A heuristic approach to handling missing data in biologics manufacturing databases

Area of Science:

• Biotechnology
• Data Science
• Process Engineering

Background:

• Biologics manufacturing generates vast historical datasets.
• Inter-batch variability and missing data are common due to diverse experiments and technical issues.
• Data pre-processing is crucial for accurate analysis and predictive modeling.

Purpose of the Study:

• To investigate the efficiency of mean imputation and multivariate regression for handling missing data in biomanufacturing datasets.
• To evaluate the performance of these imputation methods using symbolic regression and Bayesian non-parametric models.
• To identify key factors influencing the selection of appropriate imputation techniques.

Main Methods:

• Comparison of mean imputation and multivariate regression imputation.
• Application of symbolic regression models for data processing.
• Utilization of Bayesian non-parametric models for evaluating imputation performance.
• Analysis of missing data mechanisms (Missing Completely At Random, Missing At Random, Missing Not At Random).

Main Results:

• Mean imputation is efficient for smooth, non-dynamical datasets.
• Regression imputation effectively preserves data distribution and standard deviation for dynamical datasets with <30% missing data.
• The nature of missing data (MCAR, MAR, MNAR) is critical for method selection.

Conclusions:

• The choice of imputation method depends on dataset characteristics and the nature of missing data.
• Appropriate pre-processing enhances the reliability of data mining in biomanufacturing.
• Understanding missing data mechanisms is key to effective data imputation and predictive modeling.