Blog: MSA on Glucose Measurement

Introduction

This blog describes how I improved the gauge capability my blood glucose monitor to enable me to spot any changes in my fasting glucose measurement. I would like to remove the Metformin from my drug regime. Improving the gauge capability would allow me to rapidly detect any changes to my fasting glucose measurement.

I show how continually improving the measurement process can improve the gauge capability even if the gauge capability has yet to achieve best practice of <10%.

I am type two diabetic. When I was first diagnosed, I used Metformin and Gliclazide to control my blood glucose. I had a simple blood glucose monitor based on a finger prick to draw a drop of blood and a test strip and meter to make the measurement. It was not possible trend individual data points, but I could keep an eye on seven-day, thirty-day and ninety-day averages.

Over a period, I noticed the fasting readings had shifted from between 7 and 8 mm/L to between 8 and 10 mm/L. Backed up by an increase in HbA1c data my physician decided we should add Dapaglifozin to my medication. Apparently, it is not unusual for the body to become insensitive to Metformin and to add Dapaglifozin to the regime. At the time of this change there was an intention to review the data and if possible, to stop taking the Metformin.

On reading that Metformin is a serious pollutant that is not removed by wastewater treatment plants I decided to use the data extraction facility on my new glucose meter to trend out my fasting glucose check. This is the check I do each morning before eating anything. From now on I will call this my check.

I decided that it was fair to call this check a “Critical Parameter”

Methodology

Calibration is not currently possible as I do not have calibration samples.

First let’s establish the “Precision”.

Although the specification limits on the monitor are from 3.5 to 10 mm/L I prefer to use 5 to 10 mm/L as the advice for driving is “5 to drive”. From my few experiences when my “sugars have dropped” a bit this seems sensible. Below 5 mm/L I can sense a problem.

The resolution of the meter is .1 mm/L. Using the “Rule of 10”  5/.1=50. 50 being much greater than 10 indicates that the meter could well be capable for monitoring my glucose.

By reading the instructions and a bit of brainstorming we could generate a list of possible sources of variation.

Finger cleanliness, test strip age, test strip storage, blood spot size, where the blood is taken from, lancet age.

In industry we see 3 methods for analyzing measurement variation: GRR Type1, 2 and 3 study

Type 1 study means we take 25 measurements of 1 product. This method is not suitable in this case. It is not possible to take 25 measurements of 1 blood sample.

Type 2 study means at least 2 appraisers and 2 to 3 measurements per appraiser which is also not suitable.

So most suitable MSA study is the type 3 study where we take 25 drops, and each drop is measured 2 times

To perform a Type 3 gauge study, we started with taking two measurements from the same drop of blood on one finger.

Results

We have taken 25 days of measurements. The table of the measurements is given in table 1.

Table 1: Data from Type 3 Study

The first 20 days 1 drop was taken with 2 strips. The drops were taken from the left index finger. After 20 days 2 drops were taken from the left middle finger with 2 strips.

The control chart for the 25 measurements looks like this.

Figure 1: Control chart Type 3 Glucose study

From subgroup 16 we see a clear increase of the measurement variation and we see an OOC in subgroup 24. For subgroup 24 there was no obvious reason why the range was OOC so an extra measurement was taken and that measurement was 6.9, so it seems that 8 was an outlier. For first analysis we did not remove this one because there was no obvious and clear reason.

Normally we would try to eliminate out of controls on the range chart and repeat the study before we go further with our analysis but for this example we will continue the study.

In case of a good MSA result we would expect that the average chart would show a lot of out of controls. The fact that there are hardly OOC on the average chart is already an indication that the measurement system is not accurate enough.

Figure 2: GRR results compared to process variation

If we look at the results it is clear that the measurement variation is not acceptable compared to the process variation. It should be at least less than 30% and ndc should be 5 or higher.

We can also show the GRR compared to the tolerance. If we check the variation compared to a tolerance from 5 to 10 we get the following results. 

Figure 3: GRR Results compared to tolerance

We see that this percentage is also larger than 30% and ndc = 1 so also not acceptable. 

Analysis and Next Steps:

Looking at the results we expect that the biggest problem is that the second measurement (strip) from 1 drop is not always gets enough blood in the required time to get the same results.

This proved to be problematic. Sometimes I noticed that with the second strip sometimes the strip did not fill up completely and this always resulted in a higher measurements. So we moved to 2 drops of blood at the same time on the same finger. Any difference in these could be considered as contributing to the repeatability of the measurement.

• Test strip storage and age is as per the specifications.
• Hands were washed immediately before the test.
• Blood droplet size was always plenty.
• The blood is always taken from the same finger at the same time.

In future I will use lancets inside their use by date.

Initially we still had some variation in measurement up to 1.5 mm/L

We then moved to “Hospital best practice”  and cleaned the finger with an antiseptic wipe before each blood drip was made. This reduced the variation to a maximum of .6mm/L.

The results of a new study with this new practice you see below:

Figure 4: Type 3 Study Glucose with improvement method

This chart already looks a lot better. We see no ranges out of control which is good and as expected we see a lot more out of controls on the average chart indicating the measurement variation is a lot smaller than the process variation.

 In figure 5 you see the GRR results.

Figure 5: GRR Results Type 3 study

Although the results are a lot better, they are still not acceptable.

If we compare it to the tolerance the GRR is 24.8% which is marginal.

For the analysis purpose the only way we can reduce measurement variation quickly is to take 2 measurements and store the average in the control chart to monitor my glucose. This will reduce measurement variation with sqrt(2). In that case the GRR will become 24.1% instead of 34.2%, which is still not great but for analysis purposes in the short term there is at least a big improvement.

The MSA analysis we ran is typically for an MSA study in industry. The results found in the first study are often not acceptable and you need to analyze the root cause and improve the method. You often see quick improvements, but getting acceptable results is with GRR< 10% and ndc > 5 are not always easy but it is important otherwise all analysis done is based on unreliable results.

The MSA study confirmed my first guess that the measurement I performed was not acceptable. I have improved the measurement process by improving the cleanliness of my finger and ensuring the strip fills with blood correctly. Still further improvement is required. The documented literature indicates a GRR of less than 10% should be possible. The next step will be performing a GRR with calibrated samples.

Steve Murphy

26 October 2025