P305: Utilization of Artificial Intelligence (AI) by African Regulators for Risk-Based Assessment of CMC: Proof of Concept
Poster Presenter
Lorraine Danks
Senior Program Officer
Gates Foundation South Africa
Objectives
Many African regulators are challenged to adhere to their review timelines resulting in increasing backlogs; therefore, the aim of this study was to develop an AI Proof-of-Concept (AI PoC) tool to automate a risk-based assessment (RBA) framework to expedite the CMC evaluation of generic medicines.
Method
A survey was conducted among 11 African medicines regulators (East, West & South) to assess their desire for adopting AI tools. Following strong interest, an AI PoC for CMC RBA for generics was co-developed with the South African regulator and subsequently also validated within the Botswana agency.
Results
The results from the survey showed that 5 out of the 11 (46%) national regulatory authorities (NRAs) indicated that they have explored or are currently introducing AI tools into their workflows, with 3 more (27%) in the planning phase. Many of these are already generating PoC tools utilizing large language models (LLMs), with a focus on leveraging AI to optimize the time-consuming administrative processes during the initial product review and authorization. There was wide acknowledgement that bespoke, regulatory-specific algorithms could expedite access to medicines in their jurisdictions.
The South African Health Products Regulatory Authority (SAHPRA) expressed interest in automating its manual RBA framework to further reduce the CMC assessment time of generic products. As a result, an AI PoC tool (named LEXI) was developed based on SAHPRA’s RBA quality matrix, using 210 product dossiers, and was validated under the Good Automated Manufacturing Practice (GAMP5) framework. The system employed Retrieval-Augmented Generation architecture with Meta’s Llama 3, supported by regulatory data integrations from global reference institutions. Validation comprised three qualification phases - installation (IQ), operational (OQ), and performance (PQ) - and was conducted within SAHPRA and the Botswana Medicines Regulatory Authority (BoMRA).
During the external validation across 60 dossiers, the tool demonstrated 91.7% overall predictive accuracy compared with expert human assessors (97% sensitivity, 86% specificity). Criterion-level accuracy exceeded 95% for most critical quality attributes and LEXI reduced assessment time by approximately 91% (1.5 hours versus 16 hours for a manual review) while maintaining full traceability and regulatory compliance. Further tool adaptations allowed effective operation across different dossier structures and local regulatory environments, confirming its interoperability and robustness.
Conclusion
The validation of LEXI demonstrates the feasibility of integrating secure, explainable AI into regulatory workflows in low- and middle-income countries. LEXI can automatically extract, interpret, and summarize key dossier content while providing intelligent risk stratification to support prioritization. By combining natural language processing, document intelligence, and regulatory data integration, the tool delivers structured, explainable risk analyses and can efficiently process 92% of dossiers. This results in significant time savings, using the algorithm, depending on product complexity. By identifying high-risk areas in generic product applications and directing assessors to relevant critical quality attributes through concise summaries, LEXI enables a risk-based review approach that can avoid full dossier assessments. This is particularly impactful in Africa, where over 90% of NRA pipelines consist of generic applications and backlogs are widespread. LEXI, therefore, offers a practical means to accelerate reviews without compromising medicine quality, safety, or efficacy.
LEXI represents a validated proof of concept for AI-assisted, risk-based regulatory assessment in resource-constrained environments. Its development illustrates how advanced LLMs can be adapted for use in tightly controlled regulatory contexts, while maintaining essential human oversight and validation by expert assessors. The tool’s demonstrated efficiency and accuracy support its potential scalability across African NRAs, with broader adoption likely to improve regulatory transparency, reproducibility, and responsiveness as well as contributing to progress toward Universal Health Coverage. These findings align with global regulatory trends by mature regulators, emphasizing AI’s role in improving regulatory agility while maintaining scientific rigor.
