Wearing a Dexcom Without Diabetes: What the Data Can and Cannot Tell You
A Dexcom sensor worn by someone without diabetes is a reasonable measurement instrument and a poor diagnostic one. With roughly 8-10% error, it will tell you what your interstitial glucose did after a specific meal. The same holds for a specific run or a bad night of sleep. It will not tell you whether you have a metabolic disease, and the clinical evidence base for using it outside diabetes is thin 1. If you treat it as an instrument for controlled personal experiments rather than a screening test, the money is well spent. Below is how we would run it.
What the sensor measures
The sensor does not sit in blood. Its small filament sits in subcutaneous interstitial fluid. It runs a glucose oxidase reaction and reports a current that firmware converts into a glucose estimate. Dexcom G7 outputs a value every 5 minutes. The over-the-counter Stelo sensor, built on the same platform, reports every 15 minutes.
Two consequences follow from the interstitial site. First, there is a physiological lag of roughly 5 to 10 minutes between blood and interstitial glucose, and the lag widens when glucose is changing fast. Second, accuracy is worst exactly where a non-diabetic reader spends most of their time. Mean absolute relative difference, the standard accuracy metric for these systems, is reported around 8-9% for recent Dexcom sensors against a laboratory reference 2. That figure is an average across a wide glucose range. In the 70-140 mg/dL band, a 9% relative error is roughly 7-12 mg/dL, which is the same size as many of the meal differences people try to detect.
Accuracy also degrades in populations and physiologies far from the calibration set. A study of G6 Pro and G7 in patients on maintenance dialysis found meaningfully worse agreement than the manufacturer figures, with performance varying by sensor generation 3. The general lesson holds for anyone. The headline accuracy number describes a trial cohort. It does not describe you.
Which sensor to buy
For a technically fluent reader who wants data, we would buy Dexcom G7 if a prescription is available and Stelo if not.
G7 gives 5-minute resolution and a 30-minute warmup. It wears for 10 days with a 12-hour grace window and reports down to 40 mg/dL. Stelo requires no prescription in the United States, wears for 15 days, and costs less per day. The tradeoffs are real: 15-minute sampling blurs postprandial peaks, and Stelo clips its displayed range at the low end, so it is a weak instrument if you are specifically interested in low readings.
If your question is “what does rice do to me versus lentils,” Stelo is adequate. Buy the 5-minute sensor if your question involves exercise, sleep-onset dynamics or anything with sharp slopes.
Getting the data out
Three paths, in descending order of how much we like them.
The Dexcom Clarity web export is the simplest. Log in at clarity.dexcom.com, choose the date range, and download a CSV. The file contains a timestamp column, an event type column, and a glucose column in mg/dL or mmol/L. Values below the reporting floor appear as the string Low rather than a number, so parse with pd.to_numeric(errors="coerce") and decide explicitly what to do with the NaNs rather than letting pandas drop them silently.
The Dexcom Developer API is the right choice if you want automation. It is OAuth2, with a sandbox environment that returns synthetic data for development. The production endpoint serves estimated glucose values in three-hour-delayed batches, which is fine for analysis and useless for anything real-time.
Local Bluetooth capture with xDrip+ on Android gives you 5-minute values with no delay and no cloud round trip. It is unsupported, it breaks when Dexcom updates sensor firmware, and it will occasionally fight with the official app over the sensor connection. We use it anyway when we need raw backfill.
Whichever path you take, log meals and activity in a separate file with ISO 8601 timestamps in the same timezone. Untimestamped meal notes make an entire two-week dataset uninterpretable.
The metrics worth computing
Start by resampling to a regular grid. With a 5-minute sensor, df.set_index("ts").resample("5min").mean().interpolate(limit=3) fills short dropouts without inventing hours of data. Anything longer than a 15-minute gap should stay missing.
Then compute the standard set, which is well described in the diabetes literature and transfers cleanly to a non-diabetic context 4:
- Mean glucose and the glucose management indicator,
GMI = 3.31 + 0.02392 * mean_mg_dl. - Coefficient of variation, standard deviation divided by mean. Below 36% is the conventional stability threshold in diabetes, and most people without diabetes sit far below it.
- Time in range, though the standard 70-180 mg/dL band is nearly meaningless for a non-diabetic who lives in it 98% of the time. Tighten the band to 70-140 mg/dL for a usable number.
For meal experiments, the useful quantity is incremental area under the curve over two hours. Take the mean of the 15 minutes before the first bite as baseline. Subtract it from every subsequent reading, clip negatives to zero and integrate with numpy.trapz. Report peak value, time to peak, and 2-hour iAUC together. A meal with a 55 mg/dL peak that returns to baseline in 90 minutes is a different event from a 40 mg/dL peak that lingers for three hours.
On duration: fourteen days of continuous data is the usual benchmark for estimating longer-term glycemic behavior, and shorter windows estimate mean glucose better than they estimate variability 5. Plan for at least two sensors.
Failure modes that will fool you
Compression lows are the most common artifact. Lying on the sensor restricts local perfusion and produces a smooth 40-60 mg/dL trough lasting 30 to 90 minutes, almost always overnight, with a rapid recovery on waking. These are not hypoglycemia. Flag any nocturnal low that resolves within minutes of a position change and exclude it before computing variability.
Day-one noise is real. Many sensors read low or erratically for the first 12 to 24 hours after insertion while the insertion site settles. We drop the first 12 hours of every sensor from analysis.
Sensor-to-sensor variability is the reason single-sensor conclusions are fragile. Two sensors worn simultaneously can differ by 10-15 mg/dL at the same instant, and the offset can persist across the whole wear. If you want to compare two foods, eat both on the same sensor and ideally within a few days. Repeat each at least twice.
Finally, the largest source of spurious findings is confounding. Sleep, prior exercise and meal order all shift postprandial response. So do time of day and stress. Randomize the order of your test meals and keep the rest of the day similar.
Where a clinician belongs
A CGM is not a diagnostic test for diabetes or prediabetes. Diagnosis rests on fasting plasma glucose, HbA1c or an oral glucose tolerance test run by a laboratory. The interpretation belongs to a physician 1 6. Your sensor may show repeated fasting readings above roughly 100-110 mg/dL, sustained post-meal excursions above 180 mg/dL or genuine lows with symptoms. In those cases, take the data to a clinician and ask for venous testing. Do not change medication on the basis of interstitial readings.
Questions people also ask
Do I need a prescription to wear a Dexcom without diabetes? Not in the United States. Stelo is sold over the counter for adults not using insulin. The prescription G7 gives finer time resolution and a wider reporting range.
How accurate is a CGM at normal glucose levels? Modern Dexcom sensors report mean absolute relative difference near 8-9% against laboratory reference across their full range 2. At 100 mg/dL that is an error band of roughly 8-9 mg/dL, which is large relative to the differences most people are trying to detect.
How long should I wear one to learn something? Around fourteen days of good data is the standard window for characterizing typical glucose behavior, with variability metrics needing longer than mean glucose does 5.
Can I get raw sensor current instead of calibrated glucose? No. Dexcom exposes calibrated estimated glucose values only, through Clarity exports and the developer API. The raw electrochemical signal stays inside the transmitter firmware.
Is time in range a useful metric for me? Only if you redefine the range. The clinical 70-180 mg/dL target was built for diabetes management 4. A 70-140 mg/dL band produces a number with some variance for a non-diabetic reader.
Footnotes
-
Anna Wood, David O’Neal, John Furler, et al. Continuous glucose monitoring: a review of the evidence, opportunities for future use and ongoing challenges. Internal Medicine Journal, 2018. https://doi.org/10.1111/imj.13770 ↩ ↩2
-
John B. Welsh, Peggy Gao, Mark Derdzinski, et al. Accuracy, Utilization, and Effectiveness Comparisons of Different Continuous Glucose Monitoring Systems. Diabetes Technology & Therapeutics, 2019. https://doi.org/10.1089/dia.2018.0374 ↩ ↩2
-
Leila R. Zelnick, Subbulaxmi Trikudanathan, Yoshio N. Hall, et al. Accuracy of Dexcom G6 Pro and G7 Continuous Glucose Monitors in Patients Treated with Maintenance Dialysis. Diabetes Technology & Therapeutics, 2025. https://doi.org/10.1177/15209156251368934 ↩
-
Jared G Friedman, Kasey Coyne, Grazia Aleppo, et al. Beyond A1C: exploring continuous glucose monitoring metrics in managing diabetes. Endocrine Connections, 2023. https://doi.org/10.1530/ec-23-0085 ↩ ↩2
-
Tonya D. Riddlesworth, Roy W. Beck, Robin L. Gal, et al. Optimal Sampling Duration for Continuous Glucose Monitoring to Determine Long-Term Glycemic Control. Diabetes Technology & Therapeutics, 2018. https://doi.org/10.1089/dia.2017.0455 ↩ ↩2
-
Mark R. Burge, Stephen Mitchell, Alison Sawyer, et al. Continuous Glucose Monitoring: The Future of Diabetes Management. Diabetes Spectrum, 2008. https://doi.org/10.2337/diaspect.21.2.112 ↩
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.