Capability Assessments and
Process Validation Stage 3 Implementation:
1.33 and Beyond
MBSW 2012
Midwest Biopharmaceutical Statistics Workshop
May 21-23, 2012
Presenter: Krista Witkowski
Co-author: Julia O’Neill
Merck & Co., Inc.
Abstract
This talk will discuss considerations for practitioners in
pharmaceutical manufacturing as they implement the new FDA
guidance for process validation. We will focus on Stage 3 - ongoing
monitoring, or continued process verification - and how process
capability is established, evaluated, and monitored. Examples on
overcoming obstacles to implementation will be discussed, and the
use of statistical thinking in our implementation strategy is
highlighted.
2
Requirements of FDA Validation Guidance
•
FDA Guidance for Industry: Process Validation: General Principles and
Practices, published January 2011 distinguishes three stages of validation:
– Stage 1 – Process Design: The commercial manufacturing process is defined
during this stage based on knowledge gained through development and scale-up
activities.
– Stage 2 – Process Qualification: During this stage, the process design is
evaluated to determine if the process is capable of reproducible commercial
manufacturing.
– Stage 3 – Continued Process Verification: Ongoing assurance is gained during
routine production that the process remains in a state of control.
•
Further states that manufacturers should understand the sources of
variation
– Detect the presence and degree of variation
– Understand the impact of variation on the process and ultimately on product
attributes
– Control the variation in a manner commensurate with the risk it represents to
the process and product
3
Stage 3: Continued Process Verification
Process Validation
Stage 2
Stage 1
Process
Qualification
Process
Design
Continued
Process
Verification
Stage 3
4
Stage 3: Continued Process Verification
Develop
Monitoring
Plan from Control
Strategy Document.
Continually monitor
critical areas of the
process
Goal=To continually assure
that the process remains in a
state of control (the validated
state) during commercial
manufacture.
Develop Monitoring Reports
Assessing the data
on a frequent basis
(e.g., monthly, quarterly)
Make any
adjustments
to continually
assure the process
remains in a state
of control. Update
Control Strategy document
if needed
5
Understanding Variation for Pharmaceutical Processes
Issue: Statistical Process Control (SPC) procedures are generally designed
based on assumptions not typically met by pharmaceutical processes:
•
SPC Assumptions
– Independent results
– Specifications based on customer
needs
– Normally distributed results
•
Pharmaceutical Processes
– Autocorrelation
– Specifications based on process
history
– Non-normal distributions common
(e.g., lognormal)
6
Issue 1: Autocorrelation
Independent results
Autocorrelated results
Result vs. Previous Result – correlation not significant
Result vs. Previous Result – significant correlation = .35
7
One Cause for Autocorrelation
Production lots
Growth
Propagation
Purification
Production (Weeks)
A new raw material lot
introduced late in the production
cycle has little opportunity to
impact a product lot;
however, a new raw material lot
introduced early in the production
cycle has a much greater opportunity
to impact a product lot.
This creates gradual trends (autocorrelation),
rather than abrupt shifts, in product properties.
Introduction of
New raw material lot
Introduction of
New raw material lot
8
One Solution: Use long-term sigma
Independent results:
short-term and long-term
limits are nearly equal.
 ST 
 LT 
R
d2
n

i 1
xi  x 2
n 1
Long-term
Short-term
Autocorrelated results:
short-term limits are narrower
than long-term limits.
Long-term limits are more
representative of
process capability.
Short-term
Long-term
9
Example 2: Inherent mean shifts
I Chart of Process data
Mean shifts may be inherent – due to
campaign effects, raw material changes,
slight changes in processing conditions
(e.g., seasonal effects).
1.25
1.00
Individual Value
 ST
Short-term limits based on MRbar/d2
R

d2
1
1
1
0.75
1
UCL=0.888
22
22
2 2
2
2 222
2
22
0.50
0.25
_
X=0.479
2
2
2
0.00
 LT 
n

i 1
xi  x 2
1
1
1
13
Short-term
25
37
n 1
49
61
73
Observation
85
97
LCL=0.071
109
I Chart of Process data
Long term 3s limits
1.25
UCL=1.124
Results with mean shifts:
short-term limits are narrower
than long-term limits.
Long-term limits are more
representative of
process capability.
Individual Value
1.00
2
22
22
2 2
2
2 222
2
0.75
22
0.50
0.25
Long-term
2
_
X=0.479
2
2
2
0.00
LCL=-0.165
1
13
25
37
49
61
73
Observation
85
97
109
10
Understanding sources of variability
Distribution
with additional
source
ofofvariability,
σ=2
Distributionof
ofVariable
variableA,
A
reflecting
sources
variability,µ=21.5,
µ=19,
µ=20,σ=2.5
σ=2
Distribution
reflecting
allinitial
sources
of variability
LCL
LCL
14
28
UCL
16
mu20
mu20
mu20
18
mu21
mu21
20
22
mu21
24
UCL
26
mu19mu19
1
1
25.0
25.0
25.0
26
UCL=25.38
UCL=25.38
UCL=25.38
UCL=26.88
Final limits (n=90)
Early limits (n=30)
Individual Value
Individual
Value
Individual Value
Individual Value
24
22.5
22.5
22.5
22
_
_ _ _
X=20.23
X=20.07
X=20.23
X=20.23
20.0
20
20.0
20.0
18
17.5
6
17.5
16
17.5
15.0
14
5
15.0
15.0
12
1
1111
10
74
10
19
13
719
1
28
19
28 37
25
3146 5519
37556422
43647325
4973
82 28
5582
10
1337 4616
Observation
Observation
Observation
Observation
Do not set limits too early,
before all sources of variability
are captured.
LCL=15.08
LCL=13.25
LCL=15.08
LCL=15.08
11
Statistical Thinking Strategy: for Autocorrelation
•
Standard Statistical Process Control (SPC) chart assumptions:
– Observations are statistically independent – very important!
– Observations are Normally distributed – much less important.
– Limits are representative of expected performance.
•
Autocorrelation can have profound effects on the performance of SPC
charts.
•
Considerations for control chart design:
– Quickly signal real changes in results.
– Reduce false alarms.
– Make the chart easy to interpret –
• present results in original scale, and
• limits with a physical meaning.
•
Recommendation;
– Set limits using the overall standard deviation based on a “long” stable
period.
– Bisgaard and Kulahci provide an elegant justification.
12
Issue 2: Establishing Process Capability
• Two challenges:
– Fundamental questions for pharmaceutical processes:
• Are long-term shifts (for example, from raw material trends) “extraneous”
sources of instability?
• Or are they known and predictable special causes inherent to
pharmaceutical process behavior?
– Specifications may be set based on process consistency, not customer
requirements.
13
Three Approaches to Capability Strategy
Specification Spread
Short-Term =
Often underestimates total
process variation
6 * short-term Sigma
higher is better
Specification Spread
“Quality” =
6 * long-term Sigma
Business Requirements
“Business” =
6 * long-term Sigma
14
Basics of Capability Calculations
 USL  X X  LSL 
 USL  X

X  LSL
  min

C pk (or Ppk )  min
,
,



3 
 3
 UCL  LCL  / 2 UCL  LCL  / 2 
The mean and standard deviation are estimated from the centerline
and control limits of the control charts, where three sigma is half the
width of (UCL-LCL).
Well Off-target /
Too Much Variation
Relatively Close to Target /
Moderate Variation
LSL
LSL
Cpk < 1
USL
USL
Cpk = 1
Very Little Deviation From Target
LSL
USL
Cpk > 1
15
Short term vs Long term
Grp 1
Grp 5
Grp 3
Grp 2
Short Term Studies
Grp 4
Long Term Study
16
Example 2: Short term variability < Long term
Process Capability Sixpack of Process data
I Chart
Individual Value
1.0
Short-term  ST  R  0.14
d2
Capability Histogram
1
1
1
1
USL
UCL=0.888
S pecifications
USL 1
_
X=0.479
0.5
0.0
1
1
13
25
37
49
61
73
85
97
LCL=0.071
1
109
-0.00 0.15
0.30 0.45
Moving Range Chart
A D: 0.802, P : 0.037
1
0.50
UCL=0.5015
0.25
__
MR=0.1535
0.00
LCL=0
1
13
25
37
49
61
73
85
97
109
-0.5
0.0
Last 25 Observations
Long-term  LT 

i 1
Values
xi  x 
2
n 1
 0.21
Use the P-indices to provide a
realistic assessment of long-term
performance. For independent
(not autocorrelated) processes,
the P-indices and C-indices will
be nearly equal.
Within
S tDev 0.136082
Cp
*
C pk
1.28
0.5
105
110
Observation
1.0
115
Within
O v erall
S tDev 0.214831
Pp
*
P pk
0.81
C pm
*
O v erall
S pecs
0.0
100
0.5
Capability Plot
1.0
n
0.60 0.75 0.90
Normal Prob Plot
1
Moving Range
C indices underestimate total
process variation when
autocorrelation is present
(when “within subgroup”
variation is low compared to
overall).
120
Process Capability of Process data
USL
P rocess Data
LS L
*
Target
*
USL
1
S ample M ean
0.479301
S ample N
120
S tDev (Within)
0.136082
S tDev (O v erall) 0.214831
Within
Ov erall
P otential (Within) C apability
Cp
*
C PL
*
C P U 1.28
C pk
1.28
O v erall C apability
Pp
PPL
PPU
P pk
C pm
-0.00 0.15
O bserv ed P erformance
P P M < LS L
*
P P M > U S L 0.00
P P M Total
0.00
0.30
E xp. Within P erformance
P P M < LS L
*
P P M > U S L 65.02
P P M Total
65.02
0.45
0.60
0.75
E xp. O v erall P erformance
P P M < LS L
*
P P M > U S L 7680.36
P P M Total
7680.36
0.90
*
*
0.81
0.81
*
Cpk = 1.28
Ppk = 0.81
17
Example 2: Short term variability < Long term
C indices underestimate total
process variation when
autocorrelation is present
(when “within subgroup”
variation is low compared to
overall).
Long-term
 ST
 LT 
n

i 1
xi  x 2
n 1
1.25
1.124=Long term
1.00
Individual Value
Short-term
R

d2
I Chart of Process data
Short-term limits based on MRbar/d2
1
1
1
22
22
2 2
2
2 222
2
0.75
1 = USL
UCL=0.888
1
22
0.50
0.25
_
X=0.479
2
2
2
0.00
1
1
LCL=0.071
-0.165=Long term
Use the P-indices to provide a
realistic assessment of long-term
performance. For independent
(not autocorrelated) processes,
the P-indices and C-indices will
be nearly equal.
1
13
25
37
49
61
73
Observation
85
97
109
One-sided: USL = 1
Cpk = 1.28
Short-term
Ppk = 0.81
Long-term
18
Risk Strategy: Ppk Comparison of CQA’s
Capable & Stable Process (≥1.33)
Process Robustness & Simplification
Opportunities (<1.33)
Frequency of monitoring report guided by risk strategy
Ppk for 27 Critical Quality Attributes of a family of pharmaceutical products.
Each bar represents the estimated Ppk for a single CQA.
The bars are ordered from lowest Ppk (greatest risk) to highest.
Note: Ppk is long-term capability, but takes into account centering of the process
within specifications.
In cases when there is a very large range of values for Ppk, a log scale can make
this more read-able, while still maintaining the “red, yellow, green” risk categories
19
Other Choices in Capability Indicators
Process
Characterization
Summary Statistics
Proportion
Defects
Calculate
Calculate
Indicators
Cp Cpk
Pp Ppk
Xbar (Mean)
OR
 (Std. Dev.)
Calculate using
Use Z
Table or
Minitab
specifications and
process data
Calculate
Gather
Data
Convert to DPM
Process
Z
Score
DPM(Upper)
+
DPM(Lower) =
DPM (Total)
Z Score
Use Z
Table or
Minitab
ZUPPER
ZLOWER
20
Translating Pass/Fail to Ppk - type Index
• Non-normal or pass/fail data: Use a "z-score" approach
• Calculate the z-score using normal distribution theory
– Proportion good  z-score
– Translate z-score to a “Ppk-type" scale: divide by 3.
3*Ppk = z-score
Ppk = z-score / 3
• Does not account for sample size, so results should be viewed in
light of the amount of information you have
• Example:
– If 99% is "good“ (“within spec”):
• z-score is 2.33,
• Ppk = 2.33/3 = 0.78
Cp for a “6 sigma process”: Cp =
USL- LSL
12
=
= 2
6
6
21
Statistical Background on Capability
•
Capability index assesses whether a process is capable of meeting
customer requirements.
•
Capability: “the natural or undisturbed performance after extraneous
influences are eliminated”
– from the Western Electric Company Statistical Quality Control Handbook (1956)
•
“Cpk can be calculated when the process is stable. Otherwise, for
processes with known and predictable special causes and output meeting
specifications Ppk should be used.”
– from the AIAG PPAP Manual (2006)
•
Most important: PLOT THE DATA ON A CONTROL CHART.
– Exact value of capability index is secondary.
22
Issue 3: LogNormally Distributed Results
Normal Results
LogNormal Results
Error does not depend on measurement.
Characterized by constant
Standard Deviation.
Results are symmetric within limits.
Error is proportional to measurement.
Characterized by constant
Relative Standard Deviation (RSD)
Results are not symmetric within limits.
Has little impact if range of results is less than 10X.
Easily corrected by analyzing results on the log scale.
23
Solution: Log Transform Results
LogNormal Results
Same data on
different scale
Log (LogNormal Results)
Log transform makes error constant and results symmetric within limits.
24
References
•
•
•
Bisgaard, S., Kulahci, M.. (2005) Quality Quandaries: The Effect of Autocorrelation on
Statistical Process Control Procedures. Quality Engineering 17: 481-489.
AIAG. “Definition of Process Measures.” Statistical Process Control. AIAG, 1995. pp
80-81. 2nd Printing.
The Black Belt Memory JoggerTM. (2002) GOAL/QPC Six Sigma Academy. First
edition. p. 96.
25