AI-integrated IQPD framework of quality prediction and diagnostics in small-sample multi-unit pharmaceutical manufacturing: Advancing from experience-driven to data-driven manufacturing

This is the authors' abstract. We don't add key points for this paper.

The pharmaceutical industry faces challenges in quality digitization for complex multi-stage processes, especially in small-sample systems. Here, an intelligent quality prediction and diagnostic (IQPD) framework was developed and applied to Tong Ren Tang's Niuhuang Qingxin Pills, utilizing four years of data collected from four production units, covering the entire process from raw materials to finished products. In this framework, a novel path-enhanced double ensemble quality prediction model (PeDGAT) is proposed, which combines a graph attention network and path information to encode inter-unit long-range and sequential dependencies. Additionally, the double ensemble strategy enhances model stability in small samples. Compared to global traditional models, PeDGAT achieves state-of-the-art results, with an average improvement of 13.18% and 87.67% in prediction accuracy and stability on three indicators. Additionally, a more in-depth diagnostic model leveraging grey correlation analysis and expert knowledge reduces reliance on large samples, offering a panoramic view of attribute relationships across units and improving process transparency. Finally, the IQPD framework integrates into a Human-Cyber-Physical system, enabling faster decision-making and real-time quality adjustments for Tong Ren Tang's Niuhuang Qingxin Pills, a product with annual sales exceeding 100 million CNY. This facilitates the transition from experience-driven to data-driven manufacturing.