Dexcom G7 - Comparison with finger stick readings
This page is not giving you medical advice or treatment advice, and that is very much on purpose! I am not a medical professional, and even if I were, I don’t know your personal circumstances. If you need medical advice, consult your physician! If, at any time, you have reason to doubt your G7 readings, use one or more finger-stick measurements to confirm your G7 readings. Base your treatment decisions (bolus, etc.) on the finger stick reading, and in accordance with your physician’s instructions.
For users with an insulin pump controlled by the G7: Until you have reached a full understanding of your G7’s behavior and when, in your particular case, its readings can be trusted, you may need to double-check, as frequently as necessary to give you confidence, with finger stick readings. Read the page on reading comparisons for important information on interpreting the differences.
Pages
- Introduction
- How it works
- Accuracy and Calibration
- Comparison with finger stick readings
- My Sensor Failed
- Known Issues
- Contacting Dexcom
Dexcom G7 - Comparison with finger stick readings
Based on what I have already described, it should be fairly clear that comparing finger stick readings and G7 readings taken at the same time is likely to be problematic if your blood sugar level has not been stable for at least 15 minutes. Otherwise, the time lag would cause the G7 readings not to reflect blood glucose levels within the specified accuracy.
So, assuming your glucose level was stable for the last 15-20 minutes (indicated by a horizontal arrow on the G7 indicator and a fairly “flat” graph), what might you be looking at? Keep in mind:
- The finger stick reading itself might not represent the actual glucose level. It may be up to ± 20% off. If the true glucose level is 100, the finger stick might read between 80 and 120, within the prescribed limits.
- Likewise, the G7 reading may deviate from the true glucose level as well, generally by up to 20-30%. If we use 25% as the number, and with true glucose at 100, that means readings between 75 and 125 are within expected limits 90% of the time.
- When comparing these two readings (finger stick and G7), there is a probability that they differ by a certain amount and still both be within limits. These probabilities can be calculated.
I asked ChatGPT for help with the calculations, and here is what I found, assuming the limits described above. We are using this formula:
\[\text{Relative difference} = \frac{∣BG_{fs} − BG_{g7}∣}{BG_{true}} × 100%\]This makes the table independent of the true blood glucose level, indicating only the probability of finding a difference in each bin, or in that bin, or in a smaller one (third column).
| Relative difference (%) | Probability (%) | Cumulative Probability (%) |
|---|---|---|
| < 2.5 | 10.2200 | 10.2200 |
| 2.5–7.5 | 19.7803 | 30.0003 |
| 7.5–12.5 | 17.9259 | 47.9261 |
| 12.5–17.5 | 15.2134 | 63.1396 |
| 17.5–22.5 | 12.0912 | 75.2308 |
| 22.5–27.5 | 8.9994 | 84.2302 |
| 27.5–32.5 | 6.2727 | 90.5029 |
| 32.5–37.5 | 4.0944 | 94.5973 |
| 37.5–42.5 | 2.5028 | 97.1001 |
| 42.5–47.5 | 1.4327 | 98.5329 |
| 47.5–52.5 | 0.7680 | 99.3009 |
| ≥ 52.5 | 0.6991 | 100.0000 |
How to read this intuitively:
- ~30% of the time, two “perfectly in-spec” devices will differ by less than 7.5% (second number, third column)
- ~15% of the time, they’ll differ by more than 22.5% (100% - third column, sixth number)
- ~9.5% of the time, they’ll differ by ≥ 27.5% (100% - third column, seventh number)
- Differences ≥ 50% are rare (~0.7%) but not impossible
The key takeaway here is that the entire table represents measurements where both devices are within their accuracy specifications! Just because you see a pair of measurements of 177 and 222, their 45 difference means a relative difference of 18%, which can happen 12% of the time, or any difference of 22.5% or less can happen 75% of the time. Conversely, a difference of more than 22.5% can happen 25% of the time.
Home work
In real life, we won’t know the true glucose level, but it’s reasonable to assume it is the average of the two measurements. Using that approach, the so-called “symmetric difference” formula becomes:
\[\text{percent_diff}_{sym} \;=\; \frac{|BG_{fs} − BG_{g7}|}{(BG_{fs} + BG_{g7})/2}\times 100 \;=\; \frac{2|BG_{fs} − BG_{g7}|}{BG_{fs} + BG_{g7}}\times 100\]The table, however, cannot be computed yet. What you need to do for yourself is collect close to 50 pairs of readings, compute their absolute difference, and count them in each of the indicated bins. Once that is done, the count column will be filled. You can then ask ChatGPT to fill in the other two columns by giving it the following instruction (copy and paste into an editor, fill out the count column, and copy and paste into ChatGPT):
Instruction for Computing Percent and Cumulative Percent from Binned Glucose Differences
I have a table of paired glucose readings from two devices.
For each pair of readings, I computed the symmetric percent difference using the formula:
[ \text{percent_diff} = \frac{2|x - y|}{x + y} \times 100 ]
I then binned each pair into the bins shown below and manually counted how many pairs fall into each bin.
Your task:
- Treat the numbers in the Count column as fixed and correct.
- Compute Percent for each row as:
[ \text{Percent} = \frac{\text{Count}}{\text{Total count}} \times 100 ]- Compute Cumulative percent as the running total of Percent, starting from the first row and proceeding downward.
- Do not recompute the bins, reinterpret the data, or recalculate the symmetric percent differences.
- Format the completed table as Markdown.
- Round all percentage values to two decimal places.
- Verify that the final cumulative percent equals 100% (allowing for rounding).
Table to Complete