Multivariate Analysis for Manufacturing Quality Systems: Lessons from Novartis’ Tableting Operations

Sept. 30, 2005
MVA can allow you to use existing sensors and measurements to make each successive batch better and more effectively.
Final drug quality is influenced by everything that happens during manufacturing — each process step, each ingredient, the condition of the equipment and even subtle changes in the manufacturing environment itself can lead to variations in product quality. The univariate specifications presently used for characterization of raw materials cannot adequately describe their quality or influence on the final product, often allowing problems to go undetected (Figure 1, below).Only by developing multivariate models that account for each potential cause of variability, and by applying process analytics, can drug manufacturers establish a foundation for manufacturing quality systems at their facilities. MVA should be a cornerstone of any PAT program.Unfortunately, for many people today, PAT is synonymous with process analytical chemistry, when it’s really much more. Process analytical chemistry may be a fast and precise pharmaceutical quality control tool, but, without modeling, it only shows a fraction of the variability picture and provides limited benefit. Analytics alone will not provide early fault detection or improve process knowledge, and may not align well with an organization’s larger business strategy.Multivariate analysis (MVA; see "Demystifying Multivariate Analysis" for McCready's primer on this) allows manufacturers to model the influence of various processing steps and ingredients, so that major quality risk factors can be identified, even before a decision is made to invest in new analytical equipment.This article will show how combined MVA of raw materials and each production step is used to develop the foundation for a pharmaceutical manufacturing quality systems approach. Other industries, notably petrochemical, semiconductor and foods, have established such an approach, and pharmaceutical manufacturers are moving toward that still-elusive goal.The technology needn’t require investments in new sensors, although it can point to a need for such investment, and provide insights required for smarter purchases.The article will also highlight some examples of what Novartis and Umetrics have learned so far from an ongoing project at the Swiss pharmaceutical manufacturer’s Suffern, N.Y. facility. At Suffern, MVA has been used to develop models, which have been calibrated and tested on an existing tableting production line, using existing sensors and raw material measurements. The resulting models are being used successfully to monitor quality in real time, detect faults, and predict final product quality. (Editor’s Note: For more information on this project, read November/December’s issue of Pharmaceutical Manufacturing.)Why univariate analysis doesn’t workUnivariate tools are often not adequate or affective for monitoring of pharma and biotech processes, which are complex and not well understood. Much of the important information in pharma production is contained in the correlation of the variables rather than the state of specific parameters. Multivariate tools model the relationships between variables, extracting more information from a process or spectra than univariate methods. The result is a greater sensitivity to subtle changes and less risk of missing serious upsets. Multivariate methods provide visual summaries of plant performance with tools for diagnosis of changes in operation. Consider one example of a pharmaceutical process upset (Figure 2, below); by projecting the 33-dimensional space onto a 2-dimensional plane, the model shows how the plant performance drifts in time.MVA not only provides more insight into raw material quality variances, but allows users to combine spectral and wet chemistry data, and to model the relationship between raw material and final product quality. It also allows users to address the complexities of batch monitoring, including:
  • Time variance;

  • Mixed data, including initial conditions, process data and final product quality data;

  • The many unit operations, including granulation, drying, compression, and coating, that are responsible for final product quality.
A hierarchical approach is taken. Online processing steps are first modeled and monitored, then the complete batch, including interactions between raw materials and processing steps, is modeled and correlated to quality. The resulting trend lines:
  • Summarize normal variation in process parameters;

  • Track the status and evolution of the batch;

  • Allow users to distinguish between in- and out-of-control operations via control limits on summary variables (Figure 3, below).
By themselves, the multivariate summary statistics are abstract. Contributions are provided to depict which variables are responsible for a change in the MVA space and connect these multivariate metrics to process parameters to diagnose and remove process faults (Figure 4, below).Initially, the modeling process is challenged by IT infrastructure and data storage issues, as well as the many forms that data can take, whether they come from static wet chemistry operations, univariate or advanced multivariate sensors.Data must first be extracted from several repositories, whether LIMS, SCADA, MES or ERP platforms. They must then be aligned and analyzed to identify critical parameters, calibrate process signatures, build models for quality metrics and understand process variations.In a feasibility study with Novartis MVA analysis was used to help understand variance in dissolution between batches. Models were developed using historical data for a secondary processes like granulation and drying, all of which were carried out in batch mode. For each process, a set of variables (with K between 7 and 50) was measured on the process step from beginning to end, at regular intervals. In addition, “one shot” data (initial conditions and seven key raw material characteristics) and final batch quality were measured. The results were analyzed for two years worth of production, some 314 batches.The facility then developed an integrated knowledge of their manufacturing operation expanding the analysis to include all process steps and ingredients. Quality metrics such as dissolution rate were predicted by the model and compared with traditional offline analysis. Online monitoring of the complete production cycle was implemented for fault detection and process diagnosis. The challenge in creating this integrated analysis and monitoring environment was the building of an IT infrastructure connecting and more importantly organizing the many data sources of varying data formats (Figure 5, below).No sensors were added, and the exercise improved process understanding and developed multivariate signatures for documentation and communicating process behavior, building a foundation for real-time release.It is widely understood that PAT involves measuring and quantifying the influence of manufacturing on pharmaceutical product. We see the integration of new sensors for calibration of metrics indicative of quality without a clear vision or strategy of the appropriate use of the data generated. This case study demonstrates a real-life example of how data collection systems are used for execution of a PAT strategy, including improving process understanding, manufacturing efficiency and product quality.PAT programs that focus on the development of analytics often do not align well with execution of a business strategy and demonstrate limited value. Experience has shown that a clearly defined strategy that sits above the development of technology is essential. PAT programs driven from a manufacturing perspective focused on utilization of technology for the achievement of a business goal have the proper foundation to ensure success.
Figure 1. The univariate specifications presently used to characterize raw materials cannot adequately describe their quality or influence on the final product, often allowing problems to go undetected.

Figure 2. By projecting the 33-dimensional space onto a 2-dimensional plane, the model shows how the plant performance drifts in time.

Figure 3. MVA allows users to distinguish between in- and out-of-control operations via control limits on summary variables.

Figure 4. When you determine which variables are responsible for a change in the MVA space, you can connect these multivariate metrics to process parameters in order to diagnose and remove process faults.

Figure 5. This diagram shows the challenges involved in creating an integrated analysis of the complete production cycle.

About the Author

Chris McCready | Umetrics