(75b) Adaptive Strategies for Updating Unit Operation Models and in-Line Monitoring of Blend Uniformity in Continuous Pharmaceutical Manufacturing Process

Guided by U.S. Food and Drug Administration (FDA)’s Quality-by-Design (QbD) initiative^1,2 and market incentives³, the pharmaceutical industry has heavily invested in the transition of tablet production from a traditional batch operating mode to continuous pharmaceutical manufacturing (CPM) over the past decade. The benefits of employing CPM, including but not limited to the smaller footprint, reduced human impacts, increased controllability, and shortened release time, have been clearly identified⁴. To better understand this emerging process and perform quality assessments, different mathematical models for unit operations and in-line process analytical technology (PAT) tools have been developed^4,5. On the unit operation level, depending on the availability of process knowledge and the extent of data incorporation, these models can be classified as: (1) mechanistic models that strongly rely on process understanding; (2) data-driven models which depend only on the process data; and (3) hybrid models that integrate the first two categories^6,7. In terms of the in-line PAT tools which monitor product quality, the performance relies heavily on data-driven global calibration models, such as partial least squares (PLS) regression⁸.

The development and implementation of data-driven and hybrid models have a well-known challenge related to model maintenance. Since these models are typically trained based on historical data, they can only reflect the system for the range of process conditions, material properties, and environmental variables present at the time of model building⁹. However, shifts in the CPM process can happen over time due to changes such as environmental conditions, equipment wear and tear, lot-specific material properties, operating conditions, etc. These changes can happen suddenly, gradually, incrementally, or in a pattern¹⁰, impacting performance of unit operations and ultimately, the critical quality attributes (CQAs) of a product. It is noted that these changes are usually not captured in typical training datasets, leading to deteriorating model prediction accuracy¹⁰. Therefore, appropriate model maintenance strategies need to be established. For unit operation models that involve data-driven components, model maintenance can ensure accurate representation of the manufacturing process and allow dynamic system analyses. For PAT-based chemometric models, maintenance is important for the calibration models to be robust and transferable^11,12, but traditional maintenance strategies are generally complex, costly, and time-consuming^8,11,13. For example, depending on the diagnostics of a reduced model performance, more calibration data may need to be added, requiring additional experiments to be performed, consuming time and financial resources.

In this work, adaptive modeling strategies are developed and assessed, to serve as a model maintenance tool to cope with non-linear time-varying changes in process characteristics and/or operating conditions. The developed hybrid unit operation models for the continuous direct compaction (CDC) line^4,14 are used as baseline models. Adaptive windowing¹⁵ and ensembled adaptive windowing¹⁵ are selected as adaptive algorithms to be embedded into the data-driven components of the base cases. In addition to using inputs at specific time points for prediction, these algorithms incorporate change detection mechanisms, which monitor the process data streams to identify any process changes. Given such information, the algorithms can then update the data-driven components of the models. The prediction performances of the resulting adaptive models are compared against the base models in cases involving changes in process conditions, material properties, and design parameters. In terms of the PAT-based chemometric models, performance of locally weighted PLS (LW-PLS)¹⁶, moving window¹⁵, and recursive PLS (RPLS)¹⁷ are evaluated in comparison to regular PLS for in-line monitoring of low dose blends in the feed frame of a tablet press using near infra-red (NIR) spectroscopy. Performance assessment involves a cross-prediction between datasets comprising of NIR spectra collected at different feed frame speeds with physical baseline differences in the NIR spectra. Cross-prediction can help to evaluate the effectiveness of adaptive methods to prevent the anticipated deterioration of prediction accuracy. Case studies involving time-varying process condition changes are also developed for comparison. The results of adaptive unit operation and chemometric models demonstrate a more accurate prediction performance in systems with changes in operating conditions. Using these adaptive modeling strategies, the models can be updated in real-time or periodically to enhance the predictability of CQAs throughout the process development and operation phases¹⁸.

References

(1) Yu LX, Akseli I, Allen B, et al. Advancing Product Quality: a Summary of the Second FDA/PQRI Conference. AAPS J. 2016;18(2):528-543.

(2) Yu LX, Kopcha M. The future of pharmaceutical quality and the path to get there. Int J Pharm. 2017;528(1-2):354-359.

(3) Boukouvala F, Niotis V, Ramachandran R, Muzzio FJ, Ierapetritou MG. An integrated approach for dynamic flowsheet modeling and sensitivity analysis of a continuous tablet manufacturing process. Comput Chem Eng. 2012;42:30-47.

(4) Wang Z, Escotet-Espinoza MS, Ierapetritou M. Process analysis and optimization of continuous pharmaceutical manufacturing using flowsheet models. Comput Chem Eng. 2017;107:77-91.

(5) Ierapetritou M, Muzzio F, Reklaitis G. Perspectives on the continuous manufacturing of powder-based pharmaceutical processes. AIChE Journal. 2016;62(6):1846-1862.

(6) Zendehboudi S, Rezaei N, Lohi A. Applications of hybrid models in chemical, petroleum, and energy systems: A systematic review. Applied Energy. 2018;228:2539-2566.

(7) von Stosch M, Oliveira R, Peres J, Feyo de Azevedo S. Hybrid semi-parametric modeling in process systems engineering: Past, present and future. Comput Chem Eng. 2014;60:86-101.

(8) Wise BM, Roginski RT. A Calibration Model Maintenance Roadmap. IFAC-PapersOnLine. 2015;48(8):260-265.

(9) Webb GI, Hyde R, Cao H, Nguyen HL, Petitjean F. Characterizing concept drift. Data Mining and Knowledge Discovery. 2016;30(4):964-994.

(10) Gama J, Žliobaitė I, Bifet A, Pechenizkiy M, Bouchachia A. A survey on concept drift adaptation. ACM Computing Surveys. 2014;46(4):1-37.

(11) Zeaiter M, Roger JM, Bellon-Maurel V, Rutledge DN. Robustness of models developed by multivariate calibration. Part I. TrAC Trends in Analytical Chemistry. 2004;23(2):157-170.

(12) Besseling R, Damen M, Tran T, et al. An efficient, maintenance free and approved method for spectroscopic control and monitoring of blend uniformity: The moving F-test. J Pharm Biomed Anal. 2015;114:471-481.

(13) Zhong L, Gao L, Li L, Zang H. Trends-process analytical technology in solid oral dosage manufacturing. Eur J Pharm Biopharm. 2020;153:187-199.

(14) Chen Y, Ierapetritou M. A framework of hybrid model development with identification of plant‐model mismatch. AIChE Journal. 2020.

(15) Kadlec P, Grbić R, Gabrys B. Review of adaptation mechanisms for data-driven soft sensors. Comput Chem Eng. 2011;35(1):1-24.

(16) Hazama K, Kano M. Covariance-based locally weighted partial least squares for high-performance adaptive modeling. Chemometrics and Intelligent Laboratory Systems. 2015;146:55-62.

(17) Joe Qin S. Recursive PLS algorithms for adaptive data modeling. Comput Chem Eng. 1998;22(4-5):503-514.

(18) Gama J, Medas P, Castillo G, Rodrigues P. Learning with Drift Detection. 2004; Berlin, Heidelberg.