Using a CGM for Fitness: What the Data Can and Cannot Tell You

Oak
cgm glucose exercise wearables data-analysis
A running horned beast in a foggy valley with a glowing amber band on its shoulder, leaving fading luminous footprints.

A CGM is a fine instrument for training, with a hard limit: it measures interstitial fluid glucose every 1–5 minutes with a lag of roughly 5–15 minutes behind blood, and a mean absolute relative difference (MARD) in the 8–10% range in manufacturer studies of resting adults. That is enough resolution to see how a given session, meal, or carb timing moves your glucose curve over hours. It is not enough to chase a 7 mg/dL difference between two intervals, and it says nothing directly about muscle glycogen, substrate oxidation, or power output. Wear it to compare protocols against yourself over weeks, not to pilot a single set.

What a CGM measures during a session

The sensor filament sits in subcutaneous interstitial fluid, not blood. Glucose diffuses across the capillary wall into that compartment, so the sensor sees a smoothed, delayed version of plasma. Manufacturers apply a predictive filter to partly compensate, which is why sensor readings can overshoot at inflection points.

Exercise makes this worse in a predictable way. Skin blood flow, local temperature, and hydration all shift during a hard session, and glucose is changing fast. In adolescents with type 1 diabetes exercising under CGM, sensor readings tracked the direction of change but showed meaningful deviation from reference values during and immediately after activity, with the largest errors when glucose was falling quickly.1 Reviews of CGM in physical activity reach the same conclusion: the signal is useful for trend and pattern, with accuracy degraded during rapid change and around the hypoglycemic range.2 In elite athletes, the practical guidance is to treat mid-session values as directional and confirm anything decision-relevant with a fingerstick.3

Translation for a non-diabetic reader: the shape of your post-session curve, the time to return to baseline, and the size of the glucose rise after a given breakfast are real and repeatable. A 10-point difference between two Tuesdays is not.

Where to wear it, including under a barbell

The approved site matters more than people assume. A head-to-head study of abdominal versus upper arm insertion found the upper arm more accurate, with lower MARD at the arm site, particularly during glucose excursions.4 A separate study of alternative Libre sites found that non-approved placements gave systematically different readings than the upper arm, so the site is part of the measurement, not a cosmetic choice.5

Practical rules we use:

  • Pick one site and keep it for the whole experiment. Arm and abdomen are not interchangeable data sources. Comparing an arm week to an abdomen week introduces a site effect that will masquerade as a diet effect.
  • Alternate left and right arm between sensors to avoid tissue irritation, and log which side in your notes. Sleeping side matters (see compression lows below).
  • Lifting is fine. The sensor is a 5 mm filament under adhesive, not something that tears loose under load. The failure mode is mechanical: a barbell racked on the back, a strap, a knurled bar dragged up the thigh, or a wetsuit sleeve. Back-of-arm placement clears most bar paths. If you do rack position work, the back of the arm below the deltoid is still exposed. Use an overpatch (Skin Grip, Simpatch, or the manufacturer’s own) and expect to reapply after heavy sweat.
  • Swimming and contact sports are the real risks. Adhesive survives water better than it survives friction.

Getting the raw data out

The phone app is a toy. Pull the CSV.

Dexcom: Clarity web app, Export → CSV. You get one row per 5-minute reading with Timestamp (YYYY-MM-DDThh:mm:ss), Event Type, and Glucose Value (mg/dL). Values below 40 export as the string Low and above 400 as High, so cast with care instead of astype(float). Timestamps are device-local with no timezone offset, which will bite you if you train across time zones.

Abbott: LibreView, Export glucose data → CSV. Columns include Device Timestamp and Record Type, where 0 is automatic historic glucose and 1 is a manual scan. Filter to Record Type == 0 for the continuous trace, then handle the fact that historic sampling is coarser than the app display on Libre 2.

A minimal pipeline:

import pandas as pd
df = pd.read_csv("clarity.csv", skiprows=range(1,11))
g = (df[df["Event Type"] == "EGV"]
       .assign(ts=lambda d: pd.to_datetime(d["Timestamp (YYYY-MM-DDThh:mm:ss)"]),
               mgdl=lambda d: pd.to_numeric(d["Glucose Value (mg/dL)"],
                                            errors="coerce"))
       .set_index("ts")["mgdl"]
       .resample("5min").mean()
       .interpolate(limit=3))   # cap gap-filling at 15 minutes

Cap interpolation. A two-hour sensor dropout filled by a straight line will quietly flatten your variability metrics.

Metrics worth computing

Mean glucose alone is nearly useless for a metabolically healthy person. These carry more information:

  • Coefficient of variation: g.std() / g.mean() * 100. Compute it over full 24-hour windows, not partial days.
  • Time in a tight range: fraction of readings in 70–120 mg/dL, and separately the fraction above 140.
  • Incremental AUC after a standardized meal: subtract the pre-meal baseline (mean of the 15 minutes before) and integrate the positive area over 120 minutes with numpy.trapz on minute-valued x. This is the cleanest single number for comparing two breakfasts.
  • Post-exercise return time: minutes from session end until glucose is back within 10 mg/dL of the pre-session baseline.
  • Overnight minimum and its timing, which is where compression artifacts live.

Compute all of these per-sensor-session and label the sensor ID. Sensor-to-sensor bias is the largest source of drift in a multi-week dataset, and a 10 mg/dL offset between two sensors is entirely normal. Older evaluations of CGM accuracy against reference methods made the same point: agreement varies with glucose range and rate of change, so the device is better at describing patterns than at assigning a single true number.6

Designing a test you can trust

Three things make the difference between a real result and a story.

Repeat. Run each condition at least three times on separate days. With ~9% MARD plus normal day-to-day biological variation, single-trial differences under about 15–20 mg/dL in peak, or under ~20% in iAUC, are not distinguishable from noise.

Randomize order. Alternate A/B/A/B rather than a week of A then a week of B, so sensor drift and sensor changes do not line up with your condition.

Control the obvious confounders. Same wake time, same sleep duration, same caffeine, same prior-day training load. If you menstruate, note cycle day: CGM data across the menstrual cycle shows measurable differences in glucose variability between phases, enough to swamp a modest dietary effect if conditions are unbalanced across a month.7

A worked example: to compare 30 g of carbohydrate pre-session against fasted training, fix the session (same route, same power target, same time of day), run six sessions alternating conditions, and compare the distribution of post-session return times and the minimum glucose reached in the four hours afterward. That is a question a CGM can answer about you.

Artifacts that will fool you

Compression lows. Lying on the sensor compresses the tissue, reduces local perfusion, and produces a sharp drop of 20–40 mg/dL that recovers within 20–60 minutes of rolling over. They look like nocturnal hypoglycemia. Pattern: abrupt onset, flat floor, abrupt recovery, always in the same sleep window, always on the side you sleep on. Do not report these as events.

First-day drift. Many sensors read low and noisy for the first 6–24 hours. We drop the first 12 hours of every sensor session before analysis.

Sensor-change discontinuity. Insert the new sensor while the old one is still running, keep both for 12 hours, and use the overlap to estimate the offset. It costs one sensor and saves an entire longitudinal series.

Acetaminophen and vitamin C. Older Dexcom generations were sensitive to acetaminophen; high-dose vitamin C interferes with Libre chemistry. Log supplements.

None of this is a substitute for clinical measurement. If your CGM shows readings that look like real hypoglycemia, persistent fasting values above normal, or you are considering anything about diabetes risk, that belongs with a physician and a venous lab draw, not a wearable.

Questions people also ask

Can I put a CGM on my belly? Some systems are labeled for abdominal use, but the upper arm performed better in a direct comparison, with lower MARD at the arm site.4 Non-approved sites give systematically different readings.5 Pick the labeled site and stay there.

Can you lift weights with a CGM? Yes. The filament tolerates muscle contraction and load. The risk is the adhesive being scraped off by a bar, strap, or sleeve. Back-of-arm placement plus an overpatch handles nearly all of it.

Can an Apple Watch or Fitbit check blood sugar? No. No consumer smartwatch on the market measures glucose. Non-invasive optical glucose sensing has not cleared regulatory validation, and the FDA has warned against devices claiming to do it. Anything showing glucose on a watch face is relaying data from a separate CGM sensor over Bluetooth.

How fast can HbA1c change? HbA1c reflects glycation of hemoglobin over the ~120-day lifespan of a red blood cell, weighted toward the most recent 4–6 weeks, so a repeat test is informative at about 3 months. What that number should be, and what to do about it, is a conversation with a clinician.

Does exercise show up on a CGM? Yes, and the response depends on the session. Prolonged aerobic work typically lowers glucose during and after the session, while short, high-intensity efforts often produce a transient rise driven by catecholamines. CGM captures both patterns and the multi-hour post-exercise window, which is the part you cannot see with fingersticks.2 Real-time feedback of this kind has been examined as a driver of behavior change in people at metabolic risk.8

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

  1. Peter Adolfsson, Staffan Nilsson, Bengt Lindblad. Continuous glucose monitoring system during physical exercise in adolescents with type 1 diabetes. Acta Paediatrica, 2011. https://doi.org/10.1111/j.1651-2227.2011.02390.x ↩

  2. Oliver Schubert-Olesen, Jens Kröger, Thorsten Siegmund, et al. Continuous Glucose Monitoring and Physical Activity. International Journal of Environmental Research and Public Health, 2022. https://doi.org/10.3390/ijerph191912296 ↩ ↩2

  3. Gurneet Brar, Sean Carmody, Alistair Lumb, et al. Practical considerations for continuous glucose monitoring in elite athletes with type 1 diabetes mellitus: A narrative review. The Journal of Physiology, 2024. https://doi.org/10.1113/jp285836 ↩

  4. Isabelle Isa Kristin Steineck, Zeinab Mahmoudi, Ajenthen G. Ranjan, et al. Comparison of Continuous Glucose Monitoring Accuracy Between Abdominal and Upper Arm Insertion Sites. Diabetes Technology & Therapeutics, 2019. https://doi.org/10.1089/dia.2019.0014 ↩ ↩2

  5. Rosemary M. Hall, Sophie Dyhrberg, Arthur McTavish, et al. Where can you wear your Libre? Using the FreeStyle Libre continuous glucose monitor on alternative sites. Diabetes Obesity and Metabolism, 2021. https://doi.org/10.1111/dom.14630 ↩ ↩2

  6. Roger S. Mazze, Ellie S. Strock, Sarah Borgman, et al. Evaluating the Accuracy, Reliability, and Clinical Applicability of Continuous Glucose Monitoring (CGM): Is CGM Ready for Real Time?. Diabetes Technology & Therapeutics, 2009. https://doi.org/10.1089/dia.2008.0041 ↩

  7. Georgianna Lin, Rumsha Siddiqui, Zixiong Lin, et al. Blood glucose variance measured by continuous glucose monitors across the menstrual cycle. npj Digital Medicine, 2023. https://doi.org/10.1038/s41746-023-00884-x ↩

  8. Nicole Ehrhardt, Enas Al Zaghal. Behavior Modification in Prediabetes and Diabetes: Potential Use of Real-Time Continuous Glucose Monitoring. Journal of Diabetes Science and Technology, 2018. https://doi.org/10.1177/1932296818790994 ↩