A Machine Learning Platform for Imputing Missing Critical Quality Attributes in Pharmaceutical Development

Abstract

Quality control is fundamental to drug development, with critical quality attributes (CQAs) serving as key indicators of product performance. However, missing or incomplete CQA data remains a persistent challenge throughout pharmaceutical research and manufacturing. In this study, we evaluated the utility of machine learning techniques (MLT) for CQA imputation across five pharmaceutical datasets comprising 14 CQAs. Five MLT were benchmarked against conventional imputation methods to assess their imputation performance. In several datasets and CQAs, MLT achieved imputation accuracies exceeding R2 = 0.80, including under conditions of up to 30% simulated missingness. Nevertheless, performance varied substantially across CQAs and datasets, highlighting the critical influence of data quality, feature informativeness and source heterogeneity on imputation outcomes. Overall, ML-based methods outperformed several traditional approaches in the majority of evaluated scenarios, although reduced performance was observed for selected complex or weakly informative CQAs. To facilitate practical adoption, we developed a web-based software platform enabling users to apply and evaluate ML-based imputation workflows on their own datasets. These findings indicate that ML-based imputation methods can complement conventional techniques for managing incomplete CQA data and may enhance data integrity when applied under appropriate conditions. Collectively, this work supports the potential of ML-assisted imputation as a component of data-driven decision-making frameworks in pharmaceutical development. Graphical Abstract