(11e) Dissolution Prediction By Process Analytical Technology, Machine Learning and Mathematical Modeling: Toward the Real-Time Release Testing of Pharmaceuticals

Continuous pharmaceutical manufacturing, both at upstream (API production) and downstream (formulation technology) processing, is rapidly gaining interest as a promising way to improve efficiency, reduce operating costs and ensure easier scale-up and shorter time-to-market [1]. However, the risk mitigation of process failure is higher when the unit operations are integrated into one continuous process line as the failure can cumulatively disturb the whole line [2]. To fully exploit the system-wide performance of a continuous manufacturing line, while providing the required product quality, the entire process must be flexibly controlled to compensate the process variations. Consequently, the conversion from batch to continuous manufacturing also entails the necessity of changing the current quality assurance approaches. Addressing these challenges, the Process Analytical Technology (PAT) and Quality-by-Design (QbD) initiatives emerged over the last few years to support the design, analysis and control of critical quality attributes (CQA) and critical process parameters (CPP) of the pharmaceutical processes. These can also lead to the feasibility of real-time release. The real-time release testing (RTRT) methods [3] utilize a combination of measured material attributes and in-process analysis to demonstrate that the product conforms to the quality attributes defined by the regulations, without the need for end-product quality tests.

In this work, the prediction of in vitro dissolution properties of pharmaceutical tablets is studied using PAT data, machine learning and mathematical modeling of the continuous process. In vitro dissolution testing is one of the most substantial analytical methods in the pharmaceutical industry. It is widely applied during the research and development stage, such as for formulation optimization, prediction of in vivo performance or can be used as a surrogate measurement for bioequivalence. Dissolution measurement also serves as a routine quality control (QC) test during the manufacturing process to assess product performance, stability or batch-to-batch variability [4]. In this study, the dissolution of two model pharmaceutical tablet formulations was studied, an immediate release acetylsalicylic acid formulation as well as caffeine tablets with extended-release dissolution properties. The dissolution curves of the formulations were simultaneously influenced by several critical factors: the concentration of the active pharmaceutical ingredient (API), the concentration of the polyethylene oxide, used to create the extended-release matrix, the tableting compression force and the particle sizes of raw materials. Reflection and transmission NIR and Raman spectra were measured on tablets prepared by following experimental designs as well process data were acquired during the manufacturing.

Several modeling approaches were tested and compared to predict the dissolution of the tablets based on the PAT and process data. Partial Least Square (PLS) regression and feedforward backpropagation Artificial Neural Network (ANN) models were developed for the different spectroscopic measurements individually as well as by combining them together. The decrease of the root mean square errors (RMSE) of the models and the improved prediction of the dissolution curves belonging to independent validation tablets indicated a consistently enhanced performance of the ANN models, which could be explained by their capability of modeling non-linearity between the process parameters and dissolution curves. To move forward the real-time release testing approach, the developed dissolution prediction models could also be fitted into a system-wide process model of a continuous manufacturing line, which describes the variation of the identified critical parameters (e.g. API concentration, particle size) throughout the manufacturing process steps. In this way, the effect of mitigation of process parameter variability on the dissolution properties could be studied, which can also serve as a basis of process control.

The results of this work indicate that data-driven models could provide a non-destructive and fast alternative for the current quality control standard, in line with the real-time release initiative. Machine learning also served as a straightforward data fusion method of the PAT data without the need for additional preprocessing steps. The method could significantly increase the efficacy of dissolution testing, advance data processing in the PAT environment and contribute to an enhanced real-time release testing procedure, especially when the dissolution model is incorporated into a system-wide modeling approach.

[1] S.L. Lee, T.F. O’Connor, X. Yang, C.N. Cruz, S. Chatterjee, R.D. Madurawe, C.M.V. Moore, L.X. Yu, J. Woodcock, Modernizing pharmaceutical manufacturing: From batch to continuous production, Journal of Pharmaceutical Innovation, 10 (2015) 191-199

[2] A. Mesbah, J.A. Paulson, R. Lakerveld, R.D. Braatz, Model predictive control of an integrated continuous pharmaceutical manufacturing pilot plant, Organic Process Research & Development, 21 (2017) 844-854.

[3] European Medicines Agency, 2012. Guideline on Real Time Release Testing (formerly Guideline on Parametric Release)

[4] Dressman, Jennifer B., and Johannes Krämer, eds. Pharmaceutical dissolution testing. Boca Raton, FL:: Taylor & Francis, 2005.