AI/ML in Drug Manufacturing: FDA’s discussion paper and PDA’s comments
In February 2023, FDA’s Center for Drug Evaluation and Research (CDER) published a discussion paper titled ‘Artificial intelligence in Drug Manufacturing’.
This discussion paper outlines potential problem areas and gaps in policy development for the application of artificial intelligence (AI) in pharmaceutical manufacturing. With the paper the FDA requested feedback from stakeholders. While the FDA has embraced the technological and manufacturing advancement offered by AI, it has also acknowledged the lack of a proper regulatory framework in certain aspects of its applications in the pharmaceutical sector.
We are pleased to say that Ulrich Köllisch, Associate Partner at GxP-CC, co-led the effort together with Kir Henrici, Founder and CEO at the Henrici Group, on behalf of the Parenteral Drug Association (PDA) to provide response to the questions addressed in the discussion paper.
The commenting team consisted of more than 20 subject matter experts from technical and regulatory fields who are members of the association. They provided responses to the questions posed in the discussion paper, stated below.
The whole comment can be found here: https://www.regulations.gov/comment/FDA-2023-N-0487-0016
Questions Asked by FDA to the Industry
Q1: What types of AI applications do you envision being used in pharmaceutical manufacturing?
Q2: Are there additional aspects of the current regulatory framework (e.g., aspects not listed above) that may affect the implementation of AI in drug manufacturing and should be considered by FDA?
Q3: Would guidance in the area of AI in drug manufacturing be beneficial? If so, what aspects of AI technology should be considered?
Q4: What are the necessary elements for a manufacturer to implement AI-based models in a CGMP environment?
Q5: What are common practices for validating and maintaining self-learning AI models and what steps need to be considered to establish best practices?
Q6: What are the necessary mechanisms for managing the data used to generate AI models in pharmaceutical manufacturing?
Q7: Are there other aspects of implementing models (including AI-based models) for pharmaceutical manufacturing where further guidance would be helpful?
Q8: Are there aspects of the application of AI in pharmaceutical manufacturing not covered in this document that FDA should consider?
In the following we summarize some of the key points addressed in the PDA response
PDA suggested some new areas in pharmaceutical manufacturing that could benefit from the application of AI solutions, such as Natural Language Processing (NLP) for predictive analysis and Digital Twins for process control.
The use of AI technology also raises the importance of Data governance, as the frequency, quantity, and type of generated data increase. PDA recommends implementing robust Data governance and Data Management processes, due to the involvement of training and testing datasets in addition to output data. The proposed measures also include a centralized Electronic Data Lake (EDL), scalable cloud computing solutions and implementing cybersecurity assessments to ensure proper handling of the wide range of data.
In terms of actions from manufacturers, PDA provides recommendations including but not limited to implementing Machine Learning Operations (MLOps), expert audit and third-party assessments for IT suppliers, Quality management systems (QMS), sound risk assessment procedures to assess the model robustness and a real time monitoring of the model.
For validation practices, PDA suggests the use of high-quality training data sets, clear definition of the intended use of the model, description of the rationale for algorithm selection and clear documentation to ensure traceability of all data (raw data, final data) as well as handling procedures including data transformation. In addition, PDA suggests implementing periodic or real-time monitoring framework to check the model’s functional fitness.
PDA recommends that FDA should provide guidance on validation and qualification approaches including which aspect of the model should be validated, data management for models using IOT (Internet of Things) as well as recommendations to ensure process control and change control management with self-learning models. PDA also suggests that manufacturers could also benefit from guidance on different aspects such as: applicable criteria in data preparation and cleaning, algorithm selection, lifecycle monitoring and traceability and auditability of the involved steps.
PDA’s response also brings attention to some key but in terms of regulations less addressed aspects of AI applications. It is evident that human subjectivity might be in contrast with AI generated outputs and manufacturers could benefit from clearer guidance on human-aided AI systems. For continuously learning AI models, clarity on model monitoring and maintenance, and guidance on regulatory submissions/amendments for variation of manufacturing processes and changes driven by algorithms would be beneficial to the pharmaceutical manufacturers.
In addition to the above-mentioned recommendations, PDA insists that a robust training program should be implemented for involved personnel in the manufacturing unit, auditors and certifying bodies, specifically in case of continuous learning AI systems. FDA should also provide directions regarding reusing models in the productive environment after a certain period of time. Finally, the harmonization of new guidelines with other existing frameworks around Software as a Medical Device (SaMD) and applications outside of manufacturing like clinical trial should be considered by FDA.
As we are actively bringing forward regulatory guidance volunteering in industry groups like PDA, GxP-CC has extensive knowledge and experience in compliant digitalization of life science industry. We offer comprehensive and contemporary services to support you in leveraging the power of AI technology, while ensuring compliance with GxP regulations.
Contact us to learn how we can assist you with your AI/ML compliance challenges.