Model maintenance: the unrecognized cost in PAT and QbD
Multivariate calibration, classification and fault detection models are ubiquitous in QbD (Quality by Design) and PAC and PAT (Process Analytical Chemistry and Technology, respectively) applications. They occur in both the development of processes and their permissible operating limits, (i.e. models for relating the process design space to product quality), and in manufacturing (i.e. models used in monitoring and control). Model maintenance is the on-going servicing of these multivariate models in order to preserve their predictive and diagnostic abilities. It is required because of changes to either the sample matrices or the instrument response. The goal of model maintenance is to sustain or improve models over time and react to changing conditions with the least amount of cost and effort. A model maintenance roadmap is presented. It includes procedures for determining when model maintenance is required, the probable source of the model/data mismatch, and the best approaches for bringing model performance back to acceptable levels.
Model maintenance can be roughly defined as the on-going upkeep of (primarily) multivariate calibration and fault detection models in order to preserve their predictive and diagnostic abilities. The goal of model maintenance is to preserve or improve models over time and address changing conditions with the least amount of cost and effort. As noted in ASTM E2891-13 (1), model maintenance should be considered a core activity of the overall modelling activity in pharmaceutical and manufacturing operations. Regrettably, in spite of this, model maintenance is often not planned for in advance in QbD and PAT applications. Instead, it is frequently assumed that models will work indefinitely. However, it is an unfortunate fact that seemingly innocuous chang