Multi-block tensor regression for quality prediction and root cause analysis in the production of active pharmaceutical ingredients

Abstract

The large scale production of active pharmaceutical ingredients (APIs) is traditionally accomplished via batch processes. Nevertheless, their inherent complexity has limited the development and application of models for the processes as well as the use of advanced online monitoring, control and optimization strategies for their continuous improvement. The quality by design (QbD) strategy, defined by regulatory agencies, has brought a practical view into this. QbD seeks to determine methods to define the critical quality attributes (CQAs) of the product in terms of the critical material attributes (CMAs) and the critical process parameters (CPPs) of the input space along the whole process. This means not only on individual batch units but also throughout the multiple steps and stages of the production process. In this contribution, data-driven modelling methods were exploited to model two synthesis steps in the large scale production of an Active Pharmaceutical Ingredient (API). First, tensor factorization was applied to train correlation based models to capture the main directions of variability for each of the studied steps. Secondly, a partial least squares (PLS) model was trained to regress the concentration of an impurity on the product onto the latent variables of the primary monitoring models as well as the CMAs of the input materials. The proposed modelling approach was applied to data from the large scale production on an API. This resulted in an accurate model, which was validated on an independent data set and which captures meaningful correlations that helped on the root cause identification for variations encountered in the CQA of the product. These results are in line with the observations made on the process operation.