Objectives

This is the follow-up to my post from August 19th, which you can find here, if you haven’t already read it. In that post, I describe a table from a Medtronic report with a series of descriptive measures of blood glucose levels. Today, we will use Continuous Glucose Monitor (CGM) data from Karen’s Nightscout rig to create a series of clinically similar measures. Medtronic measures using CGM data can be replicated exactly using Nightscout CGM data. Medtronic measures using manual blood glucose data cannot be replicated exactly. We will use CGM data to create a clinically similar measures. Because of physiological and sample frame differences between manual blood glucose tests and interstitial tissue glucose readings taken by a CGM, they will not be identical but they will be very similar. The Medtronic report has two goals:

1. Provide Karen and her endocrinologist with a basic understanding of Karen’s blood glucose control.
2. Help them identify ways in which she can do more to control her blood glucose levels.

Aggregated measures using CGM or manual blood glucose tests have the same clinical goal, to inform Karen’s diabetes self-management. In aggregated form, CGM data is not necessarily inferior to manual tests. In fact, CGM data may be superior. I will discuss the medical and sampling differences between the two technologies another day. In the meantime, we will treat measures based on CGM data as being clinically similar to measures based on manual blood glucose tests.

The objectives of today’s post are:

1. Import CGM data from a CSV file.
2. Clean-Up / Process the data (Data Munging!) to prepare it for use.
3. Use R to produce descriptive tables which are clinically similar to three of the data elements provided in the Medtronic Statistics table.

• The Medtronic Statistics table provides descriptive statistics for a single time period. The demo code provided here returns descriptive statistics from four weeks earlier this summer. This allows us to understand if values from any given week are an aberration or are normal for Karen.
• Use Nightscout CGM data to build the measures. It’s what we have, we will use it.

1. Learn more about Karen’s average blood glucose levels and use this to discuss the diabetes management goals of this project.

All code will be shown and documented.

Code, Data, Dependencies

Code: The .Rmd file used to create this post is available on GitHub.

Data Location: The data used here can be found in the data sub-folder on the project page.

Data Warning: Our current rig has been in nearly continuous use since June 21, 2015. Data collected prior to June 2015 is inconsistent due to limitations with the previous rig and mistakes we made figuring out how to use it. I do not recommend using data collected prior to June 2015 for analysis.

Reproducibility Warning: File names and content are subject to periodic updates. Open a bug or download the repository from GitHub for long-term use.

The only dependencies needed to run the code in this document are dplyr and ggplot2 from the Hadleyverse and pander which produces the formatted table output. To compile the HTML from the .Rmd file you will need the rmarkdown package. Readers familiar with GNU Makefiles can use the Makefile for compiling this document. Others are encouraged to use a tool such as RStudio which simplifies the process.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

## Dependencies ----------------------------------------------------------------
    ## If any of these fail, run the following (capitalization counts):
    ## install.packages("name_of_package")
    library(dplyr)
    library(ggplot2)
    library(pander)

    ## Config ----------------------------------------------------------------------

    ## Prevents pander from wrapping the table @ 80 chars.
    panderOptions('table.split.table', Inf)

Data Import & Munging

The following code chunk imports the most recent CSV file available in the data folder and stores it in a data frame called “entries”. This name was chosen for consistency with Nightscout data structures.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21

## Import data------------------------------------------------------------------
    ## Retrieves a list of CSV files in the data dir.
    ## Imports the most recent file, assuming the file naming schema is followed.
    files <- dir("data", "*csv", full.names=TRUE)
    entries <- read.csv(files[length(files)], as.is=TRUE)

    ## Data Munging ----------------------------------------------------------------

    ## Removes unnecessary SGV rows.
    entries <- entries %>%
        filter(type == "sgv")

    ## Tells R the "date" column is full of dates.
    entries$date <- as.POSIXct(entries$date)

    ## Week Labels - Useful labels we will be to aggregate by later.
    entries$wk_lbl <- NA
    entries$wk_lbl[ entries$date >= '2015-06-28' & entries$date <'2015-07-05' ] <- 'Week 1: June 28 - July 04'
    entries$wk_lbl[ entries$date >= '2015-07-05' & entries$date <'2015-07-12' ] <- 'Week 2: July 05 - July 11'
    entries$wk_lbl[ entries$date >= '2015-07-12' & entries$date <'2015-07-19' ] <- 'Week 3: July 12 - July 18'
    entries$wk_lbl[ entries$date >= '2015-07-19' & entries$date <'2015-07-26' ] <- 'Week 4: July 19 - July 25'

Nightscout data is a time-series data set. The descriptive statistics shown below aggregates CGM readings by week. The “Week Labels” column (wk_lbl) is a simple way to aggregate, order and label the data. Future posts will demonstrate how to graph individual CGM readings to look for patterns. Eventually, I will demonstrate more formal time series analysis methods.

Descriptive Statistics

These are the measures of interest found in the Medtronic Statistics table. The Replicated column identifies which measures are replicated more or less exactly and which are clinically similar, due to the differences in source data as previously discussed.

In this table, Medtronic uses manual blood glucose tests far more frequently than it uses CGM data. I don’t know why, but the graphics in the report, which we will discuss soon, are all based on CGM data. This is the only place in the report which is so heavily biased toward blood glucose readings. Perhaps this serves as a point of comparison between the two methodologies, to allow users to compare results.

Sensor Avg (mg/dL)

The table shown below is calculated in the same way “Sensor Avg” is one row five. This measure is clinically similar to Avg BG measure on row 1. In either case, it is a simple measure of blood glucose control. Because it is an average, hypoglycemic episodes can mathematically cancel out periods of hyperglycemia. Unfortunately, they do not cancel out damage caused by hyperglycemia.

To use this measure to understand diabetes management, you must use both the average and the standard deviation, to account for the variance in blood glucose levels. Diabetics who swing to extreme blood glucose levels quickly, such as Karen, have high standard deviations, even if the average is relatively close to 100.

As discussed previously, the Medtronic is unable to use the CGM data from Karen’s Dexcom CGM. Had the data been available, the Medtronic report would have shown a two week average sensor reading. While useful, a single two week average lacks context. I want to understand the stability of this average over time. Is an average of 187 mg/dL normal or abnormal for Karen? Fortunately, our Nightscout data set is now large enough to return multiple weekly averages, which can put single week averages into context. The following table includes four weekly sensor averages and we can see that Karen’s weekly averages and standard deviations are fairly consistent week to week.

1
2
3
4
5
6
7

entries %>%
        filter( !is.na(wk_lbl) ) %>%
        group_by( "Date Range" = wk_lbl) %>%
        summarise("Avg Sensor Reading" = paste(round(mean(sgv, na.rm=TRUE),1),"mg/dL"),
                  "Std Deviation" = paste( round(sd(sgv, na.rm=TRUE),1),"mg/dL")
                  ) %>%
        pander()

The blood glucose level for a non-diabetic is, on average, 100. Over the course of the day, a non-diabetic could be as high as 140 and as low as 70, although such extremes values would be short-lived. This shows us that even with insulin treatment, Karen’s averages are higher than a non-diabetic. The weekly averages shown above are a little higher than the report from February, but well within the range we would expect to see based on the variance seen across weeks. The high standard deviation demonstrates that Karen experiences a wide range of blood glucose levels. Although she does not meet the definition of a “brittle diabetic”, her blood glucose levels do swing dramatically as a result of her gastroparesis, a complication of diabetes.

Number of Sensor Readings

The Statistics table does not include the number of sensor readings. It only includes the number of manual blood glucose readings, which cannot be directly replicated using CGM data. It is surprising that the Medtronic report only includes the number and daily average of manual tests. Knowing the number of sensor readings is useful for understanding the consistency with which the CGM is worn, which is obviously important for assessing the accuracy and validity of the aggregate CGM data.

This measure is not really clinically similar to Measure 2. Manual blood glucose tests are important. They are used to calibrate the CGM sensor, which helps improve the accuracy of the CGM data we are going to rely on so much here. Reporting on the number of sensor readings is as similar as we can get, but it is not clinically similar. But, as a measure, the count and distribution of CGM sensor readings is worth discussing.

When the CGM is worn correctly and is operational, it records a new reading every five minutes. But, Murphy rules. Batteries die, sensors fall out and sensors lose their connection to the rig. This is an individualized “big” data set and it isn’t even genetic! In an ideal world, the sensor will record 2,016 readings every week. As you can see, the number of actual readings is usually below 2,000. When the number of readings falls unexpectedly, such as the week starting on July 19, it may indicate a problem with the CGM or how consistently Karen is using her CGM.

1
2
3
4
5

entries %>%
        filter( !is.na(wk_lbl) ) %>%
        group_by( "Date Range" = wk_lbl) %>%
        summarize("N Entries" = format( n(), big.mark="," ) ) %>%
        pander()

In this case, the low number of CGM readings is misleading. It is not indicative of a CGM error. Not is it telling us that Karen decided to stop managing her diabetes for a week. We were in Ireland that week on vacation. Contrary to what you see, Karen did use her CGM in Europe. There are gaps in the Nightscout data because we stopped using the rig for several days that week. We would like to thank Verizon for making it outrageously complicated and expensive to set up and monitor a data connection in Europe. THANK YOU!

Time Spent Where

Duration matters. The Statistics table includes several measures to approximate how much time is spent in different blood glucose value ranges. Karen’s endocrinologist wants to know how much time she spends in the “good” range, 70-140 mg/dL. This range of values damages her body less than either hyper or hypo glycemia. In the Medtronic report, these data elements are based on manual blood glucose tests, but we can create a clinically similar measure using Nightscout data. The following table shows the percentage of CGM readings that are in the target range, above it and below it. A high proportion of readings in the 70 - 140 mg/dL range is the desired outcome.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

entries %>%
        filter( !is.na(wk_lbl) ) %>%
        group_by( "Date Range" = wk_lbl) %>%
        summarise("Sensor Readings" = n(),
                  "Readings In Target Range " = paste( format(round(sum(ifelse(sgv >= 70 & sgv<= 140,1,0))/n()*100,2),
                                                              nsmall = 2),
                                                       "%", sep=""),
                  "Readings Above Target" = paste( format(round(sum(ifelse(sgv > 140,1,0))/n()*100,2),
                                                  nsmall = 2),
                                                  "%",sep=""),
                  "Readings Below Target" = paste( format(round(sum(ifelse(sgv < 70,1,0))/n()*100,2),
                                                   nsmall=2),
                                                   "%",sep="")
                  ) %>%
        pander()

Although mathematically crude, simple proportions such as these are similar to the AUC and it is easier for many patients and doctors to understand. By including the average sensor reading for readings in a given range we can better understand the interplay between duration and severity. More on that below. Yes, this is a long post. Sorry. Thanks for hanging in there.

Readings In Target Range

The first column, the measure denominator, is the total number of CGM readings. The numerator, the number of readings in the target range (70-140 mg/dL) is the second column. A diabetic wants as many readings as possible to fall in this range. The last column is the mean sensor value for readings in the target group. An average close to 100 mg/dL is good.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

entries %>%
        filter( !is.na(wk_lbl) ) %>%
        group_by( "Date Range" = wk_lbl) %>%
        summarise("CGM Readings" = n(),
                  "Readings In Target Range" = sum(ifelse(sgv >= 70 & sgv<= 140,1,0)),
                  "% In Target Range" = format(round(sum(ifelse(sgv >= 70 & sgv<= 140,1,0))/n()*100,2),
                                                      nsmall = 2
                                                      ),
                  "CGM AVG" = paste( round(mean( ifelse(sgv >= 70 & sgv <= 140, sgv, NA), na.rm=TRUE),2),
                                    "mg/dL",
                                    sep=" "
                                   )
                  ) %>%
        pander()

On average, only a quarter of Karen’s CGM readings fall within the target range. Anything we can do to increase the number of readings in the target range will improve Karen’s diabetes control and help prevent or limit complications. The raison d’être of this project is to help Karen take corrective actions sooner, which we believe will improve her blood glucose control. This is not a wild hypothesis, Continuous Glucose Monitors were developed because of this hypothesis and it worked. Research has been shown that Type 1 diabetics using CGM data have better blood glucose control. Using this same data to make person-specific predictions is the next step in developing a more comprehensive and timely management system.

Readings Above Target Range

The first column, the measure denominator, is the total number of CGM readings. The second column, the numerator, is the number of readings above 140 mg/dL. A diabetic wants to minimize the number of readings above the target range. The last column is the mean sensor value for readings above the target range. Time spent above 140 mg/dL is damaging to the body. At extreme values, hyperglycemia can result in diabetic ketoacidosis, coma and death. A hyperglycemic average close to 140 is better than a higher average. Reducing the number of readings about 140 is ALWAYS a good thing.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

entries %>%
        filter( !is.na(wk_lbl) ) %>%
        group_by( "Date Range" = wk_lbl) %>%
        summarise("CGM Readings" = n(),
                  "Readings Above Target" = sum(ifelse(sgv > 140,1,0)),
                  "% Above Target" = paste(format(round(sum(ifelse(sgv > 140,1,0))/n()*100,2),
                                                  nsmall = 2),
                                           "%"),
                  "CGM AVG" = paste( format(round(mean( ifelse(sgv > 140, sgv, NA), na.rm=TRUE),2),nsmall=2), " mg/dL")
                  ) %>%
        pander()

Karen is hyperglycemic more than desired. And when she is hyperglycemic, these episodes are moderately severe and tend to last several hours. If we can predict the onset of hyperglycemic episodes before they begin, we may be able to increase the amount of time she spends in the desired range by taking corrective action before she is hyperglycemic. Fingers are firmly crossed.

Readings Below Target Range

The first column, the measure denominator, is the total number of CGM readings. The second column, the numerator, is the number of readings below 70 mg/dL. A diabetic wants to minimize the number of readings below the target range. When below 70 mg/dL diabetics experience confusion, anxiety and other physiological and psychological symptoms. Severe lows can result in losing consciousness or diabetic comas. Thankfully, the latter is rare. The last column is the mean CGM value for readings below the target range. A hypoglycemic average close to 70 is better than a lower average.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

entries %>%
        filter( !is.na(wk_lbl) ) %>%
        group_by( "Date Range" = wk_lbl) %>%
        summarise("CGM Readings" = n(),
                  "Readings Below Target" = sum(ifelse(sgv < 70,1,0)),
                  "% Below Target" = format(round(sum(ifelse(sgv < 70,1,0))/n()*100,2),
                                                      nsmall = 2
                                                      ),
                  "BG AVG" = paste( format( round( mean( ifelse(sgv < 70, sgv, NA), na.rm=TRUE), 2),
                                            nsmall=2
                                          ),
                                    "mg/dL",
                                    sep=" "
                                  )
                  ) %>%
        pander()

Although the amount of time spent below 70 is small, Karen’s hypoglycemic episodes can be severe. Karen’s hyperglycemic episodes are often followed by a hard crash. These crashes can ruin her entire afternoon. At their worst, they can be dangerous. Limiting the severity of these attacks would improve her quality of life and safety. If we can reduce the number of hyperglycemic episodes, we will also reduce the number of these hard crashes. Diabetes is a vicious circle. But, improvements in one area can also result in improvements elsewhere.

Karen copes with these hypoglycemic episodes remarkably well. A blood glucose level of 40 mg/dL for a non-diabetic would be profoundly disabling. In the 10+ years I have known her, Karen has never lost consciousness due to a hypoglycemic episode. She is apparently made of adamantium.

Thanks

If you are reading this. Thanks! That was a long post, but I wanted to get it done. Now we can start doing fun things!