Sample Size Calculation (Confidence Interval and Reliability) - Medical Devices

[Harini17]

Dear Statisticians and regulatory experts,

Kindly help me to solve this problem for one of the medical device project. I asked many statisticians, but none were able to give convincing answer from the regulatory standpoint as you know FDA is very strict with the evidences and rationals. I found the similar problem on this forum but couldn't find the end answer.

Problem: The manufacturer of medical endoscope device equipment (pretty expensive prototypes) and they are trying to figure out justification for using 5, 7 or 10 samples with sufficient confidence level and reliability to proceed w/ manufacturing.
Testing such as bend cycling and environmental testing etc which they can do in house, for these tests sample sizes need to be determined (for each test). *These sample sizes can be calculated based on confidence and reliability levels, but a precedent should be set for the confidence/reliability that is needed for these devices and testing based on the risk level of the testing and the expense of running these tests.

At least with the examples of similar kind with formula and method!

I have very limited statistical knowledge. Kindly help me to solve this problem with the solutions from the regulatory standpoint.

Thank you!

 

[paulag]

http://www.fdanews.com/ext/resource...mple-Size-for-Verification-and-Validation.pdf

 

[Harini17]

Thank you Paulag.
I would like to know the reliability and confidence interval percentages that we can to use for these tests, so that FDA accepts. I would like to know the FDA requirements for these tests.

Thank you!

 

[Bev D]

I am afraid you are asking for something that cannot be answered to your satisfaction. Sample sizes that will 'satisfy the FDA' are not easy or straightforward. You cannot simply plug numbers into a formula. We can give you the formula (I believe that it is in the attachment in the above post) but YOU must come up with the numbers that go into the formula based on your organizations assessment of the risk (hazard) of each characteristic that you will be testing.


The formula that incorporates Reliability with a specified confidence level is:

n = LN(1-C)/LN(R)

n= sample size
LN is the natural log
1-C is 1- confidence level (traditionally this is 95% or 99%. a higher confidence level is selected if the hazard is high)
R is the 'Reliability'. This is 1-defect rate that you can tolerate. again a severe hazard would require a pretty low defect rate

This formula requires that NO DEFECTS be found in the sample in order to accept the validation results.

This formula is based on using only pass/fail categorical results. you can get smaller sample sizes if you have continuous data where you can make predictions regarding the distribution vs. specification limits. These typically require a statistician or at least someone who is well versed in these types of statistics.

Severe hazards require large sample sizes per this Reliability/Confidence formula - although this typically results in the lowest sample sizes possible. Often companies will balk at high sample sizes and so the FDA provides the ability to provide a rationale based on balancing cost of testing vs. cost of the hazard. you can plug various numbers into the formula and see the effect for yourself.

Beyond this you have the complication of understanding that sample size applies to both the number of instruments/products and the number of USES of the product. The reliability/confidence formula typically applies to the number of instruments and the number of uses is based on the either the specified number of maximum uses or the maximum expected number of uses.

 

[Statistical Steven]

http://www.fdanews.com/ext/resource...mple-Size-for-Verification-and-Validation.pdf

Thank you for sharing my presentation. I have updated it slightly, but the ideas are the same. There is no "Standard" confidence and reliability. You need to tie it back to your risk management program so that your sample sizes (confidence and reliability) are commensurate with the risk to the patient (reliability component).

 

[Harini17]

Thank you, Bev D and Statistical Steven...I have read all your threads addressing these questions thank so much. You both are doing good work in solving these regulatory obstacles. Special thanks to Bev D for such detailed explanation!

 

[v9991]

Often companies will balk at high sample sizes and so the FDA provides the ability to provide a rationale based on balancing cost of testing vs. cost of the hazard. you can plug various numbers into the formula and see the effect for yourself.

while FDA is providing enough direction, there is not even an outline/framework or guidance from agency for same. got certain references on risk from PDA & ISPE ; but certainly not ones which reference cost of testvshazard.!

1. point is, agency has ultimate power (veto to say the system/practice is inadequate, non-compliant) but what it says is ...all theory, no reference/ example, guidance, framework; (I remember an section on risk benefit analysis on clinical trial data evaluation!., but no such thing on validations/sample size.)

2. is this one reason, why (especially) pharma is way behind in adopting these tools and techniques. (its only recently that 'statistical' rationale/analysis gaining releac env

 

[MeganG]

Thanks all for this, which has been incredibly informative for determining sample size.

In Paulag's attachment, the example for continuous data states "Unknown Standard Deviation and Delta, but want to detect a 0.5 standard deviation shift." How would one decide the shift that is necessary to detect?

 

[Bev D]

What is the difference you can tolerate? That is how you determine it.

 

[MeganG]

Thanks Bev D. I'm curious what would determine when to use the "sample size for continuous data" formula vs. the k-value formula.

Assuming the same specification as the example, for 95% confidence and 95% reliability, I would get a sample size 52 using the continuous data formula vs sample size 21 using the k-value formula.

Is it only existing information (mean, standard deviation) that justifies using the lower sample size?