As we discussed in the first article in this series, there’s a lot of marketing about having More with your CGM, whether that’s accuracy or data. First up, we’re going to look at Dexcom’s G7, which we’re told is “More Accurate”.
When Dexcom tells us that the G7 is the most accurate CGM on the market, they’re saying it has the lowest Mean Absolute Relative Difference (MARD), but as we’ve discussed many times on this site in the past, MARD is dodgy ground and reflects circumstances of a clinical trial that has been undertaken, rather than real world accuracy, so isn’t really comparable across different trial data.
Indeed, early in its life, a lot of people made a lot of noise about the data the G7 produced (not to mention other issues relating to failure rates and other factors). At Diabettech, we compared the G6 and G7, and found that it’s quite hard to provide an absolute answer to this question. The brief investigation is shown below.
Variance Evaluation Metric assessment
The yellow and blue bars at the bottom of the image show the Variance Evaluation Metric, as described in the last post on this topic. Where the blue bars are taller, the G6 variation is higher and where the yellow bars are higher, it represents more variation from the G7. As the graph shows, it’s not clear cut that one has significantly more variation than the other.
I’ve removed periods where there were compression lows for one or other sensor from both sets of data as they don’t reasonably reflect normal use (although in a real world setting they create a different type of noise). This is another area where there was a noticeable difference between the two sensors. Bearing in mind the two were sitting next to each other on my arm, the G6 suffered four times more compression lows than the G7. Whilst it’s a different type of noise, it’s still noise.
In this close up, I’ve zoomed in to a specific part of the data which demonstrates how in some cases the G6 is more variable, and in others it’s the G7. In the six hour period shown, the G7 bars are taller in the first three hours, while in the second three it’s the G6. What really matters around this is timing. The G7 variance is around a time when a meal is likely to be eaten, and also when levels are climbing. The G6 is likely to be when someone is in bed. If this was a pattern that was repeated more regularly, ie. when levels are rising, the variability of the G7 was greater than the G6, it might present a problem with dosing insulin. In this dataset, there’s not enough data to show whether climbing glucose levels correlate with higher variability in one or other sensor though.
Overall, during the period that this analysis was undertaken, the mean VEM for the G6 was 3.2 while for the G7 it was 2.8, suggesting that there was, on average, more variance with the G6.
Alongside this, the standard deviation of the data over the period from the G7 is larger than that of the G6, with the G7 showing 47.1mg/dl and the G6 43.9mg/dl. This suggests greater dispersion on the part of the G7. Again, whether that’s because of the way in which it’s more accurate, it’s unclear.
If we look at the data in the course of this capture period, we see that 56% of the time, the VEM for the G6 was greater than that of the G7, while 44% of the time it was lower than than the G7.
Overall then, that last set of numbers suggests that there’s little real difference in the data, and that the “More Accurate” G7 probably isn’t a great deal more accurate in real world use.
Effect on Open Source systems
Given the general feedback on the G7, the similarity in the VEM metric to the G6 is unexpected. While there are times that it appears to be far more variable than the G6, the data collected here suggests that in general, it probably isn’t. Unfortunately, the dataset is too small to draw any real conclusions. A full test would require multiple of each sensor to be worn simultaneously by multiple users.
Never the less, as both data sources seem to have similar variability in their data, it makes sense to use smoothing on open source AID systems where it’s available. Both sensors appear to be subject to unexpected, outsize fluctuations from time to time, and using a smoothing algorithm in OS-AID is likely to result in safer insulin delivery.
So in the case of the Dexcom G7, is More better? On the specific use case we’re looking at here, I think the answer is marginally.
Leave a Reply