Best CGM for Athletes

If you want one recommendation: use a FreeStyle Libre 3 Plus, and capture the data with Juggluco or xDrip+ rather than the official app. It samples every minute, wears for 15 days, and needs no fingerstick calibration. The one-minute cadence matters because glucose during intervals can move faster than a five-minute sensor can describe. Take the Dexcom G7 instead if you need alarms or a documented developer API. The same applies if you share data with a clinician who uses Dexcom Clarity, and in both cases you accept five-minute resolution.

Why sampling rate decides this more than accuracy

Head-to-head accuracy between the current Abbott and Dexcom sensors is close enough that it should not drive your choice. Both report overall mean absolute relative difference (MARD) near 8% in adults, measured against venous or YSI reference in mostly resting conditions. MARD is the average percent error against a laboratory reference. That number degrades during exercise, and neither manufacturer characterizes it well for athletes.

The degradation has a physical cause. A CGM electrode sits in interstitial fluid, not blood, so it reports a delayed and smoothed version of plasma glucose. The lag is commonly 5 to 15 minutes and widens when glucose is changing quickly. Exercise makes it worse by changing skin perfusion and local fluid dynamics under the sensor.

So during a 4x4 minute VO2max session, a five-minute sensor gives you roughly four points across the whole set, each already smoothed and delayed. A one-minute sensor gives you twenty. Neither is truth, but the denser stream lets you fit and correct the lag instead of guessing at it. Reviews of CGM in non-diabetic athletes consistently flag this interstitial delay as the main interpretive limit on session-level analysis.¹

The device shortlist

FreeStyle Libre 3 Plus is our default. It is a single-piece applicator worn on the back of the upper arm, with 15-day wear and one-minute readings. Warmup takes roughly one hour. Abbott only validates upper-arm placement, which constrains what you can do about compression during sleep.

Dexcom G7 is the alternative. Ten-day wear plus a 12-hour grace period, 30-minute warmup, five-minute readings. Its advantage is the ecosystem: configurable alerts, a published developer API, and Clarity exports that clinicians already read. Abdomen and upper arm are both labeled placements for adults, which gives you more options when one site keeps failing.

Skip the over-the-counter tiers for training analysis. Stelo (Dexcom) and Lingo (Abbott) are the consumer versions of the same hardware with software restrictions layered on. Stelo updates roughly every 15 minutes and is not designed to display or alarm on lows. A cadence that slow is unusable for interval work. Eversense is an implanted 365-day sensor that requires a clinic procedure and periodic fingerstick calibration, which is a poor fit for someone who wants to iterate on their own data. Supersapiens, which several older articles still recommend, wound down its operations in 2024. Do not build a workflow on it.

Getting the raw stream out

This is where most people lose their data, so plan it before you apply the sensor.

For Libre, the official LibreView CSV export gives you 15-minute historic values in rows tagged Record Type = 0, with columns like Device Timestamp and Historic Glucose mg/dL. The export does not contain the one-minute stream. The dense data exists only in the Bluetooth stream the phone app consumes. To keep it, activate the sensor with Juggluco or xDrip+ on Android and let that app own the BLE connection. Both can push to a Nightscout instance or write local SQLite. You cannot run the official Libre app and a third-party reader against the same sensor session, so decide at activation.

For Dexcom, Clarity exports five-minute values as CSV with an ISO-8601 Timestamp (YYYY-MM-DDThh:mm:ss) column and a Glucose Value (mg/dL) column. Readings below 40 and above 400 appear as the strings Low and High, which will silently poison a float cast if you do not handle them. For programmatic access, the official Dexcom API v3 serves estimated glucose values with a delay for most partner tiers. The Share endpoint used by pydexcom returns near-real-time values and is what people use for live dashboards, with the caveat that it is not a supported public interface.

Whichever path you take, normalize into one table: UTC epoch milliseconds, glucose in mg/dL, sensor session ID, and a source flag. Resample to a one-minute grid with linear interpolation, and mark interpolated rows so you never compute a rate of change across a gap.

Aligning glucose with training files

The payoff is joining glucose to power and heart rate. Export your sessions as FIT files and parse them with fitparse in Python. Iterate get_messages('record') and pull timestamp, power, heart_rate, and cadence. FIT timestamps are seconds since 1989-12-31T00:00:00Z, and fitparse already converts them to UTC datetimes. Your phone clock and your bike computer clock both drift, so verify against a known event before trusting sub-minute alignment.

Estimate the interstitial lag empirically rather than assuming a constant. Take a session with a sharp carbohydrate intake or a clear intensity step. Cross-correlate the glucose first derivative against the forcing signal across shifts of 0 to 20 minutes and take the argmax. Expect a value in the 5 to 15 minute range, varying by sensor site and session.

Then compute a small set of session metrics: incremental area under the curve over a fixed window using the trapezoid rule, glucose at the start of the main set, maximum rate of change in mg/dL/min, and the time to return to the pre-session baseline. Coefficient of variation (standard deviation divided by mean) over 24 hours is the cleanest summary of overall variability. Reference ranges for these metrics in healthy athletes are not established, so treat your own baseline as the comparator.²

Failure modes that will cost you a session

Compression-induced lows are the most common artifact. Lying on the sensor restricts local flow and produces a smooth 30 to 60 minute dip that recovers abruptly when you roll off. It looks like a nocturnal hypoglycemic event and is not one. Flag any drop that resolves in under ten minutes as suspect.

First-day readings are the least reliable of the wear period for both sensors. If you are comparing two hard sessions, apply the sensor two days before the first one.

Sweat and shear kill adhesion. Use an overpatch, and apply it before the ride. Do not wait until the edge lifts. For open-water swimming, both sensors are rated for shallow immersion for limited periods. Bluetooth does not travel through water, so expect gaps that backfill on resurfacing.

Finally, the evidence base is thin. Systematic review of CGM in non-diabetic athletes finds that most studies are small and short, with heterogeneous protocols and few performance outcomes.³ Expert commentary on non-diabetic CGM use has similarly cautioned that the tool is ahead of the interpretive framework.⁴ Use CGM to describe your own responses. Do not use it to chase a target number someone published.

If you see repeated readings outside the normal range, or symptoms during or after exercise, that is a question for a physician. A consumer CGM is not a diagnostic instrument, and glucose management in anyone with diabetes is a clinical matter requiring specialist input.⁵

Questions people also ask

Is the Dexcom G7 more accurate than the Libre 3 Plus? The published MARD values are close, both near 8% in adults under mostly resting conditions. Neither is characterized well during high-intensity exercise, so the difference is unlikely to matter for training analysis. Sampling rate and data access are the better decision criteria.

Does a CGM improve athletic performance? There is no good evidence that it does on its own. Current reviews describe CGM in healthy athletes as a promising monitoring tool with limited outcome data behind it.³ Its value is descriptive: seeing how a specific fueling protocol behaves for you.

How much does exercise distort the reading? Enough to matter during rapid changes. Interstitial lag of roughly 5 to 15 minutes plus perfusion changes under the sensor mean the trace understates both the speed and the size of fast swings.¹ Estimating the lag from your own data is more useful than assuming a fixed offset.

Can I get the one-minute Libre data without a third-party app? No. The LibreView CSV export contains 15-minute historic values. Capturing one-minute data requires an app that reads the Bluetooth stream directly, and that decision has to be made when you activate the sensor.

Footnotes

1. Mikael Flockhart, Filip J. Larsen. Continuous Glucose Monitoring in Endurance Athletes: Interpretation and Relevance of Measurements for Improving Performance and Health. Sports Medicine, 2023. https://doi.org/10.1007/s40279-023-01910-4 ↩ ↩²

2. Joanna Wasik, Marta Armuła, Claire Bongage, et al. Continuous Glucose Monitoring in Healthy Athletes: Current Evidence, Practical Applications and Limitations: A Narrative Review. Quality in Sport, 2026. https://doi.org/10.12775/qs.2026.69.74446 ↩

3. Emilia Browarska, Katarzyna Superson, Katarzyna Miemczyk, et al. Continuous Glucose Monitoring (CGM) as a Tool for Training Quality Optimization: A Systematic Review of Metabolic Management in Non-Diabetic Athletes. Quality in Sport, 2026. https://doi.org/10.12775/qs.2026.67.74262 ↩ ↩²

4. David C. Klonoff, Kevin T. Nguyen, Nicole Y. Xu, et al. Use of Continuous Glucose Monitors by People Without Diabetes: An Idea Whose Time Has Come?. Journal of Diabetes Science and Technology, 2022. https://doi.org/10.1177/19322968221110830 ↩

5. Danilo Caponio, Giuseppina Alessia Acucella, Michele Acucella. Exercise and sport in type 1 diabetes in the continuous glucose monitoring and automated insulin delivery era: a phenotype-organized, timing-aware practical narrative review. Sport Sciences for Health, 2026. https://doi.org/10.1007/s11332-026-01878-4 ↩