PQRI
Process Drift Workshop
Case Study
Denise Rivkees, R.Ph., Ph.D.
Pfizer Global Manufacturing
The Setting:
Product Quality Lifecycle Implementation
Purpose of this PQRI Presentation
• To demonstrate the use and advantages of Product Quality Lifecycle Management
• To demonstrate the use and advantages of new technology
• To demonstrate the use these tools in multiple collaborator environments over departments and organizations
• To demonstrate how the paradigm allows progress to go forward even in light of unexpected results
• To demonstrate what the data “look like” and demonstrate software differences and effects
• To demonstrate the challenges that occur
Notes:
• The presentation represents a live work-in-process that will continue after this presentation.
• The model process was out-of-trend, but no result was outside of a specification that would impact patient care.
Situation Description
• An excipient supplier changed the milling step for a raw material used in an extrusion process. Risk analysis revealed little expectation of impact to the critical quality attributes, and a straightforward like-for-like validation plan was put in place.
• During Offsite Stability Testing of the validation lots, the average 30 minute dissolution results began to drift while site retains at 30 minutes complied with the L2 dissolution criteria.
• FMEA followed the traditional investigation route to special causes:
Was it the excipient?
Was it the offsite lab?
Was it something about packaging?
Was it The dissolution method?
But the new paradigm of lifecycle management proved the root cause was none of these.
Filter Settings
- Accept: (True)
Initially, multiple studies were performed as part of an FMEA,
but the results showed nothing out of the ordinary. For example, fill weight
studies concerning the fill of individual raw materials were conducted.
Results of fill weight study:
Dissolution method
Guage R&R Studies of dissolution across sites
FOUND
All sites and methods were WITHIN NORMAL VARIATION
None of these aspects explained the change!
Example: Multiple aspects concerning the dissolution test, including the
test method and variance across multiple sites, were examined.
Ivelisse Colon-Rivera, PGRD
Process Drift Case Study Overview
Occurred during a material Change Control
A drift in Critical Quality Attribute was noted
The drift was visualized through Process Capability
Data Mining was used to identify Latent Variables
Data used for Risk Assessment
Control Strategy established
Restart Manufacturing in relatively short time period
Quality By Design DOE for Process Understanding
Identification of new latent variable
Latent variable used for Continuous Improvement
Process Capability Phase
Low variability within a lot observed from 2005
Variability within a lot increased from 2007
DATA TRENDS: IMR Chart
Validation batch
for Excipient #2
process change
at manufacturerExcipient #1
Slight change
associated
With small
variance
Compared
to spec
20092009200920082007200720052005200520052003
8
6
4
2
0
Mfg Date
In
div
idu
al
Va
lue
_X=2.22
UC L=5.76
LC L=-1.32
20092009200920082007200720052005200520052003
8
6
4
2
0
Mfg Date
Mo
vin
g R
an
ge
__MR=1.331
UC L=4.347
LC L=0
1
1
1
111
1
1
11
I-MR Chart of 30min Stdev by Mfg Date
Trending by
Israel Cotto, PGM
What is it? Matrix Composed of API in Excipient
Confocal Raman Imaging by
Don Clark, PASG
How is it made?
Three raw materials charged to blender via mill
Blend for 45 minutes
Blend charged to downstream process equipment
Extruder
Five additional major downstream processes
Process Drift Case Study Overview
Occurred during a material Change Control
A drift in Critical Quality Attribute was noted
The drift was visualized through Process Capability
Data Mining was used to identify Latent Variables
Data used for Risk Assessment
Control Strategy established
Restart Manufacturing in relatively short time period
Quality By Design DOE for Process Understanding
Identification of new latent variable
Latent variable used for Continuous Improvement
*
CONTROL STRATEGY PHASE
-5 0 5-4
-3
-2
-1
0
1
2
3
4
t3
t 4
Scatter Plot: t4 vs t
3
V090762
V09076318818V
V09037317178V
v
1.2
UNK
OOS
Data Mining Shows Outliers
Model Developed showing inverse relationship between API particle size
and excipient critical quality attribute.
Restriction of particle size and excipient attribute implemented.
Manufacturing to proceed with temporary Control Strategy.
Using these tools, the stability samples of interest led us to other lots that
were out of trend according to the model. Those lots were withheld based
on the prediction. DataMining by
Sal Garcia, PGRD
DataMining showed that the biggest contributors to
variability were particle size and excipient value
The site used the model developed from the site
data to match API and Excipient to achieve
desired dissolution
as a continuous improvement
Control Strategy
Business Factor Came Into Play:
Request to study new API source,
so we added it in and it revealed
information that we would not
have gathered had it not been in
the study
Process Drift Case Study Overview
Occurred during a material Change Control
A drift in Critical Quality Attribute was noted
The drift was visualized through Process Capability
Data Mining was used to identify Latent Variables
Data used for Risk Assessment
Control Strategy established
Restart Manufacturing in relatively short time period
Quality By Design DOE for Process Understanding
Identification of new latent variable
Latent variable used for Continuous Improvement*
Process Understanding Phase
1. Investigate a relevant design space
2. Acquire the raw materials
3. Run a pilot study
4. Based on Pilot Results, Design the DOE
5. Run the DOE
6. Analyze the data
Understand the Process!
1: Original Design Space Study Proposal
Statistical Support from
Anthony Carella, PGRD
Factorial Design 2X3
High and Low Excipient Values
High, Med, and Low API Size
Three Center Points
Eleven runs
Plus New API Source
2: API Milling Design of Experiments
Mill Study by
Al Pichieri, GMS
Jose R. Rivera, Site
Adjustable D6 Fitz Mill
Mill Settings: Low, Med, High
Produced: Low, Med, and High Particle Size
This study showed how we could manipulate the
particle size using a different mill. Useful for
continuous improvement of the API process.
3. Pilot Study: Dissolution and Water Content
*There was not enough water for success
**Routine finding for water content
Water added normally to process so we looked at three levels
In order to see the effect on the online NIR
Lot
number
Water
injected
(%)
Water
content
(%)
Dissoluti
on
At 30
minutes
0286-017 1 2.3 NA*
0286-013 2 3.0** 42
0286-015 3 3.8 39
Productio
n
Control
NA NA 61
Surprise!
Low results compared
to Production Control.
This meant we had to
expend DOE runs to
investigate equipment
differences at the pilot
study contractor.
No root cause was
Identified, but the team
agreed that the particle
size and excipient attribute
would show up in a smaller
DOE anyway, so we
Proceeded.
3 Pilot Blends
Appeared Identical by NIR: So far, so good
2nd derivative
SG 9pt.
2nd derivative SG 9pt.
Offline NIR (Bruker MPA) – 3 Lots
Appear Identical by NIR
Margot Cortese, Ron Beyerinck, Bend Research Offline NIR (ePAT601)
NIR (ABB) Extruder: 3 Lots - Comparable by NIR, non water
peaks overlapped
2nd derivative SG 9pt.
Zoom of first overtone region
H2OCH2
Zoom of second overtone region
H2O
Effect of water
variation
4. Design the study based on
successful pilot
API - Hi API- Lo
Excipient
Hi
Hi- Hi Lo-Hi
Excipient
Low
Hi- Low Low-Low
New Study Proposal
PLUS NEW API Source to examine effects
Statistical Support from
Anthony Carella, PGRD
Run
API
D (4,3)
µ
Excipient
Value
KF Water
%
Added
KF
%
Final
30 min
Dissolution
81 40±11 Low 2.15 3.4 45
82 40±11 High 2.20 3.5 44
83 89±10 Low 2.06 3.4 41
8489±10
High 1.94 3.4 43
85 52 High 3.6 3.6 48
Control NA NA NA NA 61
Study Results:
Smaller Particle Size Led to Faster Dissolution
KF data shown here for illustrative purposes because water typically
has an influence, but was not a major factor in this studySlight trend for the old API,
but something was different
about the new API.
Top-line Statistical Summary for Bend
DoE• Relevant data on the two parameters (API PS and Excipient
Value AV) . . .
• Mitigating factors
– Slower (& harder to impact) dissolution profiles using 27-mm extruder at
Bend
– Small DoE (due to resource constraints) led to weak statistical power
– Weak statistical conclusions, due to weak statistical power and to robust
manufacturing process (with respect to changes in dissolution)
• Statistical conclusions (over the parameter ranges studied) . . .
– Indication that increasing API PS decreased % dissolved at 30 minutes
– No indication that Excipient Value affected % dissolved at 30 minutes
– Indication that lot # BREC-0286-095 differed from other 4 lots
5. Analyze the Results
PAT is a major enabler! Even with weak statistical power, online trends revealed were known to beassociated with other findings throughoutall of the studies - and were therefore considered scientifically important
Online Blender NIR
Signals for alkyl groups move
Note: The NIR stopped working after
10 minutes on the first run, but the
effect of blending was already in place
after 5 minutes, as shown in trend
analysis below.
Note: When the new API batch was made,
the engineers called to say the API was more
cohesive and handled differently than ever
before. We knew at that minute why it might
yield different results. See PCAs on later
slides
1 Excipient CH2
2 Excipient CH3
3 API CH2
4 API CH3
5 API bound H20
6 Excipient CH2
These trends show powder
relfectance/absorbance
and are associated with density
and particle size.
One would expect the
middle particle size group,
the new API, to be in the
center --it was here
but not on dissolution!
Extruder Online NIR
API
Variance 1
Variance 2
Variance 3
Note: Second loading (Variance) almost a straight line after melting-
does this mean something??
DIFFERENT
SOFTWARE
NIR Peaks Assigned for API and Excipient
Zoom on API Peaks Zoom on Excipient Peaks
API+ H2O Peaks
Compritol PeaksExcipient
These two peaks directly demonstrate the trade off between
particle size and excipient value, VALIDATING the Data Mining
model
Trending with Excipient Peak at 1732 nm
(Second derivative)
Lot 81 Lot
82
Lot 81 Lot
82TREND
Visualization of raw data
producing this trend
Note large movement
In Excipient Peak
But small movement
in API peak
What changed between the two? The excipient value: and it was reflected in the
movement of the excipient alkyl peak AND the API particle size (or granule size?) as shown
In thereflectance that produces the trend line. This shows the interaction.
Trending with API Peak at 1694 nm
Lot 83Lot 84
Lot 84
Lot 83
TREND
Note less movement in
Excipient signal than last slide
Trending with Excipient Peak at 1732 nm
Lot 85
87 40 0.96 197 45
89 40 2.56 199 44
91 89 0.96 201 41
93 89 2.56 201 43
95 52 2.56 194 48
MSC
#
API
D(4,3)
Exc
AV
Sphere
D(4,3)
Disso
30min
I
N
C
R
E
A
S
I
N
G
R
E
F
L
E
C
T
A
N
C
E
The new
API absorbance
was in trend with
the other sizes,
but dissolution
was different-
meaning that
something
else
about the
new API
was different
TREND ANALYSIS
NIR Peaks Assigned for API and Excipient
Zoom in API Peaks Zoom in Excipient Peaks
API+ H2O Peaks
Compritol PeaksExcipient
Trending with API Peak at 1694 nm:
40 micron API does not produce a stable process
Lot 81
Lot 81
Lot 82
Lot 82
Regardless of mechanistics, this is not a stable process Regardless of mechanistics, this is not a stable process
Note difference induced by excipient value
Trending with API Peak at 1694 nm: Larger API Particle
Size Produces a More Stable Process
Lot 83
Lot 84
Lot 84
Lot 83
Stable Stable
Note:
Less
Variation
Than 81 and 82
Trending with API Peak at 1694 nm
Lot 85
Lot 85
52 micron (average)
Material still does not
produce a flat trend
Different Trend Analysis:
Average of last 10 Spectra of Each Blend
Lot 81, 82, 83, 84 and 85
Bend Ref 81
Bend Ref 82
Bend Ref 83
Bend Ref 84
Bend Ref 85
Note crossovers:
Why would this happen?
Look at the raw data or second derivatives for clues,
then trend those points as we did in the previous slides
PCA Score #1 Vs. Score #2
If one agrees than some small particles might aggregate and act
as larger particles, then the first PCA (measures variance) trends
with particle size this way-
PCA Score #1 Vs. Score #2
But the dissolution appears to
trend on a different axis- because it is
Influenced by a different variable:
Something to do with the API surface
Particle Size Quantitative Model Calibration
Results
97% Correlation
NIR Spectra Correlate to %Dissolution (30 Mins)
Based on PLS
98.5% Correlation!
Dissolution
Correlated
With
Online NIR Spectra
by
98.5%
NEXT:
Determine and use the new latent variable identity
to
correlate with dissolution and online NIR
In order to steer the knowledge gathering
and
CONTINUOUS IMPROVEMENT
Process Understanding!
SUMMARY
• Traditional Batch Record Control Strategy initially
• Drift
• Temporary New Control Strategy using Data Mining
• Designed Experiment for Process Understanding
• Latent Variables Identified: The latent variable was not the original excipient that precipitated the study
• Permanent Control Strategy Determined
• Continuous Improvement
Product Quality Lifecycle Implementation
Team
Jose Mercado
Vionette Padovani
Israel Cotto
Victor Ruiz
Alex Montanez
Amaryllis Roman
Denise Sanchez
Elvira Alvarez
Jose Rivera
Bend Research:
Margot Cortese
Ron Beyerinck
Greenridge Consultants:
Leah Appel
Josh Shockey
Matt Shaffer
Alan Phillips
Chi-Shi Chen
Ke Hong
Don Clark
Shailesh Hiremath
Al Pichieri
Anthony Carella
Ivelisse Colon-
Rivera
Sal Garcia-Munoz
Julian Lo
Avi Thombre