Development of A Chromatographic Database for Applying Machine Learning for Drug Analysis

Abstract

This study introduces a machine learning-based framework to optimise chromatographic data analysis in pharmaceutical quality assurance. We developed a chromatographic peak database and trained an XGBoost model— widely adopted in industry for its scalability and robustness across diverse data ranges—to predict the acceptability of chromatograms. The dataset included 100 CSV files representing real-world chromatographic waveforms, from which we extracted features such as retention time, peak height, and peak width for training. Our model achieved 65% accuracy, 3% precision, 38% recall, and an average RMSE, highlighting ML's potential to augment current manual quality assessment workflows. No benchmark model existed for this application; therefore, our comparative analysis with the Random Forest model provided the only available baseline. Unlike traditional, time-consuming manual reviews—where quality analysts may evaluate up to 100 chromatograms daily, each taking approximately five minutes—our predictive system streamlines the process by flagging only non-compliant cases, effectively reducing human workload by over 90%. This innovation lays foundational work for AI-driven chromatography, especially in areas where literature integrating ML with chromatographic analysis remains sparse.