Process Robustness Aided by Retrospective Multivariate Data Analysis of Factory Data - a Case Study

Process robustness aided by retrospective multivariate data analysis of factory data – a case study     

Sankar Raghavan, Cherokee Pharmaceuticals L.L.C. a wholly owned subsidiary of Merck & Company,  Sankar.Raghavan@merck.com , 570-271-2106

Jeff Foster, Cherokee Pharmaceuticals L.L.C. a wholly owned subsidiary of Merck & Company, Jeffrey.Foster@merck.com, 570-271-2173

Dale Kopas, dakopas@gmail.com, 609-417-1707

Ed Warner, Merck & Company, edwarner4@gmail.com, 908-670-1902

Presenter:  Sankar Raghavan

Keywords:  Multistep process, Noisy data, Multivariate analysis, Continuous improvement

Purpose:  Illustrate the power of combining process knowledge and advanced statistical analysis in addressing intricate process issues

Abstract

Quality by Design is a powerful tool that can be applied to continuously improve performance and robustness of processes.  However, with complex processes, it can be effort intensive and costly to do particularly when the process has multiple steps and the scale-up and the interaction effects are not fully understood.  Since most pharmaceutical processes are batch, factory data with its normal variability provides a vehicle to generate such data without incurring excessive cost.  Such a case study is presented – The example involves a manufacturing process for an antibiotic drug substance that involved four reactions and several down stream purification steps.  Multiple factors influenced each of these steps.  Identification of key factors that affected the process proved difficult as univariate and bivariate approaches such as SQC/SPC control charts and simple linear regression models explained only a small fraction of the process variability. Multiple interactions added to the complexity, and the presence of a high degree of co-linearity between process variables in the X-matrix lead to problems with matrix inversion in traditional multivariate regression models such as stepwise regression.  To correct the situation, data mining was attempted by advanced multivariate methods, specifically Random Forest and Partial Least Squares (PLS) to supplement process knowledge.

The paper describes the analysis of inherently noisy factory data for an entire campaign consisting of over 200 individual batches and more than 100 variables by a combination of advanced statistical methods and physical reasoning based on expert knowledge of the process.   The method was effective in explaining the broad factory trends.  Key outcomes were:

1. The methodology helped rank factors in terms of their importance.  This made it possible to focus on a subset of few factors.
2. Implementation of lessons derived from the studies has resulted in a more robust process with improvements in purity, yield and productivity.
3. The work pointed to areas of opportunity to improve the process. 

The method aligns well with the F&DA initiative to continuously improve robustness of processes through the life cycle of a process.  The methodology should be broadly applicable for the continuous improvement of complex processes.