Continuous Glucose Monitoring in Athletes Without Diabetes: What the Data Can Support
A continuous glucose monitor worn by a healthy athlete is a useful instrument for comparing one session against another, and a poor instrument for reading a single number off the screen. Manufacturer-labeled accuracy for current sensors sits near 8 to 9 percent mean absolute relative difference (MARD) against a laboratory reference in adults with diabetes. Studies that compare CGM against venous or capillary reference during exercise in people without diabetes generally report worse performance, commonly in the range of roughly 10 to 20 percent MARD, with the largest errors during rapid glucose change and at the low end of the range. Since a non-diabetic athlete spends most of the day between about 70 and 110 mg/dL, a 12 percent error is on the order of 10 mg/dL, which is the same size as the effect most people are trying to detect. The signal worth trusting is the shape of the curve over tens of minutes, repeated across sessions.
What MARD measures, and why it degrades during exercise
MARD is the mean of the absolute differences between sensor and reference, each divided by the reference value, expressed as a percent. Two properties make it awkward here. It is a relative metric, so the same absolute error looks small at 200 mg/dL and large at 65 mg/dL. And it is an average over a study population and glucose range, so a headline number tells you little about performance in the specific condition you care about.
Exercise stresses the sensor in several ways at once. The sensor measures glucose in interstitial fluid, not blood, and equilibration plus on-device filtering introduces a lag typically in the 5 to 15 minute range. That lag costs you little when glucose is flat and costs you a lot at the onset of hard intervals or after a carbohydrate feed, when the true rate of change is steep. Changes in local perfusion, skin temperature and sweating during exercise also alter the sensor environment. The practical result is that error is concentrated exactly where athletes look hardest.
What 10 percent error means inside a narrow range
Consider a typical question: does a 60 gram carbohydrate feed before a threshold session change your glucose response compared with a fasted session? A plausible true difference at 30 minutes might be 15 to 25 mg/dL. With a sensor whose exercise error is around 10 mg/dL and which also carries a sensor-specific bias, a single paired comparison is close to a coin flip.
Two things make the comparison tractable. The first is that much of the error is systematic within a single sensor wear, so within-sensor differences are more reliable than absolute values. The second is repetition. If you run the same protocol four times in each condition on the same sensor, the random component averages down while any constant offset cancels in the difference. We would not draw a conclusion from one A/B pair, ever.
If you need absolute accuracy for a specific window, pair the sensor against capillary readings from a meter on a fixed schedule. A consumer meter is itself allowed substantial error under ISO 15197, so this gives you a sanity check rather than ground truth. It will catch a sensor that is running 20 mg/dL low, which is the failure that most often produces a dramatic and false story.
Failure modes that dominate during training
Compression artifacts are the most common source of spurious low readings. Sleeping on an arm sensor, a tight shoulder strap or a pack waist belt can produce a smooth dip of 20 to 40 mg/dL lasting 30 to 90 minutes, with a characteristic fast recovery when pressure is released. If a low appears during the night or under a strap and resolves within minutes of position change, treat it as an artifact rather than physiology.
Early-wear noise is the second. Both major sensors require a warm-up of about half an hour after insertion, and accuracy on the first day is typically worse than on days three through seven. We schedule any experiment that matters for day two onward and discard the first 24 hours.
Interference is specific to the chemistry. Ascorbic acid at high dose can bias Abbott sensors upward. Older Dexcom generations were sensitive to acetaminophen. Neither should be assumed away if you are supplementing heavily.
Finally, dropouts. Bluetooth loss during a long ride leaves gaps that back-fill differently depending on device and app. Check whether missing samples are missing or were reconstructed, because a reconstructed 20 minute gap will look implausibly smooth.
Getting the data out and aligning it with training files
Export from Dexcom Clarity gives a CSV with a Timestamp column in local time and a “Glucose Value (mg/dL)” column at five minute cadence. Values outside the reportable range appear as the strings Low and High rather than numbers, which silently coerces the column to object dtype in pandas and breaks every downstream calculation. Abbott’s LibreView export has a “Record Type” column that mixes historic, scan and note rows, plus separate glucose columns per record type. Filter before you do anything else.
import pandas as pd
df = pd.read_csv("clarity_export.csv", skiprows=range(1, 11))
df = df[df["Event Type"] == "EGV"].copy()
df["glucose"] = pd.to_numeric(
df["Glucose Value (mg/dL)"].replace({"Low": 39, "High": 401}),
errors="coerce",
)
df["ts"] = pd.to_datetime(df["Timestamp (YYYY-MM-DDThh:mm:ss)"])
s = df.set_index("ts")["glucose"].resample("5min").mean()
To align with training, parse the .fit file from your head unit with fitparse or fitdecode and resample power and heart rate onto the same five minute grid. Confirm both clocks agree. Head units set time from GPS, while phones and sensors use system time, and a 60 second offset matters when you are measuring the first ten minutes of an interval.
The metrics we compute per session are deliberately simple.
- mean and standard deviation over the work interval
- coefficient of variation (standard deviation divided by mean)
- the delta from the five minutes before the start to the trough or peak
- the time to return to baseline afterward
Rate of change in mg/dL per minute, computed over a 15 minute window rather than adjacent samples, is more informative than any instantaneous value and much less noisy.
What the readings can and cannot tell you
CGM in a healthy athlete measures a fueling and stress response. It does not measure a disease. Glucose can rise during hard intervals from catecholamine-driven hepatic output and fall during long steady work as liver glycogen depletes. Both are normal. A post-exercise reading in the 60s is common and frequently an artifact or a transient. Interpreting a single low as hypoglycemia, or a single post-meal peak as impaired glucose tolerance, is not supported by the measurement precision available.
If your sensor repeatedly shows fasting values above the normal range, or you have symptoms during or after exercise, that is a conversation for a physician with a proper laboratory test. A venous plasma glucose or HbA1c, run in a clinical lab, answers the diagnostic question that CGM cannot. Nothing here is medical advice, and we do not recommend any treatment, drug or dose.
The useful application is comparative. Hold the protocol fixed and vary one input. Repeat enough times to see through the noise, then read differences rather than absolutes. Interpreted that way, CGM tells you something real about how a particular fueling strategy behaves in your body. Interpreted as a blood test worn on the arm, it mostly generates stories.
Questions people also ask
What is the reported MARD for CGM in non-diabetic athletes during exercise? Published exercise studies generally report worse accuracy than the labeled resting figures of roughly 8 to 9 percent, commonly landing in the range of about 10 to 20 percent MARD. Performance degrades most during rapid glucose change and at lower glucose values, which is where a relative metric like MARD is least informative.
Is CGM approved for people without diabetes? Several sensors are now cleared for over-the-counter use in adults not taking insulin, depending on region and device generation. Regulatory clearance for sale is a separate matter from validated accuracy during exercise, which remains thinner.
Why does my glucose go up during hard intervals? High-intensity work raises catecholamines, which increase hepatic glucose output faster than muscle uptake can clear it. A rise of 20 to 40 mg/dL during sprints or short intervals is a normal response in people without diabetes.
How do I tell a compression low from a real low? Look at the shape and the context. Compression lows are smooth and occur while pressure is applied to the sensor site. They often last 30 to 90 minutes and recover quickly once the pressure comes off. A capillary check at the time settles it.
Does wearing two sensors help? Yes, if you care about the size of an effect rather than its direction. Two sensors on the same person give you an estimate of between-sensor disagreement, which is often 10 mg/dL or more and sets a floor on what any single sensor can resolve.
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.