a probability based equivalence test of nir vs hplc analytical methods in a continuous ... ·...
TRANSCRIPT
Areti Manola, Jyh-Ming Shoung, Stan Altan
Janssen R&D
A probability based equivalence test of
NIR vs HPLC analytical methods in a
Continuous Manufacturing process
validation study
Midwest Biopharmaceutical
Statistics Workshop
Muncie, Indiana
May 17, 2016
Outline
2
1. Overview of Continuous Manufacture
2. Process Performance Qualification
Verification of HPLC – NIR calibration
o Study Design
Single vs Multiple Analytical Runs
o Method Equivalence
3. Relative Performance Index at the individual analytical
determination level
Comparison of ratios of probabilities
Bayesian approach
4. Case Study
5. Summary
P method) | |y(|
Recent Industry Announcements
3
Though making the switch from batch to continuous manufacturing may be difficult, costly and time consuming, pharma manu- facturers and CMOs should begin to consider the switch as in the long-run it will end up saving companies time, money and space, FDA’s CDER Director Janet Woodcock told congressmen in a hearing Thursday. http://www.in-pharmatechnologist.com/Processing/FDA-calls-on-manufacturers -to-begin-switch-from-batch-to-continuous-production
http://www.fiercepharmamanufacturing.com/story/vertex-jj
-gsk-novartis-all-working-continuous-manufacturing-facilities/2015-02-09
Engineering Definition of Continuous
Manufacturing vs Batch Manufacturing
4 FDA Perspective on Continuous Manufacturing, Sharmista Chatterjee, Ph.D.
IFPAC Annual Meeting Baltimore, January , 2012
Example of Continuous Manufacturing
with On-line Monitoring
5 FDA Perspective on Continuous Manufacturing, Sharmista Chatterjee, Ph.D.
IFPAC Annual Meeting Baltimore, January , 2012
Ideal Future Vision of CM
6
Traditional
Release Approach
7
Granulation
Milling
Drying
Milling
Coating
Compression
Lubrication
Blending
Laboratory
Dissolution control
through DoE end point
Drying time by NIR
Blend Uniformity control
through NIR
NIR Composite Assay
Dosage Uniformity by tablet weight control
Tablet Thickness
Control through LoD
Weight, Thickness,
Hardness, DT, friability
Standard DP release Tests
Physical Description
Composite Assay(HPLC)
Impurity (HPLC)
Content Uniformty (Mass)
Real Time
Release Approach
Advantages of Continuous Manufacture
Scientific/Engineering Operational/Business
Integration of QbD concepts
Cohesive development, quality
and technical operations
Application of new
methodologies, technologies
and equipment
Improved process capability
Real time understanding of
process integrating process
engineering, chemometrics
and statistical considerations
Reduces costs
Streamlined facility lay-out
Reduction of raw materials and intermediates inventories
Flexibility in supply size
Payoff
Reduction of overall drug substance and drug product development time
Improve time to market
Assures supply of high quality product
Process Engineering Aspects
Unit Operations Model: Mixer Residence Time Distribution Model
9
• Critical process parameters (CPPs)
• Data (DEM predicted particle velocities)
Outputs
• Powder properties • Unit responses:
• Flow Rate • Mixing (RTD)
Unit: Mixer
Detector (i.e., PAT tools)
Pulse of Tracer (i.e., API)
RTD models are used to determine:
1. Disturbance dissipation along process
2. Raw material traceability
3. Scale up requirements
0
( )MRT t E t dt
0
( )(t)
( )out
C tE
C t dt
10
Chemometric Aspects
Process Analytical Technology (PAT)
• Near Infra Red (NIR) calibration modeling for PAT
during Process Design
PAT PLS = assess comparability to HPLC method
• Gage R&R designs relating HPLC to NIR during
development and subsequently during process validation
11
Statistical Aspects
Process Validation
• Stage 1 Process Design – DoE, data analysis and interpretation
• Stage 2 Process Performance Qualification
Statistical sampling protocols for Large n sampling plans
IPC sampling
Verification of HPLC – NIR calibration
• Stability protocol and sampling designs
• Definition of sampling frequency and sample size
• Stage 3 Continued Process Verification (cPV) - Process Capability
Verification of HPLC – NIR calibration
12
Calibration model is developed during Process Design, a Gage
R&R design can be used to assess comparability
Blocked design is 3 concentrations, at target tablet weight, tablets are
the blocks leading to essentially a paired comparison design
HPLC analytical run effect is an important consideration
During the stages of PV, the calibration model will be tested in
production
Equivalence measures will be calculated
Schuirmann’s test at the method mean level
We propose a Relative Performance Index using a Bayesian
approach to the assessment of an individual analytical
determination falling within a prespecified limit of the true value
for comparing NIR vs HPLC – major paradigm shift from the
method mean level to the analytical determination level
Model must acknowledge HPLC analytical run design
Relative Performance Index
13
Assuming the HPLC is the gold standard method, the probability of a
single analytical determination y from HPLC (or NIR) falling within
some interval of the true value µ is calculated as follows:
where (•) is the CDF of standard normal distribution.
The Relative Performance Index is defined as follows:
HPLCHPLC σ
ΔΦ
σ
Δ Φ PH HPLC)| |y(|Pr_
NIRNIR σ
biasΔΦ
σ
biasΔ Φ PN NIR) | |y|(Pr_
HN Pr_/Pr_Rel_Pfm
Method Comparison using the Relative Performance
Index
14
Given delta, bias and Method Variability
OC Curves of probability of falling within delta of true value
Rel_Pfm across delta of true value
Criterion for equivalence
Pr(Rel_Pfm≥1)≥PC, where PC is a desired probability level
HPLCPr_
NIRPr_
Case Study - Data Description
15
A single CM batch was sampled as follows:
20 locations chosen equispaced throughout the CM run
3 tablets per location
Tested by both NIR and HPLC methods
NIRHPLC NIRHPLC
102
100
98
102
100
98
102
100
98
NIRHPLC
102
100
98
NIRHPLC NIRHPLC
Location = 1
Method
Ass
ay
Location = 2 Location = 3 Location = 4 Location = 5
Location = 6 Location = 7 Location = 8 Location = 9 Location = 10
Location = 11 Location = 12 Location = 13 Location = 14 Location = 15
Location = 16 Location = 17 Location = 18 Location = 19 Location = 20
1
2
3
Tablet
Statistical Model
16
Variance components model:
Yj(i),k = assay of jth tablet (j=1,2,3) from ith
(i=1,2,…,20) location from kth (k=1,2 for
HPLC, NIR) method,
Mk = overall mean from kth method,
Li = random effect of ith location: ~ N(0, L2),
Tj(i) = random effect of jth tablet from ith
location: ~ N(0, T2),
j(i),k = residual error from kth method:
~ N(0, k2).
Preliminary analysis showed no location effect,
therefore the random effect of location was
dropped from final model.
kijijikkij TLMy ),()(),( Sampling
Design
Location
Number of
Tablets
Sampled
1 3
2 3
3 3
…
i 3
…
18 3
19 3
20 3
REML Parameter Estimates
17
Effect Parameter Estimate (se)
95% Confidence
Interval
Lower Upper
Fixed
HPLC 100.01 (0.10) 99.81 100.32
NIR 100.18 (0.05) 100.08 100.29
Bias* (NIR-HPLC) 0.17 (0.09) -0.02 0.36
Random
(SD)
Tablet 0.35 0.26 0.53
Residual (HPLC) 0.70 0.58 0.86
Residual (NIR) 0.20 0.11 1.04
*The 90% confidence interval for Bias = (0.01, 0.33)
JAGS – Posterior Samples
18
A Bayesian simulation of the posterior distribution and credible intervals based on the previous model was done using JAGS with vague priors:
Mean[HPLC], Mean[NIR] ~ N( Mean=100, SD=10 )
SD_Tablet ~ U(0, 5)
SD_HPLC, SD_NIR ~ U(0, 5)
60,000 posterior samples:
Number of chains=3
Burn-in = 20000
No. of sample = 20000
thin=25
JAGS – Parameter Estimates and Credible
Intervals
19
Effect Parameter Mean (Median)
95% Credible
Interval
Lower Upper
Fixed
HPLC 100.02 (100.02) 99.81 100.22
NIR 100.19 (100.19) 100.08 100.29
Bias* (NIR-HPLC) 0.17 (0.17) -0.02 0.36
Random
(SD scale)
Tablet 0.36 (0.36) 0.21 0.47
Residual (HPLC) 0.72 (0.71) 0.59 0.89
Residual (NIR) 0.19 (0.20) 0.02 0.37
*The 90% credible interval for Bias = (0.01, 0.33)
Normal Density Plots of HPLC and NIR centered on
the true mean Given Estimated mean bias and median of sigmas for HPLC and NIR methods
20
210-1-2
2.0
1.5
1.0
0.5
0.0
X
Norm
al D
ensi
ty
NIR Bias=0.17
0 0.7144
0.17 0.1973
Bias SD
OC Curves of Probability of Falling within delta of
True Value Given Estimated mean bias and median of sigmas for HPLC and NIR methods
21
Relative Performance Index across delta values Given Estimated mean bias and median of sigmas for HPLC and NIR methods
22
Summary of Posterior Distribution (JAGS) of Relative
Performance Index with Various deltas
23
delta Mean Median Maximum Minimum Pr(Rel_Pfm ≥ 1)
0.05 2.21 2.01 24.06 0.00 0.819
0.10 2.25 2.05 12.13 0.00 0.849
0.15 2.29 2.11 8.46 0.00 0.890
0.20 2.31 2.15 6.64 0.00 0.929
0.25 2.28 2.17 5.78 0.00 0.959
0.30 2.21 2.15 4.91 0.00 0.980
0.35 2.11 2.10 4.24 0.00 0.991
0.40 2.01 2.02 3.72 0.00 0.996
0.45 1.90 1.91 3.33 0.00 0.999
0.50 1.79 1.80 3.01 0.00 0.999
0.55 1.70 1.70 2.75 0.86 >0.999
0.60 1.61 1.61 2.54 0.89 >0.999
0.65 1.53 1.53 2.36 0.92 >0.999
0.70 1.47 1.46 2.21 0.95 >0.999
0.75 1.41 1.40 2.08 0.97 >0.999
0.80 1.35 1.34 1.97 0.99 >0.999
0.85 1.31 1.30 1.87 1.01 1.000
0.90 1.26 1.26 1.79 1.02 1.000
0.95 1.23 1.22 1.71 1.02 1.000
1.00 1.20 1.19 1.64 1.02 1.000
Comparison of Schuirmann’s Test and Relative
Performance Index for Method Comparison
24
Test Criterion delta= 0.25 delta = 0.50
Schuirmann’s 90%Credible
Interval of Bias
90%CI =
(0.01, 0.33)
90%CI =
(0.01, 0.33)
Fail Pass
Relative
Performance Index
Pr(RPI ≥ 1) ≥ 80% Pr(RPI ≥ 1) = 0.96
Pr(RPI ≥ 1) = 1.0
Pass Pass
Example of 3 analytical run design
25
Design consisted
of 20 locations, 3
tablets at each
location
This has a natural
correspondence
to a 3-analytical
run design
Leads to an
orthogonal design
Location
Method
HPLC NIR
Run
1
Run
2
Run
3
1 X X
X
X
2 X
X X
X
3 X
X
X X
…
18 X X
X
X
19 X
X X
X
20 X
X
X X
Example of 3 analytical run design
26
NIRHPLC NIRHPLC
102
100
98
102
100
98
102
100
98
NIRHPLC
102
100
98
NIRHPLC NIRHPLC
Location = 1
Method
Ass
ay (
%LC)
Location = 2 Location = 3 Location = 4 Location = 5
Location = 6 Location = 7 Location = 8 Location = 9 Location = 10
Location = 11 Location = 12 Location = 13 Location = 14 Location = 15
Location = 16 Location = 17 Location = 18 Location = 19 Location = 20
1
2
3
Tablet
Statistical Model incorporating analytical
run design
27
Variance components model:
where
Yj,i(k) = assay of jth tablet (j=1,2,…,n) by kth (k=1,2) method
measured in the i(k)th analytical run,
Mk = fixed effect of kth method,
Tj = random effect of jth tablet: ~ N(0, T2),
ρi(k) = random effect of ith analytical run from kth method: ~ N(0, ρk
2),
j,i(k) = residual error: ~ N(0, k2).
)(,)()(, kijkijkkij TMy
Relative Performance Index incorporating
multiple HPLC analytical runs
28
Assuming the HPLC is the gold standard method, the probability of a
single analytical determination y from HPLC (or NIR) falling within
some interval of the true value µ is calculated as follows:
where (•) is the CDF of standard normal distribution.
The Relative Performance Index is defined as follows:
22
,
22
,
HPLC)| |y(|Pr_
HPLCHPLCσσ
ΔΦ
σσ
Δ Φ PH
HPLCHPLC
22
,
22
,
NIR) | |y|(Pr_
NIRNIRσσ
biasΔΦ
σσ
biasΔ Φ PN
NIRNIR
HN Pr_/Pr_Rel_Pfm
Example of 3 analytical run design
- ML Estimates
29
*The 90% confidence interval for Bias = (- 0.57, 1.13)
Effect Parameter Estimate (se)
95% Confidence
Interval
Lower Upper
Fixed
HPLC 99.97 (0.29) 98.73 101.21
NIR 100.25 (0.08) 99.93 100.58
Bias* (NIR-HPLC) 0.28 (0.29) -0.98 1.54
Random (SD)
Tablet 0.32 0.20 0.74
Run (HPLC) 0.47 0.25 2.35
Residual (HPLC) 0.71 0.59 0.91
Residual (NIR) 0.49 0.38 0.70
Example of 6 analytical run design
30
Design consisted of 20 locations, 3 tablets at each location
A 6-analytical run design can be achieved with an incomplete blocking scheme
Leads to an orthogonal design
Location HPLC Analytical Run
1 2 3 4 5 6
1 1 2 3
2 1 2 3
3 1 2 3
4 1 2 3
5 1 2 3
6 1 2 3
7 1 2 3
8 1 2 3
9 1 2 3
10 1 2 3
11 1 2 3
12 1 2 3
13 1 2 3
14 1 2 3
15 1 2 3
16 1 2 3
17 1 2 3
18 1 2 3
19 1 2 3
20 1 2 3
Example of 6 analytical run design
31
2.01.51.0 2.01.51.0
102
100
98
102
100
98
102
100
98
2.01.51.0
102
100
98
2.01.51.0 2.01.51.0
Location = 1
Method
Ass
ay (
%LC)
Location = 2 Location = 3 Location = 4 Location = 5
Location = 6 Location = 7 Location = 8 Location = 9 Location = 10
Location = 11 Location = 12 Location = 13 Location = 14 Location = 15
Location = 16 Location = 17 Location = 18 Location = 19 Location = 20
1
2
3
Tablet
Example of 6 analytical run design
- ML Estimates
32
*The 90% confidence interval for Bias = (-0.42, 0.54)
Effect Parameter Estimate (se)
95% Confidence
Interval
Lower Upper
Fixed
HPLC 100.19 (0.23) 99.59 100.80
NIR 100.25 (0.08) 100.06 100.45
Bias* (NIR-HPLC) 0.06 (0.24) -0.56 0.68
Random (SD)
Tablet 0.31 0.19 0.77
Run (HPLC) 0.52 0.31 1.48
Residual (HPLC) 0.71 0.58 0.91
Residual (NIR) 0.49 0.38 0.70
Summary
33
CM is being actively encouraged by the FDA; companies are now
engaged in weighing its costs/benefits
CM offers many scientific and business advantages; major
quantitative stakeholders are process engineers, chemometrician,
statisticians working together to ensure quality
Equivalence of NIR to gold standard HPLC can be established
through a Relative Performance Index evaluated through
Bayesian calculations
• Made possible because Tablet dispersion can be removed
orthogonally given the paired comparison design
• Provides a natural interpretation of method performance
• Advantage to consider HPLC analytical run design to minimize the
effect associated with analytical run
A THANK YOU to
34
Steve Novick
Tara Scherder
Eric Sanchez
Gilfredo Alfredo
For their assistance and support with the presentation.