Development of a machine learning framework for multivariate monitoring of a continuous direct compression manufacturing process

Pharmaceutical processes are undergoing a transition from traditional batch to continuous operations, in order to increase the manufacturing efficiency, improve the production capability and ensure high product quality. Process Analytical Technology (PAT) tools, derived from the combination of advanced analytical and modelling techniques, play a crucial role in this transition, since they can aid process understanding, reduce costs and time of scale-up and improve consistency, already from the early development stages. The aim of this work is to develop a PAT tool for the in-depth understanding, the product quality monitoring and the multivariate monitoring of a process for the continuous manufacturing of tablets, based on a large amount of data collected during an experimental campaign carried out in an industrial direct compression line. Despite the presence of multiple units working in different time regimes and time windows, and the large variability in the data, it is shown that the application of an unsupervised machine learning technique such as principal component analysis (together with pre-treatments, such as unfolding, scaling and Savitzky-Golay methods) allows to model the important phenomena characterizing all the parts of the process and, at the same time, the quality of the product through Near-InfraRed spectroscopy. Although preliminary information about the normal operating conditions or the specification of the product are not available in the early process development stages, it is proved that the proposed technique, in combination with statistical hypothesis testing, is able to define an effective monitoring model to detect both quality inconsistency and process faults and to diagnose their root cause. Performance of the models are tested on different validation datasets with satisfactory results through data fusion in multi-block approaches to realize a comprehensive monitoring of the system.