GLU is described in the paper: L.A.C. Millard et al. GLU: A tool for analysing continuously measured glucose in epidemiology, bioRxiv, 2018
GLU takes continuous glucose monitoring (CGM) data as input, and derives a set of characteristics describing these data:
- Median absolute deviation (MAD)
- Area under the curve (AUC)
- Proportion of time spent in low, normal and high values
- Standardised glycaemic variability percentage (sGVP)
- Fasting proxy
- Meal statistics:
- Time to peak
- 1-hour postprandial AUC
- 2-hour postprandial AUC
- Medication statistics
- 1-hour postprandial AUC
- 2-hour postprandial AUC
- Exercise statistics
- 1-hour postprandial AUC
- 2-hour postprandial AUC
GLU supports CGM data from three devices specifically:
- Medtronic iPro2
- Abbott Freestyle libre
- Dexcom G2
- A device non-specific generic format (see below)
This project is licensed under The MIT License.
The package can be installed within R directly from GitHub.
To do this you will need devtools:
install.packages('devtools')
After installing devtools you will be able to install GLU:
library(devtools)
install_github("MRCIEU/GLU")
To update GLU to the latest version, rerun the install_github command.
There are two main functions in GLU: runGLUForFiles and runGLUForDirectory, that run GLU for either a specific set of files, or for all files in a particular directory.
Both these functions take an indir compulsory argument that specifies the directory containing the CGM data files.
The runGLUForFiles has an additional files compulsory argument, for providing the names of files that should be processed.
The compulsory and optional arguments are as follows.
| Arg | Description |
|---|---|
| indir | Directory where processed CGM data files (i.e. output from step 1) are stored. |
| files | List of file names to be processed by GLU, where these files reside in the indir directory (required only for the runGLUForFiles function). |
| Arg | Description |
|---|---|
| pregnancy | Set to TRUE to use pregnancy thresholds for proportion of time spent in low, normal and high summary variables. |
| diabetes | Set to TRUE to use diabetes thresholds for proportion of time spent in low, normal and high summary variables. |
| outdir | Directory where derived CGM characteristics data files should be stored. Default is indir. |
| nightstart | Time night period starts (HH:MM). Default is 23:00. |
| daystart | Time day period starts (HH:MM). Default is 06:30. |
| dayPeriodStartTime | Day period start time (HH:MM). Default is nightstart. |
| firstvalid | Use hour and minutes of the first valid glucose time points as the day period start time. |
| save | Set to TRUE to save derived (resampled) CGM data. |
| saveevents | Save summary variables of each individual event, rather than just the averages for each day and across days. |
| freq | Frequency between time points in CGM data (minutes). Default is 5 minutes. |
| imputeApproximal | Use approximal imputation method (see paper). |
| imputeOther | Use other day imputation method (see paper). |
| hypothreshold | Threshold between hypoglycaemia and euglycemia, to overide the defaults |
| hyperthreshold | Threshold between euglycemia and hyperglycaemia, to overide the defaults |
| timeformat | Time format in CGM data. Default='%d/%m/%y %H:%M:%S' (format for Medtronic ipro2 data). |
| outlierthreshold | Value k used for outlier detection threshold d=k*SD. |
| device | Medtronic iPro2=0; Dexcom G2=1; Abbott Freestyle Libre=2; other(general format as above)=3. |
For example, to run GLU for all files in a directory, where the CGM data is from Abbott Freestyle Libre, that records glucose values every 15 minutes, we can you a command like this one:
library('GLU')
runGLUForDirectory(indir='/path/to/in/dir/', outdir='/path/to/out/dir', device=2, freq=15)
To then use approximal imputation rather than the 'complete days' approach, we can add the impute argument:
library('GLU')
runGLUForDirectory(indir='/path/to/in/dir/', outdir='/path/to/out/dir', device=2, freq=15, impute=TRUE)
If you would like to run GLU for all files in a directory then use the runGLUForDirectory. The arguments are the same as above but without the files arguments.
Example commands that run GLU on example data are provided in Section 7 below.
If using GLU with CGM data from other devices, you can convert your data to a CSV (comma seperated values) file with the following columns, and then specify the device as 'other' using the device argument.
Required columns:
| Column | Description |
|---|---|
| time | Required in format 'dd/mm/yy HH:MM:SS'. |
| sgReading | Sensor glucose levels. |
Optional columns:
| Column | Description |
|---|---|
| meal | Optional column, which if not blank indicates a meal timepoint. |
| exercise | Optional column, which if not blank indicates a exercise timepoint. |
| medication | Optional column, which if not blank indicates a medication timepoint. |
| bgReading | Blood glucose levels. |
NB: If you are using a common format not currently supported by GLU, you can send us some example data and we can add preprocessing for this format to the GLU tool.
We remove all rows without either a sensor glucose value or an event (meal/exercise/medication). We resample to 1 minute epochs.
- Outlier detection
Outlier detection is performed as follows:
- Calculate the difference of each SG value from the previous and next values in the SG trace.
- Calculate the SD of the distribution of these differences
- Identify time points where these differences are > 4SD from the mean difference; these are potential outliers that should be checked.
- Dealing with missing data
The `freq' argument can be used to specify the epoch frequency (interval between adjacent timepoints) in the CGM data.
A day is included if there are `1440/freq' SG values, as this means that there are no missing SG values. The derived variables below are only calculated using the included days.
The impute' argument can be used to impute missing blocks in the CGM data, using an approximal' imputation approach (see GLU paper for more information).
- Manual review
We recommend that researchers visually review each participant's data.
GLU outputs CGM trace plots and Poincare plots to do this, and examples can be found in the examples subdirectory.
- Median absolute deviation (MAD)
The median of the absolute difference of each sensor glucose value from the median sensor glucose value.
We calculate this for each day and then calculate the mean across all days, and across all nighttime and daytime periods, respectively.
- AUC, per minute for each valid day, and mean across all valid days
- The AUC of the SG values
- Trapezoid method (linear interpolation between the discrete glucose measurements) as described here.
We calculate this for each day and then calculate the mean across all days, and across all nighttime and daytime periods, respectively.
- Proportion of time spent in low, normal and high SG ranges, per valid day and mean across all valid days
Different thresholds are appropriate for pregnant women versus the general population.
Pregnancy thresholds: low<3.5 mmol/l, 3.5<=normal<=7.8 mmol/l, high >7.8 mmol/l. General population thresholds: hypoglycemia<3.9 mmol/l, 3.9<=normal<10.0 mmol/l, hyperglycemia>= 10.0 mmol/l.
To calculate these proportions we linearly interpolate between adjacent SG values to calculate the proportion of time within each range.
The following pseudocode (for pregnancy thresholds) demonstrates the process we use.
lowT=3.5 # threshold value between low and normal ranges
highT=7.8 # threshold value between normal and high ranges
INIT lowTime=0; normalTime=0; highTime=0;
SGmin=min(SG(i), SG(i+1))
SGmax=max(SG(i), SG(i+1))
# both SGmin and SGmax are in same range (either low, normal or high)
IF SGmin<lowT & SGmax<lowT THEN lowTime+=T(i+1)-T(i); # both in low range
ELSE IF SGmin>=lowT & SGmin<=highT & ISIGmax>=lowT & SGmax<=7.8 THEN normalTime+= T(i+1)-T(i); # both in normal range
ELSE IF SGmin>highT & SGmax>highT THEN highTime+= T(i+1)-T(i); # both in high range
ELSE IF SGmin<lowT THEN # SGmin value is in low range and SGmax is in either mid or high range
lowTime+=(T(i+1)-T(i))*(lowT-SGmin)/(SGmax-SGmin)
IF SGmax<=highT THEN
normalTime+=(T(i+1)-T(i))*(SGmax-lowT)/(SGmax-SGmin)
ELSE
normalTime+=(T(i+1)-T(i))*(highT-lowT)/(SGmax-SGmin)
highTime+=(T(i+1)-T(i))*(SGmax-highT)/(SGmax-SGmin)
ELSE IF SGmin<=highT THEN # ISIG values are in normal and high ranges resp.
normalTime+=(T(i+1)-T(i))*(highT-SGmin)/(SGmax-SGmin)
highTime+=(T(i+1)-T(i))*(SGmax-highT)/(SGmax-SGmin)
RETURN lowTime, normalTime, highTime
where i is the index along the SG sequence, T(i) is the actual time at position i.
We calculate this for each day and then calculate the mean across all days, and across all night-time and day-time periods, respectively.
- Standardised glycaemic variability percentage (sGVP), per included day and mean across all included days
Intuitively, sGVP is based on the length of the line of the glucose trace. GVP is the length of the line relative to the length of a completely flat glucose trace. A trace with no variability has GVP=0, and the more variability (both in amplitude and undulations) from a flat trace, the higher the GVP value. The sGVP measure is the GVP calculated using a standardised glucose trace, to reflect the degree a trace undulates. Standardisation is performed as:
(SG - median(SG))/MAD
We calculate this for each day and then calculate the mean across all days, and across all night-time and day-time periods, respectively.
- Nocturnal Glucose
The lowest mean 30 minute consecutive period during the night-time.
- Meal time to peak
The number of minutes from the meal to the next peak SG.
In the simple case the peak is the nearest subsequent time t, after the meal where SG(t-1) < SG(t) > SG(t+1).
In the case where the peak is on a plateau, then (i.e. there are multiple timepoints on the peak with the same peak value) then we define the peak time point as the nearest time point to the meal on this plateau.
- Post-event AUC
An event is either eating, exercise or medication.
Post-event AUC can be either 1-hour or 2-hour. E.g. 1-hr postprandial AUC, 1-hr post-exercise AUC, 1-hr post-medication AUC etc.
The n-hr post-event AUC is the mean of the SG values during the 15 minute period occuring n hrs after the event.
To try the tool with example data use the following commands.
Run GLU for all files in a directory:
library('GLU')
datadir=system.file("extdata", package = "GLU")
runGLUForDirectory(indir=paste0(datadir,'/medtronic-ipro2/original/'), outdir=paste0(datadir,'/medtronic-ipro2/derived/'), save=TRUE, saveevents=TRUE)
Run GLU for a specific file by specifying this file:
library('GLU')
datadir=system.file("extdata", package = "GLU")
runGLUForFiles(files='data_export-999998.csv', indir=paste0(datadir,'/medtronic-ipro2/original/'), outdir=paste0(datadir,'/medtronic-ipro2/derived/'))
library('GLU')
datadir=system.file("extdata", package = "GLU")
runGLUForDirectory(indir=paste0(datadir, '/freestyle-libre/original/'), outdir=paste0(datadir,'/freestyle-libre/derived/'), device=2, imputeApproximal=TRUE, freq=15, nightstart='00:00')
library('GLU')
datadir=system.file("extdata", package = "GLU")
runGLUForDirectory(indir=paste0(datadir,'/dexcom-g6/original/'), outdir=paste0(datadir,'/dexcom-g6/derived/'), device=1, nightstart='16:00', imputeOther=TRUE)
GLU can also be directly from the command line, using the run.R script.
For example, to process both example Medtronic Ipro2 data files:
Rscript run.R --indir="/path/to/in/dir/" --outdir="/path/to/out/dir/"To process a specific CGM data file, specify the filename argument:
Rscript run.R --indir="/path/to/in/dir/" --outdir="/path/to/out/dir/" --filename="infile.csv"See all the arguments available by calling run.R with no arguments:
Rscript run.R