PAT for the On-line Characterization of Continuous Manufacturing Systems

Transcription

1 PAT for the On-line Characterization of Continuous Manufacturing Systems Thomas O Connor, Ph.D. Office of Pharmaceutical Science FDA/PQRI Conference: Innovation in Manufacturing and Regulatory Assessment September 16 th,

2 2 Outline Role of PAT for Continuous Manufacturing Process Monitoring Statistical Process Control (SPC) Lessons for SPC from Chemical Industry Real Time Release Testing Concluding Remarks

3 3 Process Analytical Technology Process analytical technology (PAT) is a system for designing, analyzing, and controlling manufacturing through timely measurements (i.e., during processing) of critical quality and performance attributes of raw and in-process materials and processes, with the goal of ensuring final product quality PAT Guidance PAT tools Multivariate tools for design, data acquisition, and analysis Process analyzers Process control tools Continuous improvement and knowledge management tools

4 Role of PAT for Continuous Processes Process development Continuous and fast response of process to factor changes allows efficient experimentation Increases process understanding within the range of conditions studied during development Commercial Manufacturing Process Monitoring Assure desired product quality is being consistently manufactured Identify non-conforming material Real Time Release Testing (RTRT) Evaluate the quality of final product based a combination of measured material attributes and process controls, ICH Q8 R(2) Continuous manufacturing naturally lends itself to RTRT approaches 4

5 5 Sampling Considerations for PAT Measurements Probe/sample location representative of the entire vessel Minimize the effect of the probe on the process Required measurement frequency based on process dynamics Measurement resolution should sufficient for the detection of a pulse of variability from a process disturbance Sample volume/mass Representative of a unit dose Sample Interface Remains constant during the process run (e.g., no fouling) Environment factors (e.g., temperature, humidity) Evaluate reproducibility of online measurements Include sampling considerations On-line and in-line measurements may reduce but do not eliminate sampling errors

6 Conceptual Implementation of PAT for Continuous Manufacturing Receiving Feeder Granulation Blending Drying Milling At-line Chemical Properties & Physical Properties for Raw Materials Concentration & Uniformity (Multi-component) Chemometric model Feed Frame Tablet Press Online Assay Chemometric model Weight & Hardness Coating Blending Digital Imaging Surrogate Dissolution Model (release) Raw material characterization, process data, & chemometric model outputs integrated into a supervisory control and data acquisition (SCADA) system Particle size distribution Moisture content Real-Time Release Testing 6

7 7 In-Process Control Requirements To assure batch uniformity in-process controls shall be established CFR (a) Examinations to be conducted on appropriate samples of in-process materials for each batch Requires higher frequency measurements for continuous processes compared to batch processes Controls shall monitor and validate the performance of the manufacturing processes that may cause variability in the drug product Valid in-process specifications shall be consistent with the release specification CFR (b) Limits shall be derived from acceptable process variability estimates where possible Rejected in-process materials shall be identified and isolated CFR (d) PAT tools can utilized to meet the regulatory requirements for in-process monitoring Approaches may include multivariate process monitoring

8 8 Approaches for Process Monitoring Statistical Quality Control (SQC) Variability in quality attributes of the product are monitored over time Appropriate limits are defined based on the statistical analysis of historical operations Does not guarantee that the process is in control (lagging indicator) Statistical Process Control (SPC) The variability in critical process parameters are monitored over time Monitoring the process expected to supply more information (e.g., detection and diagnosis) May generate a large number of univariate control chart that need to be monitored Multivariate Statistical Process Control (MSPC) Takes advantage of correlations between process variables Reduces the dimensionality of the process into a set of independent variables May detect abnormal operations not observed by SPC

9 9 Multivariate Statistical Process Control (MSPC) Reduction in dimensionality X1 & X2 are highly correlated Potential to enhance fault detection capabilities

10 Control Strategy Implementation Process monitoring strategy may depend upon control strategy implementation MSPC strategy may be more suitable for Level 1 and 2 control strategies Level 1 Real-time automatic control + Flexible CPPs to respond to variability in CMAs Level 2 Reduced end product testing + Flexible CMA & CPP within design space Level 3 End product testing + tightly constrained material attributes and process parameters Control Strategy Implementation Options 1 1. Yu, L et. al. AAPS J Vol

11 11 Multivariate Methods Latent Variable Models Process data and product quality data are decomposed into scores (T) and loading (P, W) X= TP T + E Y= TQ T + E Scores are related to samples and represent the projection of each sample into the space defined by the new latent variables Loading are related to variables and represent the relationship between the process variables and the new latent variables Principal Component Analysis (PCA) Determines the latent variables that maximize the variation in process data captured by the model Well suited for process monitoring applications Partial Least Squares (PLS) Determines the latent variable that maximizes the correlation between the process and product quality data Well suited for predicting quality attributes from process data (e.g. soft sensors, surrogate models etc.)

12 12 Process Monitoring Statistics MSPC model statistics can be calculated to assess the overall variability of the process Alleviates the need to monitor all the combinations of latent variables on x- y plots Hotelling T 2 The distance of the current operating point from the mean of the historical normal process data Able to detect process stretches Process moves into new operating region; the relationship between the process variables may have not changed Squared Predicted Error (SPEX) The sum of the squared error between the measured process data and the predicted process data Able to detect a change in the relationship between process variables

13 13 Diagnosing Process Faults Hotelling T 2 Sum of the squared score values normalized by the amount of variation captured by each latent variable; T 2 = A i=1 Squared Predicted Error t2 i λi Sum of the squared error between the model predicted and measured process n variables; SPEX= x m x 2 p i=1

14 Perspective on MSPC for Continuous Processes from Chemical Industry Statistical Quality Control routinely utilized to assess performance of on-line analyzers Focus on analyzers used to control end product quality Control charts plot the error between on-line and periodic off-line measurements MSPC utilized for process fault detection to address process safety and reliability issues Complementary to the alarm and advance control systems Considerations for the implementation of MSPC for continuous process Variable selection for MSPC models Data selection for building MSPC models Variable manipulations MSPC model validation On-line implementation 14

15 15 Variable Selection for MSPC Models Utilize process knowledge to select variables Risk assessment approaches may be used to identify failure modes Identify process variables that signal each failure mode Examine process variable cross-correlation matrix Statistical approach used to identify process variables that are related Dependent upon the process data selected for analysis Investigate control variables as well as process variables Tightly controlled process variables typically provide little information Information may be contained in the manipulated variable (e.g., valve opening etc.) L The % valve opening contains information about the variation in the process

16 16 Data Selection for MSPC Model Building Historical data utilized should capture the normal variations expected during the process lifecycle Should include data from routine operations (e.g. feeder refills, etc.) Process data from disturbances should be removed from the model building data set Minimize amount of steady-state data included in the model building data set Steady state data masks the relationship between variables Variation is mostly due to noise Statistic filters can be used to remove steady state data Majority of variation reflects noise Majority of variation reflects physical relationship

17 17 Data Selection Considerations Cont d All the data from process development studies may not be suitable for building MSPC models Relationships between variables are influenced by the process design and the process control structure Control structure may not be finalized till the end of process development MSPC models may need to be re-vamped after process control projects are implemented MSPC models may be initially constructed from process performance qualification data Process qualification combines the actual facility, utilities, equipment with the commercial manufacturing process, control procedures, and components to the end product MSPC models may be updated as part continuous process verification Unlikely that all the sources of normal variation will be experienced during the process qualification process

18 18 Considerations for Variable Transformations Process data is typically auto-scaled Mean-centered to focus on variation (i.e. x =0) Scaled by the standard deviation to equally weight each variable (i.e. s=1) The mean and std. dev. of each process variable is calculated from the model building data set A constant mean may not be suitable for monitoring a continuous process Filter mean to remove the impact of slow process shifts or a change in operating condition (e.g., high vs low flow rates) Filter constant should be based on the process time constant Do not utilize mean updating to monitor long term issues (e.g., catalyst aging, fouling) Increased sensitivity over a process run

19 19 Variable Transformations Cont d Weight variables to increase or decrease their influence in the MSPC model Considering increasing the weight of variables that are early indicators of faults identified by the risk analysis Considering decreasing the weight of redundant measurements (e.g., multiple temperature measurements in an unit operation) Equivalent to manipulating the standard deviation for the process variable PCA/PLS multivariate models are linear Utilize transformation to incorporate non-linear relationship PCA/PLS multivariate models are static Need to compensate for time lags for upstream and downstream process variables The MSPC models should utilize the dynamically compensated variable

20 20 MSPC Model Validation Utilize independent data sets for validation studies Validate models with normal operating data Goal is to minimize false positive alerts Validate models with operating data during process disturbances (if available) Assess detection ability and timing for process faults Distribution of process diagnostic statistics have fat tails Better performance obtained by setting SPEX and T2 limits based on validation studies rather than the std. dev. obtained form normal operating data Minimizes false positives with minimal impact on detection capability or detection timing

21 21 On-line Implementation Considerations Ensure reading of process variables are synchronized Facilitated be integrating all process and quality data into a single source (e.g., process historian, supervisory control system, etc.) Ensure process data are not compressed Process historians may utilize compression for storing data Data compression can mask the relationship between process variables Need to account for missing and bad data May be able to use model estimates for the missing data Need to establish criteria when the status of the model will become bad (e.g., number of bad inputs, missing critical inputs) Utilize bumpless transfers when missing process data becomes available Incorporate MSPC model into the Quality Management System Establish work process for responding to alerts Establish work process for assessing the health of the MSPC model

22 22 Relationship between RTRT, Process Monitoring, and PAT Real-time-release-testing (RTRT), when used, is part of the Control Strategy Can include some or all of then finale product CQAs Science and risk based approaches generally required for both MSPC and RTRT Not all Process Analytical Technology leads to RTRT PAT systems can be designed to control CQAs of raw materials or inprocess materials and not contribute to RTRT Some PAT tools may be utilized for both process monitoring and RTRT MSPC approaches that establish a process signature are an evolving approach for RTRT

23 23 Benefits of RTRT Provides for increased assurance of quality More process data collected Provides increased manufacturing flexibility and efficiency Shorter cycle time Reduced inventory Reduction in end product testing Reduction in manufacturing costs

24 24 Consideration for the Point of Testing Is there a potential for the measured CQA to change downstream from the measurement point? For example, Blend desegregation Loss of weight (e.g., chipping) after weighting step Hydrolytic degradation during aqueous film coating Is identity determined at a point that is visually unique? Mitigation of potential human and/or system error Unique identifiers on the intermediate when measured (e.g., embossing, size, shape) Risk assessment is valuable to exploring potential failure modes

25 25 Considerations for RTRT Specification Specification still required in an RTRT approach (CFR (d) and CFR (a)) Should be representative of actual measurement Can include in-process measurements (e.g., NIR measurements for assay of uncoated tablets) Can include surrogate measurements (e.g., models for dissolution) Methods should be appropriately validated (including models used as surrogate measurements) Alternatives can be included for stability testing Utilization of appropriate statistical criteria for large sample sizes

26 26 Concluding Remarks Continuous manufacturing process facilitate the adoption of PAT tools for the development of process understanding, process monitoring, and real time release testing PAT may required to meet the regulatory requirements for in-process monitoring MSPC offers several advantages over univariate monitoring of process data MSPC is a complementary process control tool to an automatic process control system MSPC approaches that establish a process signature are an evolving approach for RTRT