Calibration of multivariate predictive models the study of factors influencing the prediction ability of Raman spectroscopy applied to pharmaceutical tablets
The use of Process Analytical Technology (PAT) is nowadays well spread throughout the pharmaceutical industry, mostly because they allow gaining useful process insight, better monitoring throughout the manufacturing steps, lowering costs associated with the use of wet-chemistry based testing method, and increasing overall efficiency. Most of the PAT methods are based on multivariate predictive models (MVPM).The calibration of such MVPM is a step crucial to the success of the MVPM-based PAT method development. From what could be gathered of today's practices in MVPM calibration, the current trends are aligned on the mantra"the more the better". This means that best-practices suggest that using more calibration points, with more samples, manufactured as closely as possible to the commercial samples, with more precise measuring reference methods, will lead to a more precise, better performing MVPM-based PAT methods. Obviously, it also leads to a costlier and longer to develop method, thus limiting the potential applications as well as lowering the overall gain to expense ratio of these methods.The aim of this work was to investigate the development protocol of MVPM-based PAT methods in order to identify the factors that were most susceptible to influence the performances of the developed method, with the ultimate goal to propose an optimized protocol. A base case of MVPM-based PAT method was used to perform the assessment: the use of Raman spectra collected from intact pharmaceutical tablets to predict their content in 4 minerals and 1 vitamin. From this development case study, two minerals were chosen to be part of the investigation: one that was in high concentration in the product and one that was in low concentration. Eight (8) factors were identified and tested separately on the two data sets created.The tests were guided by a design of experiment that allowed optimizing the total number of predictive models that needed to be made, all the while retaining sufficient data to allow minimal confusions within second-order factorial interactions. Analysis of results gathered from more than 85 multivariate predictive models allowed identifying factors that had an influence on both type of calibration (high and low raw material concentration).The type of equipment and batch size used for calibration samples manufacturing, as well as the number of points and replicates use (to a certain extent) or the calibration algorithm and reference method did not influence the MVPM accuracy when dealing with highly concentrated raw material. For a low-concentration raw material, the significant factors were the type of equipment and number of calibration points used. In both cases, the distribution of the calibration points along the chosen concentration span showed a significant influence on the method performances. Having identified the factors influencing the MVPM-based PAT method, an optimized protocol for future development could be proposed. Using this alternate approach should reduce the required invested man- and lab-time by approximately 30 %. A 31 % reduction in development-related costs is also expected. These results can be achieved with a manageable accuracy decrease varying between 0 and less than 2 % (absolute values), depending on the type of raw material concentration monitored. This study fills a gap in the published literature. No extensive and comprehensive studies had been made, to the author's knowledge, regarding the overall development process of such MVPM-based methods. Several hints for future research and developments are also proposed to follow-up with this project.
- Génie – Mémoires