The pharmaceutical industry was for a long time founded on rigid rules. With the new PAT initiative, control is becoming significantly more flexible. The Food and Drug Administration is even encouraging the industry to use methods like machine learning. The authors designed a new data mining method based on inducing ensemble decision trees from which rules are generated. The first improvement is specialization for process analysis with only a few examples and many attributes. The second innovation is a graphical module interface enabling process operators to test the influence of parameters on the process itself. The first task is creating accurate knowledge on small datasets. The authors start by building many decision trees on the dataset. Next, they subtract only the best subparts of the constructed trees and create rules from those parts. A best tree subpart is in general a tree branch that covers most examples, is as short as possible and has no misclassified examples. Further on, the rules are weighed, regarding the number of examples and parameters included. The class value of the new case is calculated as a weighted average of all relevant rule predictions. With this procedure the authors retain clarity of the model and the ability to efficiently explain the classification result. In this way, overfitting of decision trees and overpruning of the basic rule learners are diminished to a great extent. From the rules, an expert system is designed that helps process operators. Regarding the second task of graphical interface, the authors modified the Orange explanation module so that an operator at each step takes a look at several space planes, defined by two chosen attributes (Demšar et al., 2004). The displayed attributes are the ones that appeared in the classification rules triggered by the new case. The operator can interactively change the current set of process parameters in order to check the improvement of the class value. The task of seeing the influence of combining all the attributes leading to a high quality end product (called design space) is now becoming human comprehensible, it does not demand a highdimensional space vision any more. The method was successfully implemented on data provided by a pharmaceutical company. High classification accuracy was achieved in a readable form thus introducing new comprehensions.
TopIntroduction
Donald E. Knuth said almost twenty years ago (D. Knuth, interview, 1993): “I think the most exciting computer research now is partly in robotics, and partly in applications to biochemistry. Biology is so digital, and incredibly complicated, but incredibly useful.”
In 2004, the United States Food and Drug Administration (FDA) issued a document “PAT — A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance.” This document was written as guidance for a broad industry audience in different organizational units and scientific disciplines (“Guidance for industry PAT”, 2004). To a large extent, the guidance discusses principles with the goal of developing regulatory processes of drug production that encourage innovation.
As the FDA states, the conventional pharmaceutical manufacturing is generally accomplished using batch processing with laboratory testing conducted on collected samples to evaluate quality (“Guidance for industry PAT”, 2004). This conventional approach has been successful in providing quality pharmaceuticals to the public. However, today significant opportunities exist for improving pharmaceutical development, manufacturing, and quality assurance through innovation in product and process development, process analysis, and process control. This is where machine learning might help.
Unfortunately, the pharmaceutical industry has been generally hesitant to introduce innovative systems into the manufacturing sector for a number of reasons. One often cited reason is regulatory uncertainty, which may result from the perception that the existing regulatory system is rigid and unfavorable for the introduction of innovative systems. For example, many manufacturing procedures are treated as being frozen and many process changes are managed through regulatory submissions.
Because of the hesitancy of the pharmaceutical industry, the document encourages new production techniques with common name Process Analytical Technology (PAT). Its focus is innovation in development, manufacturing and quality assurance by removing “regulatory fear/uncertainty”, utilizing science and risk-based approach to regulatory requirements and oversight. This will provide a flexible and less burdensome regulatory approach for well understood processes, creating an environment that facilitates rationale science, risk, and business decisions.
Therefore, the pharmaceutical industry needs a system for designing, analyzing and controlling manufacturing process (Schneidir, 2006). The goal of PAT is to understand and control the manufacturing process in real time. The system must be able to make the recommendation during the process to achieve higher quality of the end product. The system should follow the performance attributes, raw and in-process materials and processes. A PAT system should use multiple tools for understanding and controlling the manufacturing process: multivariate tools for design, data acquisition and analysis, process analyzers, process control tools, continuous improvement and knowledge management tools.
The FDA expects an inverse relationship between the level of process understanding and the risk of producing a poor quality product. The well understood process will require less restrictive regulatory approaches to manage change.
If enough data were gathered, PAT suggests constructing a design space (Desai, 2006). That is a multi-dimensional space that encompasses combinations of product design, manufacturing process design, critical manufacturing process parameters and component attributes that provide assurance of suitable product quality and performance. Therefore, design space is part of the space for which there is data (usually in a specific interval). Control space is defined similarly. It is the multi-dimensional space that encompasses process operating parameters and component quality measurements that assure process or product quality. It is a subset of the design space, as can be seen in Figure 1. It is considered as a “high quality area” for process parameters.
Figure 1. Design and control space. Control strategy is to adjust the parameters of the points in design space to get into control space.
The control strategy is to mitigate risks associated with the batch failure when the critical and non-critical process parameters fall outside the control space but stay within the design space.