Sampling plan for in-process QC (medical devices)

Achilleas

Starting to get Involved
Hello,

In a typical manufacturing process, QC will inspect product samples at the start and at periodic intervals.
e.g. 10 samples at first-off and 5 samples every 2 hours.

Are there any best practices in determining the sampling scheme and Ac/Re thresholds?
Would standards intended for AQL based lot-by-lot inspection lot inspection (ISO2859-1 or C=0 by Squeglia) be applicable to determine the total sample size and the Ac/Re thresholds?
In case there are defects identified during the inspection passing the Ac/Re thresholds at the end of the production run, shall the whole batch be rejected or only the specific period that these defective products have been produced?

Thanks for your help!
 

Bev D

Heretical Statistician
Leader
Super Moderator
Lot by lot inspection (which the “AQL” based sampling plans are designed for) is intended to be a random sample taken from the whole lot. If teh number of defects found equals or exceeds teh reject number the entire lot is rejected and subject to 100% screening or other remediation activities.
What you describe is NOT random sampling; it is directed testing. The idea is that certain defects are not random but can occur at startup,material changes, equipment adjustment etc. The sampling frequency is based on the physics of the process and how the quality of the part can change. In this case small samples are appropriate as the defect rate is often quite high and is not random. It is statistically appropriate for a non-random defect system. It is also extremely pragmatic as it has small sizes and the number of ‘at risk’ parts is quite low and escapes are less likely than with the AQL/RQL system which is actually quite flawed in practice.

Do not conflate the two schemes. Purpose matters and physics matters.
 

Achilleas

Starting to get Involved
Thank you, I understand that the “AQL” based sampling plans are not appropriate in case of periodic in-process testing.
However, how could we statistically justify the number of samples taken per x hours?
Do you believe that an auditor would request such justification?
 

Bev D

Heretical Statistician
Leader
Super Moderator
As I said the in process sampling plan is based on the physics of the process: there are two factors. The first is that the defects that in process inspection is looking for are NOT RANDOM but specific and 100% (or nearer 100%). You only need 1 or two samples to see a 100% defect rate. that is certainly statistically true. The second reason is to see shifts, drifts and cycles (based on continuous data) that would demonstrate that the process is changing BEFORE you create defects. This is SPC and the justification for the small sample sizes is documented in any good SPC book or paper.
I have never seen an auditor ask for statistical justification of these types of inspections.

As a side note, the AQL insection tabels are not statistically based but since they have been ‘commonly accepted’ auditors tend to not question that when they should. :mad:
 

Mike S.

Happy to be Alive
Trusted Information Resource
As a side note, the AQL insection tabels are not statistically based but since they have been ‘commonly accepted’ auditors tend to not question that when they should. :mad:

I have had auditors ask questions about my sampling plans. While there are appropriate questions related to "shalls" in the applicable standard, we don't want auditors creating their own requirements.

I don't know what the OP's applicable standard requires, but AS9100 8.5.1.c.2 requires "ensuring that when sampling is used as a means of product acceptance, the sampling plan is justified on the basis of recognized statistical principles and appropriate for use (i.e., matching the sampling plan to
the criticality of the product and to the process capability)."
 

Achilleas

Starting to get Involved
I understand and agree with the SPC approach.

The first is that the defects that in-process inspection is looking for are NOT RANDOM but specific and 100% (or nearer 100%).

Let's take examples of defects: a wrong automated assembly of a cap to a bottle, or particle contamination on a produced part in a cleanroom. Both defects can be random; they may occur on a limited number of samples during production. So we need to determine an appropriate sampling size for in-process quality controls, that has an adequate probability to detect the defects for the desired quality level.
 

Bev D

Heretical Statistician
Leader
Super Moderator
Are you using inprocess inspection as the only acceptance inspection? or is it in addition to acceptance sampling inspection? This is critical.

Your last sentence is not necessarily true. For low rate random defects, there is no sample size except 100%. you should be using other process controls to prevent or detect defects. for example vison systems to catch crooked caps and filters, clothing and air showers, makeup and hair restrictions to prevent contamination in a clean room.
 

Achilleas

Starting to get Involved
Are you using inprocess inspection as the only acceptance inspection? or is it in addition to acceptance sampling inspection? This is critical.

Sorry, I should have mentioned this. We are using in-process inspection as the only acceptance inspection. One approach is making sure that the total samples inspected match at the minimum the samples indicated from an AQL based plan. For example, if the AQL-based sampling plan indicates 80 samples and the batch is produced in 10 hours, we should inspect a minimum of 8 samples per hour.

you should be using other process controls to prevent or detect defects. for example vison systems to catch crooked caps and filters, clothing and air showers, makeup and hair restrictions to prevent contamination in a clean room.

Sure, such controls are in place, the quality inspection is additional.
 
Are there any best practices in determining the sampling scheme and Ac/Re thresholds?
Yes. Your AQL should be determined based on risk, meaning, if the defect could cause patient injury, you will test more samples than if the defect would have minimal effect.

Would standards intended for AQL based lot-by-lot inspection lot inspection (ISO2859-1 or C=0 by Squeglia) be applicable to determine the total sample size and the Ac/Re thresholds?
For medical devices, C=0 is the usual method. Nobody wants to be in a court room explaining to the judge why they accepted a defect.

In case there are defects identified during the inspection passing the Ac/Re thresholds at the end of the production run, shall the whole batch be rejected or only the specific period that these defective products have been produced?
Is this a homogeneous batch? AQL sampling must be done on random samples from a homogeneous batch. This in-process inspection is strange to me. How can you determine the proper AQL when your lot size is in flux? I would consider this in-process monitoring (SPC as stated before). If you find a defect, you stop production and open a nonconforming product report. If you don't find a defect, you perform QC testing on an appropriate number of random samples from the lot after production of that lot is done.

Is this a destructive test? Do you know much about process validation?
 

Bev D

Heretical Statistician
Leader
Super Moderator
There are other problems in trying to combine 'AQL' sampling with in-process sampling beyond what indubioush states.
The reject number is based on the entirety of the sample for the lot. In-process sampling as you describe is intended to stop, fix and remediate as soon as any of the small samples find ONE defect. period. If applying SPC concepts you would stop and fix the process as soon as you detect an out of control condition even if no defects exist, effectively starting a new lot. What I - and I think others are saying - is that in-process inspection is great - it finds defects early and it reduces the probability of a 'defective' lot. But it is not lot acceptance testing. You need both. Not one mish-mash.

I will also now get on my small soap box and remind everyone that lot size has nothing to do with the effectiveness of a sample size (except for very small lots where sampling without replacement actually matters). Using lot size was a result of negotiations, not statistics. hence why the AQL tables are not statistically justifiable - they only accepted because, well we've always done it and mythology is more powerful than science for too many people.

I will also get on my bigger soap box and remind everyone that AQL stands for ACCEPTABLE quality level, it is the defect rate that you will ship most of the time when it is present. What most people actually want is the RQL or Rejectable Quality Level.
 
Top Bottom