Application Constraints of Linear Multivariate Regression Models for Dielectric Spectroscopy in Inline Bioreactor Viable Cell Analysis

In the context of data-driven bioprocess modeling, selecting appropriate regression models remains a critical challenge. This is especially the case when dealing with time-dependent process dynamics and complex measurement data. The practical relevance of this study lies in its critical assessment of the application constraints associated with multivariate linear regression models in bioprocess monitoring of cell culture processes. The applicability of Partial Least Squares and Ridge regression was evaluated for different cultivation phases. The results emphasize that no single linear modeling approach is universally suitable for capturing the complex behavior of mammalian cell cultures. This is why we present an enhanced segmented modeling approach by learning the optimal transition point from data and introducing a gradual model switch, allowing for smoother and more robust adaptation to process dynamics. This segmented model led to improved predictive performance compared to single-model regression across the entire process duration. Nevertheless, the heterogeneity of the 11 mammalian cell culture datasets used in this study posed significant challenges, with the best-performing models achieving prediction error of around 0.31 of the average offline viable cell density. These results underline the potential of phase-adaptive modeling, while also emphasizing the need for further optimization to robustly handle diverse bioprocess conditions.