Investigating the rules of 500 and 1800 looking at real world data

For a long time, clinicians have used the rules of 1800 or 100 (depending on which units you use) for Insulin Senstivity Factor (ISF), or correction factor, and 500 or 300 for figuring out starting levels for your Carb Ratio (CR) at different times of day when starting users on insulin. They may also use the “2.6 rule” to figure out the carb ratio.

Back in 2010, John Walsh et al wrote this paper looking at the distribution of CR and ISF based on data taken from the Cosmo pump in 2007 and 2008. It surmised that people used magic numbers and old settings in setting up their insulin pump bolus calculator and seemed unlikely to change.

Rolling forward 10 years, capture of pump and glucose data is rather easier, and some would say routine, and Tidepool is one centre of this captured data. in 2018, they released the following data, looking at 803 pump users. 

They summarise this as:

What interested me in this dataset was that it might present a way to look at whether the the relationship between median total daily dose and the median Carb Ratio and Sensitivity aligned with the models that are often used to get someone set up. While there are clear flaws in an analysis that only looks at these factors, it represents an starting point for some analysis.

The Data

In the first table below is the median data we can extract from Tidepool’s post. It shows the median Total Daily Dose (TDD), ISF and CR for the different age groups of the participants.

Age Group	TDD	CR	ISF
0-5	15	18	205
6-8	18	20	140
9-11	28	15	80
12-14	46	11	55
15-17	50	10	40
18-20	42	11	50
21-24	44	9	40
25-29	45	9	40
30-34	39	10	50
35-39	45	9	35
40-49	40	10	50
50-59	41	10	45
60-69	37	10	40
70-85	27	12	35

Going on from this, using the TDD I calculated the equivalent values for ISF and CR that the formulas provide, and charted them against the provided values:

Insulin Sensitivity Factor

As we go through the age groups, we can see that the median ISF and output of the rule of 1800 based on median TDD have a pretty close correlation throughout adulthood in the tidepool population. Perhaps not hugely surprisingly, children and teenagers see greater variation from these calculated values, as do those over the age of 60.

Carb Ratio

Median carb ratios seem to have significantly greater variation from the rule of 500 based off median TDD, and we can see that this is consistent across most age groups, with only those in the 12-17 age group having anything approaching the calculated value. 

Variance from model

For both the values, there is some variance from the standard model, and here we seek to see how great that is. 

I’ve included +/- 15% bands on the chart to show where the used values deviate significantly from the calculated values. It’s clear here that the median used CR value is regularly 20% lower than the median TDD calculated CR. It’s also possible to see that the ISF value seems to be much more varied than the CR.

Discussion

Ignoring the concerns related to using median data and that we can’t see whether participants are using one or more ISFs and CRs throughout the day, we can see that real world values from Tidepool’s dataset are substantially different from the values calculated by the rule of 500. In addition, there is a reasonable amount of variation in the ISF value, and the variance of CR from the model doesn’t seem to be correlated to any great extent with the variation of ISF.

If we look at the CR values specifically, we can see that the median values fall in general, around 20% lower than the calculated values. with the exception of those aged from 12-20, and this raises a number of questions. 

With CR , we can see in the data set that from the age of 12-69, the real world CR remains very close to10g/iu, +/-1g/iu. I speculate that most people are put on 10g/iu rather than a rule of 500 value when diagnosed and that this doesn’t get changed very much. Given that a large proportion of the data comes from those within 5 years of diagnosis, this is may explain why. We can also see that this lies some 20% off the rule of 500 model in many cases.

The original data source also stated that the split of basal/bolus amongst adult participants was generally 50/50 (although with a fair amount of variation). The combination of the two data points therefore raises questions about both these factors. Does the idea of a 50:50 split between basal and bolus pervade decisions that people make in relation to how they adjust doses? Do people leave a standard value in place for their carb ratio because it is simply easier when it is “about 10g/iu”? Or is it that people learn how to adjust basal rates to deal with hypos and then don’t think of looking at mealtime bolusing as the culprit?

ISF, meanwhile, seems to be much more closely correlated, even if there is a reasonable amount of variation across the population. Generally, 50% of adults over the age of 20 and below the age of 60 are less than 15% away from the model calculated values. Given that there appears to be a high correlation relating to 50:50 split of basal and bolus, it is possible that this also affects ISF and how people go about calculating that, indeed, if the CR is generally 20% below the model value, does that mean that the corresponding adjustments to basal give incorrect ISF values?

However we look at this, the data doesn’t give us a huge amount of insight into how people have their pump set up. It does, however, raise a number of questions that probably need further investigation:

1. Does clinician advice to achieve a 50:50 basal bolus split drive behaviours that may not produce an optimal insulin delivery model?
2. As John Walsh et al stated in 2010, people use magic numbers. This data suggests that this trend may not have changed that much.  
3. There is no link in the Tidepool data between CGM data and time in range and users TDD, CR and ISF, which may perhaps produce some interesting correlations. This would be a useful overlay.
4. If there was a way to run an algorithm over each individual participant’s dataset to determine optimal outcomes, what would we find about their set up and would a user by user calculated dataset produce something that fitted the two rules at the start?

I think it’s fair to say that the data Tidepool has gathered suggests that the rule of 500 and the rule of 1800 are not generally followed in the real world and that raises two things:

1. Should a better way to guesstimate initial settings be created, based on the vast amounts of data that are now generated by many thousands of users?
2. Is better education needed for users (and maybe HCPs as well) relating to what to do about adjusting dosing settings?

Either way, now that we are in the world of better data collection, the time seems right for a review of these long standing paradigms.