Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Limites Microbianos Por Poisson

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

JVT_May2007.

qxd

4/23/07

8:04 AM

Page 196

Application of Poisson
Distribution in Establishing
Control Limits for Discrete
Quality Attributes
B Y P R A M O T E C H O L AY U D T H

INTRODUCTION
This article is written as a complement to one recently
published entitled Establishing Alert Limits for Microbial Counts in Purified Water, which appeared in this Journal Volume 13, Number 1, November 2006. In the current
paper, the scope will be expanded i.e., establishing control
limits for discrete quality data attributed to products, raw
materials, and critical system outputs, which also include
the microbial counts in Purified Water as discussed in the
previous article. The establishment criterion will remain the
same i.e., using historical data and the Control Chart principle where, in this article, the control limits are simply derived from the Poisson distribution that is corresponding in
distribution pattern to the original Binomial distribution.

DISCRETE QUALITY ATTRIBUTES


Control limits for discrete quality attributes may be
found in the following areas:
Trend Data Analysis For example, establishing
trend limits for:
Microbial counts in raw materials, products,
and Water for Pharmaceutical Use (WPU)
Environmental Monitoring (EM) microbial
levels in production areas
196

Journal of Validation Technology

Validated Products For example, establishing:


Control limits for numbers of tablets rejected
from an online Metal Detector during tablet
compression cycle
Control limits for numbers of containers rejected from visual inspection of sterile production batches
Release limits for microbial counts in nonsterile products
Validated Critical Systems For example, establishing:
Alert limits for microbial levels in cleanroom
environment
Alert limits for microbial counts in Water for
Pharmaceutical Use (such as Purified Water or
Water for Injection)

UNDERSTANDING OF POISSION
DISTRIBUTION
When we repeatedly take samples of the same size n =
90 from a population containing only conforming (say, 95%
of the population) and non-conforming (5%) items, the
counts of non-conformities will distribute according to both
Binomial and Poisson distributions (Figure 2). Poisson is a
special case of Binomial where n and p 0.
(Continued on page 198)

JVT_May2007.qxd

4/23/07

8:04 AM

Page 197

Pramote Cholayudth

Figure 1
Binomial and Poisson Probability Density Functions (PDFs)

Figure 2
Binomial and Poisson Distributions

Probability (%)

(Non-Conforming Rate: p=5% =0.05, n=90)


22%
20%
18%
16%
14%
12%
10%
8%
6%
4%
2%
0%

Binomial Dist, Mean = np = 4.5, SD = (npq)^0.5 = 2.07


Poisson Dist, Mean = np = 4.5, SD = (np)^0.5 = 2.12

9 10 11 12 13 14 15 16 17 18 19 20

Number of Success

M a y 2 0 0 7 Vo l u m e 1 3 , N u m b e r 3

197

JVT_May2007.qxd

4/23/07

8:04 AM

Page 198

Pramote Cholayudth
(Continued from page 196)

The two distribution curves illustrated in Figure 2 indicate that the parameter np = 90x0.05 or 4.5 is sufficient to
construct a Poisson distribution that is quite similar in shape
to the Binomials. It is also suggested by statisticians that np
< 5 can generate the common shape between Binomial and
Poisson distributions. (Interested readers may discover
more in Statistics textbooks.)
Using the Control Chart principle and data above, comparison of the control limits for np Chart (Binomial distribution) and c Chart (Poisson distribution) will be as noted in
Figure 3.

(c )

ESTABLISHING CONTROL LIMITS


Control limits for discrete quality attributes may be established as below:
Collect appropriate amount of data (at least ten individual test or measurement results)
Compute for the average value of all results ( c )
Compute for the control limits using c 3 c
principle
However, c 4 c and c 6 c

may be

applied, as alert and action limits, respectively,


to the microbial levels to compensate the
possible sampling and testing technique errors
Since alert and action limits for microbial
levels in aseptic areas are also required, the
limits

c2 c

and c 4 c

may be

applied respectively (see Figure 12)


Justify the established limits before use by plotting
c Chart using that particular data
Revise the limits as appropriate when the database
is larger

Figure 3
Control Limits Based on np and c Charts

np Chart Approach

c Chart Approach

The results in Figure 3 imply that the control limits in discrete numbers established from
both Binomial and Poisson approaches are practically the same especially for np < 5.
198

Journal of Validation Technology

JVT_May2007.qxd

4/23/07

8:04 AM

Page 199

Pramote Cholayudth

MICROBIAL ALERT LIMITS

PARTICLE COUNT IN CLEANROOM

Comparison of alert limits, using np Chart according to


the previously published article (JVT Vol. 13, No. 1, Nov.
06) and c Chart for microbial counts in Purified Water is
provided in Figure 4.
In the c Chart Approach column, the limits data (in discrete numbers) are the same as those in the np Charts.
Therefore the c Chart method can be practically used in
place of the np method and also has an advantage over it

since only the averaged count of non-conformities ( c ) is


employed without taking the sample size into account. In

Figure 6 the c value up to 7.5 (> 5) still generates the same


discrete data limits (0, 18) although the distribution area has
slipped from the Binomials.

Particle count in non-unidirectional cleanroom (turbulent air) is typical discrete data. In theory, its distribution
should follow a Binomial/Poisson distribution. Measuring
the particle counts at a fixed location in a cleanroom area at
time series will generate different measured values that
require to be evaluated. Since an area may have k sampling
locations based on the area size (i.e., rounded integer of
=
square root of the area in m2), the averages ( x ) for these
locations measurements require to be computed for the
=
grand average ( x ) and standard deviation (SD). The 95%
upper confidence limit (95% UCL) using the formula
x + t 0.05, k 1 (SD / k ) is then computed to demonstrate
that it still meets the designed cleanliness requirement.
(Continued on page 202)

Figure 4
Comparison of Control Limits Using np and c Charts Principle

np Chart Approach

c Chart Approach

M a y 2 0 0 7 Vo l u m e 1 3 , N u m b e r 3

199

JVT_May2007.qxd

4/23/07

8:04 AM

Page 200

Pramote Cholayudth

Figure 5
Alert Limits for Microbial Count in Purified Water at Points of Use

Microbial Count Distribution in Purified Water


(Count Mean: 2.06 CFU/mL, Action Limit: 100 CFU/mL)
30%
Binomial

Frequency (%)

25%

Poisson

20%
15%
10%
5%
0%
0

10

12

14

16

18

20

Microbial Count (CFU/mL)

Figure 6
Alert Limits for Microbial Count in Purified Water at Washing Area

Microbial Count Distribution in Purified Water (+Hose)


(Count Mean: 7.5 CFU/mL, Action Limit: 100 CFU/mL)
16%
14%

Binomial
Binomial
Poisson
Poisson

Frequency (%)

12%
10%
8%
6%
4%
2%
0%
0 2

8 10 12 14 16 18 20

22 24 26 28 30 32 34

Microbial Count (CFU/mL)

200

Journal of Validation Technology

JVT_May2007.qxd

4/23/07

8:04 AM

Page 201

Pramote Cholayudth

Figure 7
Alert Limits for Microbial Count in Purified Water after RO Unit

Microbial Count Distribution in Purified Water (after RO)


(Count Mean: 3.9 CFU/mL, Action Limit: 100 CFU/mL)
25%
Binomial

20%
Frequency (%)

Poisson
15%
10%
5%
0%
0

10

12

14

16

18

20

22

24

Microbial Count (CFU/mL)

Figure 8
Alert Limits for Microbial Count in Purified Water after EDI Unit

Frequency (%)

Microbial Count Distribution in Purified Water (after EDI)


(Count Mean: 5.3 CFU/mL, Action Limit: 100 CFU/mL)
20%
18%
16%
14%
12%
10%
8%
6%
4%
2%
0%

Binomial
Poisson

10

12

14

16

18

20

22

24

Microbial Count (CFU/mL)

M a y 2 0 0 7 Vo l u m e 1 3 , N u m b e r 3

201

JVT_May2007.qxd

4/23/07

8:04 AM

Page 202

Pramote Cholayudth
(Continued from page 199)

Computing the UCL is based on the Central Limit Theorem


i.e., the mean (averaged) values of samples (of appropriate
size) even taken from a non-normal population will follow a
normal distribution. The air turbulence will influence the
particle count measurement biased far beyond what it should
be. So, computing using the Binomial or Poisson principle is
not practical since the resulting UCL will be much lower
than that from the Central Limit Theorem principle.

ESTABLISHING CONTROL LIMITS FOR


REJECTED CONTAINERS
The number in each batch of sterile product containers
rejected due to their failure to pass the visual inspection is
also a discrete type data. Establishing the acceptance limits
for these rejected containers is essentially required. In a
pharmaceutical plant where the author provides consulting
service, one set of data for containers rejected after visual
inspection from ten batches of an Infusion product has
been collected. Their control limits are established as seen
in Figure 9.
In addition to the established overall control limits (38,
86), control limits for individual types of defects (particles,

glasses, and fibers) are separately established due to their


different sources of contamination (Figure 10). Such limits
are useful as the tools for analysis and improvement of individual defect rates. It was suggested that Production continuously improve to reduce the numbers of those defects
and subsequently adjust the control limits using the defect
data after the improvement. Training and qualification (OQ
and PQ) of the visual inspection process was also recommended. Figure 10 demonstrates how the greater values, up

to 62, of the averaged counts of non-conformities ( c ) are


still justified for using the Poisson distribution approach.
Using the overall limit information, standard yield limits
for production batches can be established as in Figure 11.

LIMITATION OF POISSON
APPLICATION
There are some limitations for the application use of
Poisson distribution. Some types of counts of non-confor
mities (or numbers of defects) data with the average ( c )
greater than x or smaller than y may fail to generate the
proper control limits since Poisson distribution may not
exist. The values of x (expectedly high) and y (expect-

Figure 9
Control Limits for Rejected Containers

Batch #
1
2
3
4
5
6
7
8
9
10
Average ( c )
UCL
LCL

Particles
13 (min)
32 (max)
18
21
19
26
22
18
27
19
21.5
35
8

UCL = c + 3 c , LCL = c 3 c
202

Journal of Validation Technology

Glasses

Fibers

14
7 (min)
16
8
12
12
15
20
14
21 (max)

16
17
19
38
9 (min)
39 (max)
36
33
25
34
26.6
42
11

13.9
25
3

Overall
43
56
53
67
40 (min)
77 (max)
73
71
66
74
62
86
38

JVT_May2007.qxd

4/23/07

8:04 AM

Page 203

Pramote Cholayudth

edly low) require to be identified. From a preliminary trial


and error experiment, x is probably 100 and y is 1. In any
case, it is strongly recommended to justify the established
control limits prior to use and periodically review during
their use.
Appropriate handling of the data is very important e.g.,
for microbial counts in Purified Water, the zero result (denoted as < 1) is preferably transformed to 1 before computation of the alert limits. Regarding microbial levels in an
aseptic processing environment, it is more justified to di-

rectly use the zero (0) for computing the alert and action
limits. In Figure 12, two approaches for establishing control
limits for environmental microbial levels are provided. The
cumulative frequency approach was introduced by Robert
A. Fry 6 i.e., 95% and 99% confidence levels are set as the
alert and action limits, respectively, based on non-normal
data. Two sets of limit results (cumulative and c Chart) are
to be evaluated for justification by plotting as c Charts. The
most valid set of limits is selected.
For the high-valued data of which the value is greater

Figure 10
Distributions of Rejected Containers

Poisson Distributions: Rejected Containers


Glasses
Mean = 13.9

12%

Fibers
Mean = 26.6

10%
Frequency (%)

Particles
Mean = 21.5

8%

Total
Mean = 62.0

6%
4%
2%
0%
0

5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95
Number of Rejected Containers

Figure 11
Yield Information: Standard Yield Limits

Yield Descriptions

Quantities (containers)

Theoretical Yield
3,921 ( = 400x1000/102)
Number of QC Samples
24
Control Limits Rejected Containers
UCL: 86
LCL: 38
Standard Yield Limits
UL: 3,859* (98.42 %)
LL: 3,811** (97.19 %)
Batch size: 400.- liters, target fill volume: 102 mL/container
* UL (Upper Limit) = 3,921- 24 - 38 = 3,859, ** LL (Lower Limit) = 3,921- 24 - 86 = 3,811

M a y 2 0 0 7 Vo l u m e 1 3 , N u m b e r 3

203

JVT_May2007.qxd

4/23/07

8:04 AM

Page 204

Pramote Cholayudth

than 100, the Poisson distribution will have a limitation to


exist. However, bear in mind that the data of interest is the
non-conformity (defect) data; these should be as low as possible. Establishing control limits for high defect numbers
means we are accepting high defect rates poor quality
which is the wrong concept. In any case, if such limits are
really required, it is necessary to collect the data from
smaller samples so that reduced defect numbers fulfilling
the Poisson PDF requirements are obtained. The new data
can be then converted to control limits.

CONCLUSION
Historical, discrete quality attribute data in sufficient

quantity are averaged ( c ) to estimate the mean of the entire


Poisson distribution, the basic control limits are then com

puted using c 3 c principle. These particular data must


be obtained from the same size of samples taken at random.
The established control limits are justified before use and
periodically evaluated (i.e., reviewed and revised) as appropriate. Justification is undertaken by plotting the limits-establishing data on the c Chart to verify the fitting ability of
the limits. All control limits, established based on data, are
reviewed or revised on a periodic basis. The evaluating frequency is another interesting issue where its justification is
required. From a quality assurance (QA) point of view, it is

good practice to carry out the periodic evaluation of such


control limits.

ABOUT THE AUTHOR


Pramote Cholayudth is Executive Director of
Valitech Co., Ltd., a well-established validation and
compliance consultant services company for the
Pharmaceutical Industry in Thailand. He is a guest
speaker on Process Validation to Industrial Pharmaceutical Scientists organized by the local FDA. Mr.
Cholayudth was a full-time lecturer in a School of
Pharmacy in a private university for four years (19982001). Prior to entering the academic arena, he
spent 23 years in the Pharmaceutical Industry with
Bayer Laboratories (1974-1981) and OLIC (Thailand)
Limited (1981-1997) a leading, and the largest,
pharmaceutical toll manufacturer for multinational
companies. Pramote is the author of Concepts and
Practices of Pharmaceutical Process Validation. He
is currently a respected, contributing member of
JVT's Editorial Advisory Board. Pramote can be contacted by fax at 662-740-9586, by e-mail at
cpramote2000@yahoo.com, or at the following mailing address:

Figure 12
Approaches for Establishing Control Limits for Microbial Levels

Cumulative Frequency Approach


Results
(CFU/Plate)

%
Occurrence

0
1
2
3
4
5

86
9
2
1
1
1

Cumulative
Frequency (%)
86
86+9 = 95
95+2 = 97
97+1 = 98
98+1 = 99
99+1 = 100

c Chart Approach
Control Limits
(Poisson)

Average
(CFU/Plate)
0.00*
0.09*
0.04*
0.03
0.04
0.05

Alert Limit:
Overall
= 0.25

= 0.25 + 2
=1
Action Limit:
= 0.25 + 4
=2

Alert Limit: 1 CFU/plate (95% Confidence)

Alert Limit: 1 CFU/plate

Action Limit: 4 CFU/plate (99% Confidence)

Action Limit: 2 or 3** CFU/plate

* 0x(86/100) = 0.00; 1x(9/100) = 0.09; 2x(2/100) = 0.04, ** =

204

Journal of Validation Technology

0.25

0.25 + 5 0.25

0.25

JVT_May2007.qxd

4/23/07

8:04 AM

Page 205

Pramote Cholayudth

Pramote Cholayudth
6/756 Number One Complex,
Bangkok-Ram 2 Road,
Pravate District, Bangkok, 10250
Thailand

REFERENCES
1. WHO Technical Report Series # 929, Thirty-ninth Report,
Annex 3: WHO Good Manufacturing Practices: Water for
Pharmaceutical Use, World Health Organization, 2005.
2. Center for Drug Evaluation and Research (CDER),
Guidance for Industry, Sterile Drug Products Produced
by Aseptic Processing - Current Good Manufacturing
Practice, September 2004.
3. Ad Hoc GMP Inspections Services Group, EC Guide to
Good Manufacturing Practice, Revision to Annex 1,
Manufacture of Sterile Medicinal Products, May 30, 2003.
4. The United States Pharmacopoeial Convention, Inc.,
General Information Chapter <1116> Microbiological Evaluation of Clean Rooms and Other Controlled Environments,
United States Pharmacopoeia 27th Edition, Philadelphia,
PA: National Publishing.
5. Besterfield, D. H., Quality Control, Sixth Edition, Prentice
Hall, 2001.
6. Fry, R. A., CQA Microbiology Support, Presentation on
Environmental Monitoring: Establishing Action and Alert
Levels, PDA Spring Conference, Orlando, Florida, March
11-13, 2002.
7. Wadsworth, H. M., Stephens, K. S., and Godfrey, A. B.,
Modern Methods for Quality Control and Improvement,
Second Edition, John Wiley & Sons, 2002.

Article Acronym Listing


EDI
EM
IPC
LCL
OQ
PDF
PQ
QA
RO
SD
UCL
WPU

Electrodeionization
Environmental Monitoring
In-Process Control
Lower Control Limit
Operational Qualification
Probability Density Function
Performance Qualification
Quality Assurance
Reverse Osmosis
Sample Standard Deviation
Upper Control Limit
Water for Pharmaceutical Use

M a y 2 0 0 7 Vo l u m e 1 3 , N u m b e r 3

205

You might also like