What are the acceptance criteria for blend and content uniformity now that FDA has withdrawn its 2003 draft guidance?

1What the acceptance criteria are now.

No regulation or current FDA guidance sets a numeric blend uniformity limit. 21 CFR 211.110(a)(3) requires in-process controls for “adequacy of mixing to assure uniformity and homogeneity”, 211.160(b) requires scientifically sound sampling plans, and 211.165(d) requires acceptance criteria adequate to assure that batches meet “appropriate statistical quality control criteria” before release.¹ Each firm sets its own criteria, statistically, before the batches run.

FDA has withdrawn two blend drafts. Its August 1999 draft on blend uniformity analysis for ANDAs was withdrawn in May 2002, with FDA noting that “An evaluation of uniformity of a blend may be necessary to fulfill such requirements.”² The draft on powder blends and stratified in-process dosage unit sampling, announced on 7 November 2003,³ was withdrawn on 7 August 2013 in a review of draft guidances issued before 2010.⁴

FDA’s questions and answers give two reasons. The draft let only one of three blend replicates per location be assayed, whereas “The Agency currently recommends that all replicate samples taken from various locations in the blender be evaluated to perform a statistically valid analysis.” And its routine batch criteria rested on USP <905>, which is “not a statistical sampling plan”, so “FDA no longer supports their use for batch release”.⁵

2The ISPE alternative approach.

After the withdrawal, an ISPE technical team developed an alternative approach, published in 2014, with acceptance tables built on ASTM E2709 and E2810, the standards FDA’s questions and answers point to for stratified sampling.⁹⁷¹⁰⁵

The route after a failed Stage 1 matters. An SD of 3.1 to 5.0% across all replicates still accepts the blend, but the dosage units must then be assayed at all 40 locations. Above 5.0%, a variance component analysis and investigation decide: a product or process cause means the blend is not uniform, while an analytical or sampling cause sends the decision to dosage units at 40 locations.⁷ ISPE’s Stage 1 assays one replicate per location, which FDA no longer favours; assaying all replicates from the start meets both, as the example below does.⁵

ISPE’s limits are standard deviations as a percentage of target. The 2003 draft and CPG 460.600 used the relative standard deviation (RSD), the SD divided by the mean.⁶¹³ Near 100% the two barely differ; at a mean of 95% of target, an SD of 3.0 is an RSD of 3.0 ÷ 95 = 3.16%, so state which one each limit uses.

All the ISPE tables give at least a 95% probability that a sample from the batch passes the USP test, at 90% confidence for qualification and 50% confidence for routine batches.¹⁰ ISPE explains 50/95 as “50% confidence of having at least 95% probability of passing the USP test on a future/alternate sample from the batch”, and says its routine criteria “may have reduced assurance”.¹⁴⁸

3Protocol checklist.

1. Name the blend locations, at least 10, on a drawing of the blender, including worst-case positions and more than one depth, and take at least 3 replicates from each.⁷¹⁵¹⁶
2. Justify the thief, the number and size of samples and their preparation for analysis, and show the thief’s reliability during method development.⁵¹⁶
3. Assay every replicate and plan a variance component analysis that separates within- and between-location variability.⁵
4. Fix the statistical method and acceptance criteria before the run, with the confidence level set by the attribute’s risk.¹⁷¹⁵
5. Sample dosage units across the whole run, stratified by location or by run time, with at least 3 units from each location.⁷¹⁵
6. If in-process dosage units stand in for blend samples, justify it by comparing blend and dosage unit results, “weight corrected if necessary”. A tablet of a 20 mg in 100 mg product that weighs 98 mg and assays 19.4 mg is 97.0% of label claim, but 19.4 ÷ (98 × 0.20) = 99.0% weight corrected. Release results stay uncorrected.¹⁴
7. Write the route for a blend SD above 3.0%, following ISPE’s Stage 2 above.⁷
8. Say how an individual result outside its limit is handled: never averaged away or dropped by an outlier test, but investigated as below.¹⁸
9. State the routine sampling plan that follows qualification, and what ends any period of enhanced testing.¹⁴

4How to calculate the acceptance value.

USP <905>, Ph. Eur. 2.9.40 and JP 6.02 share harmonised acceptance criteria and, under ICH Q4B Annex 6, are interchangeable subject to conditions, including one FDA exception covered below.¹² The acceptance value is:

AV = |M − X̄| + ks

Test 10 units. If AV ≤ 15.0 the batch passes. If not, test 20 more: the 30 pass if AV ≤ 15.0 and no unit is below 0.75 M or above 1.25 M.¹¹ For tablets and hard capsules, content uniformity is required below 25 mg or 25% of the unit; at or above both, uncoated and film-coated tablets and hard capsules may use mass variation instead.¹¹ FDA “finds unsuitable for regulatory purposes” the not more than 2% RSD exception to that threshold in the JP and Ph. Eur., so below 25 mg or 25% the US test is always content uniformity.¹²

Passing <905> remains the compendial release test, but FDA says criteria built on it give “only limited statistical assurance” that batches meet specifications and statistical quality control criteria.⁵ That is why the PPQ and routine plans above sit beside it. For products outside the USP, FDA’s compliance policy guide 460.600 (1987, still listed as final) requires content uniformity specifications, with individual unit assays and an RSD limit, for tablets and capsules with less than 50 mg of any active ingredient.¹³

5Example: PPQ blend, tablets and release.

The product and data are invented. PPQ batch 2 of a 5 mg tablet was sampled with a thief at 10 locations in the bin blender, three replicates each, as % of target.

• Replicate 1 alone has a mean of 99.81% and an SD of 1.22%, which passes ISPE Stage 1 at ≤ 3.0%.
• All 30 results, as FDA recommends, have a mean of 99.99%, an SD of 1.09% and a range of 97.9 to 102.0%.
• A variance component analysis of the 30 results gives a within-location SD of 0.70% and a between-location SD of 0.86%. The location effect is statistically significant (F = 5.58 against a 5% critical value of 2.39), though small, and only the replicates could show it.
• The tablets, sampled as 3 units at each of 20 locations, have a mean of 99.6%, a pooled within-location SD of 1.4% and an SD of location means of 0.8%, with every unit between 96.1 and 103.2%. ISPE’s 20 × 3 table at 90% confidence gives limits of 88.2 to 111.8% for that combination, so the batch passes.¹⁹
• The tablets show no location effect, since noise alone would give an SD of location means of 1.4 ÷ √3 = 0.81, so the report records that the small blend location effect did not carry into the dosage units.
• At release, 10 tablets have a mean of 99.86% and an SD of 1.208%. The mean is between 98.5 and 101.5, so M = X̄ and AV = 2.4 × 1.208 = 2.90, well inside 15.0. Had the mean been 97.0% with the same SD, M would be 98.5 and AV = 1.5 + 2.90 = 4.40.
• Had those 10 results come from a routine stratified sample, ISPE’s 50% confidence table would allow an SD of up to 4.799% for 10 units with a mean of 99.9%, so 1.21% passes with room to spare.²⁰

6When a blend location fails or looks odd.

Treat each result as a result. FDA’s OOS guidance says “the use of averages is inappropriate when performing powder blend/mixture uniformity or dosage form content uniformity determinations”, because the test measures variability, and that outlier tests “have no applicability in cases where the variability in the product is what is being assessed, such as for content uniformity”.¹⁸ So a high location is not averaged with its replicates or discarded; it is investigated like any out-of-specification result. If the investigation finds the sampling method “inherently inadequate”, FDA says a new one “must be developed, documented, and reviewed and approved by the QU”.¹⁸

A variance component analysis of blend and dosage unit results then shows where the variability sits.

FDA accepts that a thief can disturb the powder bed and segregate it, but says thief sampling “remains widely used and provides reliable results in many cases”. Its suggested alternatives are process analytical technology, with each probe’s effective sample size and position justified, and statistical process control of the blending.⁵

Switching to dosage units is not an escape route. ISPE expects “a good faith effort to assess BU”, except where the blend could put operators at risk, and records that FDA inspectors found firms “immediately defaulting to IPDUs” (in-process dosage units) without one, one reason the 2003 draft was withdrawn.¹⁴ Keep the record of what blend sampling was tried first.

Continuous manufacturing moves the test. FDA’s January 2025 draft guidance on 211.110 says batch manufacturing should “typically” assess blend uniformity “before in-process materials continue to the compression step”, while a continuous line could test at the tablet press feed frame or after compression, and “control strategies that rely solely on current process models would be insufficient to satisfy the requirements of § 211.110.”²¹

7What inspectors cite.

FDA cites the sampling plan as often as the result. EyePoint Pharmaceuticals had qualified the current process for its YUTIQ implant on a single batch. FDA’s letter of 12 July 2024 said:¹⁶

“This did not scientifically demonstrate blend uniformity due to a sampling plan that failed to take into account the depth of the mixing bowl, and lacked justification for the number of samples collected, the size of each sample, and the preparation of the samples for analysis.”

Glenmark rejected 14 batches of low-dose desmopressin acetate tablets from 2018 to 2021 for out-of-specification and content uniformity results. FDA’s letter of 22 November 2022 records that, after the inspection, the firm:²²

“discovered a failing stratified sample assay result for desmopressin acetate 0.2mg tablets batch 20220121 and initiated a voluntary recall of this batch on June 10, 2022.”

The requirement is not limited to tablets. CTK OTC Laboratories, which makes topical creams including sunscreens, was told on 19 May 2025 that “you were unable to provide blend uniformity studies to demonstrate your mixing processes assure uniformity and homogeneity.”²³

8How the EU, UK and PIC/S treat it.

In the EU, Ph. Eur. monographs, including 2.9.40, have mandatory applicability.¹² EMA’s process validation guideline lists “unit dose products containing drugs in low content (≤2% of composition)” among specialised dose forms that could be non-standard, for which production-scale validation data, normally on at least three batches, go in the marketing authorisation dossier unless otherwise justified. That is the low-dose territory where blend uniformity is hardest. The guideline names near-infrared measurement of blend homogeneity as an enabler of continuous process verification, without setting criteria.²⁴ MHRA points UK manufacturers to Volume 4 of the EU rules and its own Orange Guide.²⁵

PIC/S PI 006-4, in force since 1 October 2026, asks for both intra- and inter-batch variability to be considered, and for “the sampling method (e.g. worst-case timepoints, locations, or stratified such as machine run time-related), the number of samples taken, the way of processing the obtained data and the acceptance criteria (e.g. a minimum Cpk)” to be justified in advance. It also expects extra samples during qualification, such as uniformity tests.¹⁵

9How it runs in Seal.

In Seal, a PPQ protocol holds its sampling rationale, acceptance criteria and planned analysis, and each requirement is reviewed against the full evidence set, so a missed sampling point stays an open protocol exception until it is assessed. In the oral solid dose batch record, blend uniformity is a check before compression, created from the process definition, and the batch step knows whether a result permits continuation or triggers a hold.

Neil can prepare a capability review of process measurements such as uniformity results, keeping batches before and after a process change apart before any capability index is calculated. A qualified reviewer chooses the statistical approach.