Chris Burgess, PhD, is an analytical scientist at Burgess Analytical Consultancy Limited, ‘Rose Rae,’ The Lendings, Startforth, Barnard Castle, Co Durham, DL12 9AB, UK; Tel: +44 1833 637 446; chris@burgessconsultancy.com; www.burgessconsultancy.co

Chris Burgess

Regulations require drug manufacturers determine the detection limit (DL) and quantitation limit (QL) of impurities in APIs and drug products. There are multiple ways of doing so, some of which are described in the International Conference on Harmonisation (ICH) guideline, ICH Q2 (R1) Validation of Analytical Procedures: Text and Methodology. Primarily, these methods are based upon the statistical uncertainties from a linear least squares regression calibration curve. However, different calculation methods give different answers. The following describes an example and a recommended best practice method.

View the original article published in the November 2021 issue of 

Chris Burgess is managing director, Burgess Analytical Consultancy Limited, Barnard Castle, Co Durham, UK.

Articles

Different methods to calculate limits for impurities will give different answers. The prediction interval method may be ideal.

From PharmTech: Detecting and Determining Quantitation Limits for Impurities

For pharma companies and regulators to have meaningful discussion on the results of analyses, they should first agree on the meaning of the terms used. As detailed in Part 1 of this paper, the pharmaceutical industry currently does not have a globally agreed nomenclature for sampling and testing. 

Sampling and Testing Nomenclature for Pharmaceutical Analysis: A Simplified Approach

Pharmaceutical Technology Regulatory Sourcebook eBook

Inconsistent sampling and testing processes that may result from this lack of standardized practices can lead to errors, resulting in data integrity issues and potentially incorrect decisions about pharmaceutical quality.

In the second part of this paper, the author proposes a simplified overall process flow with associated definitions for pharmaceutical sampling and testing.

Pharmaceutical Technology’s September 2021 Regulatory Sourcebook eBook

When referring to this article, please cite it as C. Burgess, “Sampling and Testing Nomenclature for Pharmaceutical Analysis: A Simplified Approach," 

Pharmaceutical Technology Regulatory Sourcebook

A simplified process flow with associated definitions for pharmaceutical sampling and testing is proposed. 

In pharma applications, consistent sampling and testing nomenclature is critical for data integrity and compliance with specifications and regulatory standards. Inconsistent sampling and testing processes can lead to errors, resulting in data integrity issues and potentially incorrect decisions about pharmaceutical quality.

Sampling and Testing Nomenclature for Pharmaceutical Analysis: A Review of Current Practices

Guidance documents within the pharma industry and other regulated sectors, however, offer conflicting information. This review of regulatory and compendial standards reveals the scope of the problem.

This article provides a review of available analytical nomenclature, recommended naming for individual steps, and terms used in the process flows. Part two of this paper proposes a simplified approach for sampling and testing nomenclature.

When referring to this article, please cite it as C. Burgess, “Sampling and Testing Nomenclature for Pharmaceutical Analysis: A Review of Current Practices," 

Inconsistent sampling and testing processes can lead to errors, resulting in data integrity issues and potentially incorrect decisions about pharmaceutical quality. 

The title of this article is deliberatively provocative. Pharmacopoeias contain the legal standards for the quality of medicines with monographs for active ingredients, excipients, and finished products. Analysts should understand how a monograph, together with the associated general notices and general chapters, relates to their responsibilities under good manufacturing practices for ensuring data integrity. However, many pharmacopeial tests therein involve classical chemical analysis (e.g., color, appearance, titration) and rely on an analyst’s observation with little, if any, documented evidence for review. Hence, do pharmacopoeias inadvertently facilitate data integrity violations? 

In this article the authors look at the following topics:

Interaction of the monographs, general notices, and general chapters

Examination of a monograph and the interpretation into an analytical procedure to see how the testing structure in a pharmacopeia can inadvertently facilitate data integrity violations

Practical application of the requirements that data are attributable, legible, contemporaneous, original, and accurate (ALCOA) as well as complete, consistent, enduring, and available (ALCOA+) to pharmacopeial analysis to ensure data integrity.

Pharmaceutical Technology's Regulatory Sourcebook


eBook: Regulatory Sourcebook, October 2020
October 2020
Pages: 30–40

When referring to this article, please cite it as C. Burgess and R.D. McDowall, “Do Pharmacopoeias Inadvertently Facilitate Data Integrity Violations?"

 Pharmaceutical Technology Regulatory Sourcebook 

Analysts should understand how a monograph, together with the associated general notices and general chapters, relate to their responsibilities under good manufacturing practices.

Do Pharmacopoeias Inadvertently Facilitate Data Integrity Violations?

	The reportable value from an analytical procedure has three elements in its error structure: the error of the procedure itself, the error associated with the sampling of the batch, and the error associated with the manufacturing process itself. The first two errors need to be controlled and be sufficiently small as to ensure that any out-of-specification (OOS) result is unlikely to be from these causes. To do this, the developed procedure must enable the target measurement uncertainty (TMU), as defined in the analytical target profile, to be achieved on a routine basis via the analytical control strategy. The United States Pharmacopeial Convention (USP) Validation and Verification Expert Panel has been working on a new general chapter regarding analytical procedure lifecycle management (1–5).
	
	For many existing procedures, however, the knowledge base regarding these three components is lacking. Therefore, one must resort to performing designed experiments to separate these variance components.
	Consider the analysis of a drug product for an analyte, X, from an established manufacturing process and procedure that has a registered specification of 95-105% of claim. The USP analytical procedure uses a single sample from a batch and a singlet determination. This product has a history of frequent OOS results.
	Examination of the last 10 batches, which are without adverse trend, reveals 100.0% is retained and 2.977 is replaced by 2.98. The process capability report based on these values is shown in 

. Ten batches are used for illustrative purposes, and for reliable values of Cpk, Ppk, etc., many more batches would be needed.
	Based on the available information, however, it would be expected that 9.3% of batches would be OOS. 
	The issue now becomes one of identifying the size of the contributory sources of variance. In this instance, there are three: the analytical testing procedure itself, VT; the sampling process, VS; and the production process, VP. Because these sources of variation are independent, the total variance, V, is simply the sum of the individual variances.

	One can design a simple experiment to estimate the various contributions by:
	
	Sampling and testing 10 batches once; V =VT+VS+VP
	From one batch, sampled 10 times, test each sample once; V =VT+VS
	From one sample, test it 10 times; V =VT.
	
	For this example process, the results are shown in 


	It is critical to estimate the size of variance contributions of the analytical testing procedure, the sampling process, and the production process. 
	
	Contributory variances can be estimated using simple arithmetic (


	It is immediately apparent that the majority of the variation (59.4%) is coming from the process itself and about one-third from the analytical testing procedure. The sampling variance contributes only a little. One can calculate the individual standard deviations and, with the means draw, the overall picture of the variability (

	From an analytical perspective, the ideal but unobtainable situation would be to a target measurement uncertainty of zero. Therefore, the question becomes one of how small a TMU for the procedure would be needed. In addition, would replication help to avoid OOS results due to the testing variance?

	First, however, one should be confident that this simple approach is consistent by using a different approach.

	Confirmation of the error structure using Monte Carlo Simulation

	There are three components to the error structure coming from the testing, the sampling, and the process itself. A value from each of the three distributions can be selected, at random, as shown in 

and for each of these values, the error associated with each value is calculated.

	The analytical testing error, eT, would be the mean (99.3) minus the value selected and similarly for sampling, eS, and manufacturing, eP’ using the appropriate mean values of 100.6 and 100.0, respectively. Note that these errors may be both positive and negative.

	Then the observed reportable value would be (100+eT+ eS+ eP). If the Monte Carlo Simulation is used (6),  then the results from 1 million iterations should give data that closely match those found in Figure 1, which they do. The result is shown in 


	As the sampling and production variations are not controllable during testing, an attempt might be made to increase replication and/or to improve the target measurement uncertainty. However, 

 clearly demonstrates that increased replication does not give a major reduction in % OOS of the reportable values. Even if there were zero testing error, the predicted % OOS of the reportable values is still 4% (

	It is critical to estimate the size of variance contributions of the analytical testing procedure, the sampling process, and the production process. In this simple approach, the analytical testing variance is estimated from the repeatability, which is smaller than the intermediate precision, which would be a better estimate. A more complex design can be made to estimate the intermediate precision. 

	In this example, the combined production and sampling variance is so large that the process will always produce more than 4% OOS reportable values in the long term, irrespective of the testing variance. The production standard deviation needs to be reduced to approximately 1% and the testing to the same; and the sampling standard deviation needs to be reduced to 0.6 before a Ppk of 1.1 is obtained, predicting about 0.1% of OOS reportable values. Control strategies need to be in place to achieve these values.

	1. G. P. Martin et al., “Proposed New USP General Chapter: The Analytical Procedure Lifecycle <1220>,” 

 43(1) 2017.
	2. G. P. Martin et al., “Lifecycle Management of Analytical Procedures: Method Development, Procedure Performance Qualification and Procedure Performance Verification,” 

 39(5) 2013.
	3. K L Barnett et al., “Analytical Target Profile: Structure and Application Throughout the Analytical Lifecycle,” 

 42(5) 2016.
	4. C. Burgess et al., “Fitness for Use: Decision Rules and Target Measurement Uncertainty,” 

42(2) 2016.
	5. E. Kovacs et al., “Analytical Control Strategy,” 

42(5) 2016.
	6. Companion by Minitab v5.1.1.0 

	When referring to this article, please cite it as C. Burgess, "Reportable Values: Where is the Variation Coming From?," 

This article looks at a simple structured approach to assigning variance contributions and to assuring that the analytical procedure is fit for purpose.

Reportable Values: Where is the Variation Coming From?

	One of the barriers to be overcome in the wider application of measurement uncertainty (MU) to reportable values is the need to perform calculations that can be daunting to many in the analytical laboratory. Performing these calculations need not be so daunting, however. An earlier column showed how the basics of simple error budget calculations based on the Eurachem/CITAC guideline (QUAM) could be performed (1, 2). A new guide was published in 2015 extending MU concepts into decision making and setting target measurement uncertainty (3). Decision rules involving compliance with specification using guard band principles requiring MU have been published as well (4, 5). The importance of being able to calculate MU easily, therefore, is increasing. A new cloud-based software tool (Minitab DEVIZE) has become commercially available that makes the process of calculating MU simpler and much quicker (6). This article looks at some examples comparing the results from the traditional QUAM model with Monte Carlo Simulation (MCS) using this tool.
	
	


	In the original QUAM guidance, MCS modeling was discussed briefly and a technical methodology given in Appendix E3 (2), which is quite mathematical intensively and employs Microsoft Excel.
	
	The following principles behind MCS for MU, however, are simple:
	 

		Identify the sources of error and their data shape (distribution).

		Decide how they are connected and related. (Both these steps are the same as the QUAM model.)

		Select, at random, a value from each of the error sources and calculate a value for the measurand in accordance with the defined interrelationships.

		Repeat the calculation N times (usually N is 10000 or more).

		Calculate the overall mean and standard deviation of the measurand from all N simulated values.

Example one: Making a standard solution of a certified reference material (CRM).


	Suppose that one has a CRM with a certified value of 99.46% and a MU of 0.16 (k=2). A mass of 101.32 mg of this CRM is dissolved in water, transferred quantitatively to a 100.0 mL Grade A volumetric flask, and the solution is made up to the calibration mark with water. The measurand, the concentration in mg/L, is directly given by 


	
	However, the process involved in making the solution is subject to uncertainties in both the actual amount of the CRM and the volume of diluent. If these uncertainties are quantified, one will be able to estimate the overall measurement uncertainty of the solution. This process with its error sources is represented diagrammatically in 

	
	The error budget constituting the overall uncertainty for the CRM concentration is constructed from the interrelation of the uncertainty sources represented by the individual bricks in 

, the inputs for the MCS model can be defined. The inputs may be fixed (e.g., the mass of the CRM and the coefficients of expansion of water and borosilicate glass). Variables are specified with the appropriate distributional model (e.g., the laboratory temperature is assumed to follow a uniform distribution of between 18 °C and 22 °C, while the repeatability of filling the volumetric flask to the calibration mark is normally distributed with a standard deviation of 0.02 mL). The completed MCS input model is shown in 

	
	Based upon these inputs, the relevant outputs and the MCS model can be defined as:

		The amount of CRM corrected for the uncertainty of the balance and the uncertainty of the CRM purity

		The volume correction of the flask itself with respect to temperature and calibration

		The final diluent volume taking into account the uncertainty of filling to the calibration mark

		The final concentration of the CRM solution.

	These outputs to be calculated by the MCS model are shown in 

. The interrelationships between the inputs and outputs are illustrated in 


	
	An advantage of the MCS model over the conventional QUAM method becomes apparent when the analytical system becomes more complex.
	
	Now it is possible to run the model N times to generate the data, in this instance, 50,000 times, and see the output frequency distribution for CRM concentration (

). Even 50,000 iterations take only a few seconds in the cloud using DEVIZE.

	
	When the QUAM calculation for MU is performed, the mean values are essentially identical, but the value for s is a little larger at 1.27. In this example, it was assumed in the MCS model that the specification for the final concentration had to lie within the range 990 to 1010 mg/L. Inspection of 

 clearly indicates that the model predicts that a proportion (approximately 1.2%) of the concentration data will lie outside the top limit.

	An advantage of the MCS model over the conventional QUAM method becomes apparent when the analytical system becomes more complex. To illustrate this advantage, the MCS model is applied to a complex standard analytical method (7) used in the determination of leachable cadmium from ceramic and glass cooking ware from a food safety perspective. The example was chosen not because of its detailed application but because of its complexity and the fact that the QUAM gives as worked example A5 (2)i thereby allowing a direct comparison between the QUAM methodology and that of the MCS model.

Example two: Leachable cadmium (Cd) determination by atomic absorption spectrometry (AAS)


	In this example, more factors are involved than in the standard solution example. The procedure involves the following:

		A complex sample preparation involving acid concentration, time, and temperature factors

		An AAS calibration for cadmium requiring a linear least squares regression model

		Measurement of the cadmium concentration in the sample solution

		Calculation of the measurand, which is the amount of cadmium released in mg dm-2.

	Full details are given in the QUAM guideline, which requires seven pages to describe the process and calculations necessary to calculate the overall combined uncertainty (standard deviation) (see 

. The inputs required for MCS model for this example are shown in 

	
	The following two assumptions were made:

		The area variation for non-circularity of the sample can be calculated from a triangular distribution of the diameter ±0.01 from the measurement error and assuming the area of an ellipse.

		As the absorbance value of the Cd sample was not given, it was calculated from the concentration in the QUAM and assigned a s value of 0.001 based on pooling the linear least squares (LLSQ) calibration replicates for 0.1 and 0.3 mg L-1 .

	The output calculations for the MCS model and the interrelationships are shown in 

, respectively. Note that Cd conc is called C

	The MCS model was run 50,000 times, and the results for the mean Cd released (r) and its associated standard deviation calculated from 


	
	
	The excellent agreement between the MCS model and the QUAM model is shown in 

. In addition, the DEVIZE software provides an option to show a sensitivity analysis. This effectively shows a ranking of the various factor effects (see 

	
	Hence, of the 12 factors listed, the three major contributors to the overall standard deviation of the measurand are the temperature factor, the intercept of the calibration curve, and the measuring cylinder. This conclusion is consistent with that from the uncertainty budget in Figure A5.8 from the QUAM.


	In the original QUAM guidance, MCS modeling was discussed in a quite mathematical manner using an Excel spread sheet. In this column, the MCS tool from Minitab, DEVIZE, has been shown to produce similar results without the need for individual calculations in a QUAM uncertainty budget in a more user-friendly graphical format.

	The more accessible and simpler approach of MCS might encourage the calculation of MUs in more complex analytical processes on a routine basis.

 37 (9) pp. 62-73 (September 2013).
	2. S L R Ellison and A Williams (Eds), 

Eurachem/CITAC guide: Quantifying Uncertainty in Analytical Measurement,

 Third edition (2012), www.eurachem.org/index.php/publications/guides/quam
	3. R. Bettencourt da Silva, A. Williams (Eds), 

Eurachem/CITAC Guide: Setting and Using Target Uncertainty in Chemical Measurement 

(1st ed., 2015), www.eurachem.org/index.php/publications/guides/gd-stmu
	4. C. Burgess, 

 38 (10) (October 2014).
	6. Minitab Devize, www.minitab.com/en-us/products/devize/?WT.srch=1&WT.mc_id=SE022648&url_id=160744988&gclid=CPWs9Ln88sMCFSOy2wodNQsAHQ.
	7. British Standard BS 6748:1986, Specification for limits of metal release from ceramic ware, glassware, glass ceramic ware and vitreous enamel ware.


	When refering to this article, please cite it as C. Burgess, "Measurement Uncertainty without the Math," 

Performing a measurement uncertainty calculation is often seen as problematic.