Retrospective Soft-Sensor Assessment of Chinese Hamster Ovary Cultures Using Campaign-Grouped Cross-Validation

Background: Retrospective estimation of viable cell density (VCD) from bioreactor signals requires careful time alignment and campaign-held-out testing. Objective: The study aimed to evaluate the association of seven concurrent inline sensor signals with CHO-K1 VCD estimation and to assess a separate, reduced growth model on external aggregate trajectories. Methods: I reanalyzed 378 offline observations and 274,781 inline records from 24 CHO-K1 cultivations in 12 campaigns. A seven-signal estimator underwent post-hoc nested leave-one-campaign-out assessment in nine batch campaigns. Separately, CHO-K1-trained early-growth coefficients were applied without refitting to three NISTCHO configuration means from one independent study, with observed day-zero VCD used for initialization. Results: Complete seven-signal matches were available for 350 offline observations. Across the nine batch campaigns, campaign-averaged mean absolute error (MAE) decreased from 1.346 x 10⁶ cells/mL with time-only splines to 0.956 x 10⁶ cells/mL with time plus seven signals (28.9% exploratory reduction), but early-period gains were less consistent. In the separate cross-dataset transfer assessment, mean MAE was 0.382 versus 0.800 x 10⁶ cells/mL for the constant baseline across days 1-3; the transferred model performed slightly worse than the constant baseline for pDs. Conclusions: Benefits were retrospective and context dependent. The original seven-sensor estimator was not validated externally, and neither analysis demonstrates prospective forecasting or industrial control.