accurate enough to treat a single reading as a precise number, given MARD in the high single digits or low teens plus the first-day warmup drift.4
Is a post-meal spike to 150 mg/dL a problem?
Transient excursions above 140 mg/dL occur regularly in non-diabetic people with normal HbA1c, based on percentile reference data from thousands of individuals.6 Interpret single excursions against those percentiles rather than against a fixed cutoff, and take persistent elevation to a clinician.
Footnotes
-
Xi Liao, Yongwei Li, Siwen Tang, et al. Continuous glucose monitoring in non-diabetic populations: a systematic review of observational and interventional studies with meta-analysis. European Journal of Medical Research, 2026. https://doi.org/10.1186/s40001-026-03920-0 ↩
-
David C. Klonoff, David Ahn, Andjela Drincic. Continuous glucose monitoring: A review of the technology and clinical use. Diabetes Research and Clinical Practice, 2017. https://doi.org/10.1016/j.diabres.2017.08.005 ↩ ↩2
-
Giacomo Cappon, Martina Vettoretti, Giovanni Sparacino, et al. Continuous Glucose Monitoring Sensors for Diabetes Management: A Review of Technologies and Applications. Diabetes & Metabolism Journal, 2019. https://doi.org/10.4093/dmj.2019.0121 ↩
-
David Rodbard. Continuous Glucose Monitoring: A Review of Successes, Challenges, and Opportunities. Diabetes Technology & Therapeutics, 2016. https://doi.org/10.1089/dia.2015.0417 ↩ ↩2 ↩3
-
Filip Wilczek, Jan Gerrit van der Stouwe, Gloria Petrasch, et al. Non-Invasive Continuous Glucose Monitoring in Patients Without Diabetes: Use in Cardiovascular Prevention—A Systematic Review. Sensors, 2025. https://doi.org/10.3390/s25010187 ↩
-
Ayya Keshet, Smadar Shilo, Anastasia Godneva, et al. CGMap: Characterizing continuous glucose monitor data in thousands of non-diabetic individuals. Cell Metabolism, 2023. https://doi.org/10.1016/j.cmet.2023.04.002 ↩ ↩2 ↩3
-
David Rodbard. Continuous Glucose Monitoring: A Review of Recent Studies Demonstrating Improved Glycemic Outcomes. Diabetes Technology & Therapeutics, 2017. https://doi.org/10.1089/dia.2017.0035 ↩
-
Nour Ahmed, Mohamed Faisal Elzein Ali, Muhanad Noureldin Hamed Mohamed, et al. Use of Continuous Glucose Monitoring in Non-diabetic Individuals for Cardiovascular Prevention: A Systematic Review of Its Impact on Guiding Lifestyle Interventions. Cureus, 2025. https://doi.org/10.7759/cureus.94460 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.