(141c) Data-Driven Real-Time Operation Support for Decontamination Processes in Biopharmaceutical Drug Product Manufacturing

Manufacturing and process control tools have been continuously evolving and becoming more extensive and sophisticated. An ever-increasing number of sensors are constantly monitoring production processes and providing real-time measurements of process variables such as pressure, humidity, and temperature. This large data inventory is mostly unused except for backtracking purposes in the case of an incident occurring or for good manufacturing practice (GMP) related revalidation efforts. This data could, however, be highly valuable for anticipating failures and for predictive maintenance applications. This work describes a framework for further utilizing this data. It presents an industrial case study for the application of the framework within the decontamination process of the sterile manufacturing of biopharmaceutical drug products at the Hoffmann–La Roche facility, located in Kaiseraugst, Switzerland. Decontamination is performed after each product batch in the manufacturing environment such as clean rooms or isolators to kill any micro-organisms. The process includes dehumidification, conditioning, decontamination with hydrogen peroxide, and aeration.

Our suggested framework allows for probability-based performance evaluation and forecasting of processes in commercial operation. Application of the framework would enable operators to predict batch failures and to take pre-emptive action. The framework employs machine learning algorithms within a three-step procedure to classify the data based on batch quality, train a model to make predictions and finally a decision-making step based on the results of the previous two steps.

The first step of the procedure involves statistically defining quality boundaries using pre-recorded data, e.g., temperature, pressure, hydrogen peroxide content. Due to the high correlation of the recorded data, a principal component analysis is used to obtain a reduced set of independent variables (principal components). The results are used to define batch quality boundaries. The recorded data is then split into training and test data for a machine learning algorithm. Several algorithms such as k-nearest neighbors and random forest are compared to find the most suitable algorithm for the studied case data. The trained model should then predict process parameters for the remaining process duration based on the input from only a few seconds of measured operation parameters. The third step of the procedure provides a predictive monitoring tool which combines the previously defined operation boundaries with the predictions from the specified machine learning algorithm. The tool would allow for predicting failure probabilities along with the expected failure time. A reverse PCA is then performed to pinpoint critical process parameters. This allows operators to take precautionary preventive measures, e.g., adjustment of the pumps and valves or interruption of the process.