Sampling Plan Generation for Medical Devices: Confidence, Reliability and AQL

Every medical device company has to answer the same awkward question: how many units do we need to test? Too few and the conclusion is not defensible; too many and you waste product, time and money. A good sampling plan is a statistical argument that links sample size to risk. This guide shows how to generate sampling plans for verification, validation and incoming or final inspection, with worked numbers you can check yourself.

Start with the question, not the formula

Before choosing a method, decide three things:

1. What is being decided? Accepting a production lot, or proving a design requirement is met?
2. What type of data? Attribute (pass or fail, count of defects) or variable (a measured value such as force or length)?
3. How severe is the failure? Link the confidence and reliability you need to the severity and probability ratings in your ISO 14971 risk file. Higher-severity harm justifies higher confidence and reliability.

Regulators expect a documented statistical rationale for sample sizes. The legacy FDA Quality System Regulation expressed this in 21 CFR 820.250, and ISO 13485 expects methods, including sampling, to be justified and documented. Whatever you choose, write the reasoning down.

Method 1: Confidence and reliability for attribute data

For design verification and process validation, the most defensible approach is usually to demonstrate that at least a stated proportion of units (reliability, R) conforms, with a stated confidence level (C). With zero failures allowed, the success-run formula gives the sample size:

n = ln(1 − C) / ln(R), rounded up.

Confidence	Reliability	Sample size (0 failures)
90%	90%	22
95%	90%	29
90%	95%	45
95%	95%	59
99%	95%	90
95%	99%	299
99%	99%	459

If you allow one failure, the required sample rises. For 95% confidence and 90% reliability, n is 46; for 95% confidence and 95% reliability, n is 93. These come from the binomial distribution. A plan that allows failures is more tolerant of bad luck but costs more units, so decide this in the protocol, not after the data arrives.

A common way to pick levels: use a risk-based table in your procedure, for example higher confidence and reliability for high-severity hazards and lower levels for low-severity ones. The levels must be justified by your own risk acceptance criteria rather than copied from another company.

Method 2: Variable data and tolerance intervals

When you measure a value, you can use fewer units than an attribute test by using the data’s distribution. Check normality, then use a one-sided or two-sided tolerance interval (or a k-factor approach) to show that a stated percentage of the population falls within specification at a stated confidence. If data is not normal, transform it or use a non-parametric method. Capability indices (Cpk, Ppk) are also used, but they need enough data to be meaningful and should be reported with confidence bounds.

Measurement system quality matters as much as the sample size. Run a gauge repeatability and reproducibility (GR&R) study first, because a noisy measurement can make a good process look bad.

Method 3: AQL sampling for lot acceptance

For incoming inspection or final lot release, standard tables such as ISO 2859-1 (equivalent to ANSI/ASQ Z1.4) are widely used. The steps are:

1. Determine the lot size and the inspection level (general level II is the usual default).
2. Read the sample size code letter from the table. For example, a lot of 3,201 to 10,000 units gives code letter L, and a sample of 200 on a normal inspection.
3. Choose the acceptable quality limit (AQL) for each defect class, such as critical, major and minor. Medical devices usually allow no critical defects.
4. Read the acceptance and rejection numbers from the single-sampling table for the code letter and AQL.
5. Apply switching rules: tightened inspection after repeated rejections, reduced inspection after sustained good performance if your procedure allows it.

Remember that AQL sampling controls the probability of accepting a lot with a given defect rate, and it does not prove that every unit conforms. For critical characteristics, do not rely on AQL alone. Use 100% inspection, a stronger plan or a validated process. ISO 3951 provides equivalent sampling plans for variable data. Always confirm the current edition and read the acceptance numbers from the standard’s own tables.

Documenting a sampling plan

• The population and what each sample unit represents (device, component, lot).
• Data type, method and formula or table used.
• Confidence, reliability or AQL levels, with a link to the risk rationale.
• The acceptance and failure criteria, and the action on a failure (investigate, retest rules, CAPA).
• How units are chosen: random or representative across lots, operators, machines and time.
• A statement of who approved the plan and when, before testing starts.

Common mistakes

• Choosing a sample size first (for example 30) and justifying it afterwards.
• Re-testing failures until a pass appears. Define the retest rules in advance.
• Using one sample from a single lot to represent variation across lots.
• Ignoring destructive-test cost in planning, then cutting the sample under pressure.
• Applying a plan meant for lot acceptance to a design verification claim.

Sampling underpins design verification and process validation, so get it right once and reuse the approach in your procedures.

Frequently asked questions

Is a sample size of 30 acceptable?

Only if you can justify it. Thirty is not a regulatory requirement. A rationale tied to confidence and reliability, or to a tolerance interval, is more defensible.

What confidence and reliability should I use?

Use levels linked to the severity of harm in your risk management file. Higher severity usually calls for higher confidence and reliability. Define them in a procedure so teams apply them consistently.

What does AQL mean?

AQL is the acceptable quality limit, the worst defect rate that is considered acceptable as a process average for lot acceptance sampling. It is not a guarantee about individual units.

Can I use fewer samples with variable data?

Often yes, because each measurement carries more information than a pass or fail. This holds only when the measurement system is capable and the distribution assumptions are checked.

Statistical methods should be reviewed by a qualified person for your specific application.

Talk to the AptSkill team

Want a template or training on sampling plans for your product? Tell us what you are validating.