Estimation of design space for an extrusion-spheronization process using response surface methodology and artificial...

Area of Science:

• Pharmaceutical Development
• Chemical Engineering
• Data Science

Background:

• Quality by Design (QbD) principles are crucial in pharmaceutical development, emphasizing mathematical modeling of critical factors and attributes.
• Current methods for defining the process design space (PDS) using response surface methodologies (RSM) can lead to uncertainties due to polynomial model inaccuracies, especially at PDS edges.
• Artificial neural networks (ANNs) are often used to enhance RSM, but comparative studies on different experimental design layouts are lacking.

Purpose of the Study:

• To investigate the impact of various experimental design (DoE) layouts on model predictability.
• To compare the sensitivity of different modeling approaches based on experimental data organization.
• To evaluate the influence of DoE layout on the calculated process design space (PDS) size and position.

Main Methods:

• Comparison of multiple DoE layouts: 2-level full factorial, Central Composite, Box-Behnken, 3-level fractional, and 3-level full factorial.
• Application of response surface methodologies (RSM) and artificial neural network (ANN) based models for data analysis.
• Assessment of model predictability and sensitivity concerning experimental data set organization.

Main Results:

• The size of the calculated PDS can vary by over 40% depending on the polynomial model used, with higher-level layouts causing significant shifts in PDS position, particularly with RSM.
• ANN-based models demonstrated superior predictability compared to traditional RSM polynomial models.
• Both modeling methods showed considerable sensitivity to the organization of the experimental data set.

Conclusions:

• The choice of experimental design layout significantly impacts the accuracy and reliability of the process design space (PDS) determination in pharmaceutical development.
• Artificial neural networks offer improved model predictability for defining the PDS compared to response surface methodologies.
• Experimental designs that better represent extreme factor values are recommended for robust PDS calculations and enhanced model sensitivity.