Development and validation of an artificial intelligence proof-of-concept tool for risk-based quality assessment of generic medicines: a South African case study.

Introduction: Regulatory authorities in African countries face persistent challenges in ensuring the timely authorization of quality generic medicines due to limited technical capacity and extensive review timelines. The South African medicines regulator (SAHPRA) piloted a risk-based assessment (RBA) framework to expedite the Chemistry, Manufacturing, and Controls (CMC) evaluation of generic medicines. Building on this initiative, an artificial intelligence (AI)-driven tool, termed LEXI, was developed and validated to automate SAHPRA's RBA triage for generic medicines and was externally validated at two African NRAs. Methods: LEXI was designed based on SAHPRA's established RBA quality matrix, using 210 of its product dossiers, and was validated under the Good Automated Manufacturing Practice (GAMP5) framework. The system employed Retrieval-Augmented Generation (RAG) architecture with Meta's Llama 3.1 large language model, supported by regulatory data integrations from the European Directorate for the Quality of Medicines (EDQM) and the World Health Organization Prequalification Programme. Validation comprised three qualification phases-installation (IQ), operational (OQ), and performance (PQ)-and was conducted across two African national regulatory authorities (NRAs): SAHPRA and the Botswana Medicines Regulatory Authority (BoMRA). Results: During external validation across 60 dossiers, LEXI demonstrated 91.7% overall predictive accuracy (55/60; 95% CI: 81.6-97.2%), with sensitivity of 96.9% (31/32; 95% CI: 83.8-99.9%) and specificity of 85.7% (24/28; 95% CI: 67.3-96.0%), with a median 91% time reduction; performance varied by criterion and dossier structure. Criterion-level accuracy exceeded 95% for most critical quality attributes (CQAs). Adaptations allowed for effective operation across different dossier structures and local regulatory environments, confirming its interoperability and robustness. Discussion: The validation of LEXI highlights the feasibility of secure, traceable AI integration into regulatory workflows. The tool augments human expertise by automating risk classification and dossier triage, enabling consistent, evidence-based decision support. These findings align with global regulatory trends at mature authorities, emphasizing AI's role in improving regulatory agility while maintaining scientific rigor. Conclusion: LEXI represents a validated proof-of-concept for AI-assisted, risk-based regulatory assessment in resource-constrained settings. Its demonstrated efficiency and accuracy support its potential to enhance regulatory workflows and reduce review backlogs of generic medicines, pending further validation across a broader range of product types, including complex formulations, and diverse regulatory environments.