The FDA’s regulatory framework for Artificial Intelligence (AI) in the drug product life cycle has transitioned from early exploration to a structured, risk-based credibility assessment framework. According to the latest draft guidance, this framework applies to the use of AI to produce information or data intended to support regulatory decision-making regarding the safety, effectiveness, or quality of human and animal drugs, biological products, and combination products.
The Risk-Based Credibility Assessment Framework
The core of the FDA's current regulatory thinking is a 7-step process designed to establish trust in AI model performance for a specific Context of Use (COU).
Define the Question of Interest: Clearly state the specific decision or concern the AI model is intended to address.
Define the Context of Use (COU): Describe the specific role and scope of the model, including whether it is the sole determinant or used alongside other evidentiary sources.
Assess the AI Model Risk: Risk is determined by two factors: Model Influence (the weight of AI evidence relative to other data) and Decision Consequence (the significance of an adverse outcome from an incorrect decision).
Develop a Credibility Assessment Plan: Outline the activities to establish model credibility, such as describing model architecture, data management, and training processes.
Execute the Plan: Implement the predefined credibility activities.
Document and Discuss: Create a Credibility Assessment Report documenting results and any deviations from the plan.
Determine Adequacy: Final determination of whether the model is appropriate for the COU; if not, the sponsor may need to increase rigor, add data, or mitigate risk through controls.
Data Integrity and the "Fit for Use" Standard
The FDA emphasizes that AI performance relies heavily on data quality. Data must be "fit for use," meaning it is both relevant (representative of the target population or process) and reliable (accurate, complete, and traceable).
FAIR Principles: The broader scientific context supports this via the FAIR Guiding Principles—ensuring data is Findable, Accessible, Interoperable, and Reusable for both humans and machines.
Data Drift: Regulators are particularly concerned with data drift, where model performance changes over time because new inputs differ from the original training data.
Life Cycle Maintenance in Manufacturing (Industry 4.0)
In the context of pharmaceutical manufacturing, AI is a key enabler for the Industry 4.0 paradigm, supporting Advanced Process Control (APC) and smart monitoring.
Continuous Learning: Systems that adapt to real-time data present unique challenges for oversight. The FDA expects these changes to be managed within a pharmaceutical quality system (e.g., ICH Q10) to ensure the model remains fit for use throughout its life cycle.
Cloud and Edge Computing: The use of third-party cloud data management for AI models requires robust quality agreements and introduces new complexities for FDA inspections regarding data traceability and cybersecurity.
Regulatory Engagement Strategies
The FDA strongly encourages early engagement to align on model risk and credibility plans. Various specialized programs exist depending on the AI's application:
C3TI & CID: For AI in clinical trial innovation and novel designs.
ISTAND & DDT: for qualifying AI-based drug development tools.
ETP & CATT: For AI applications in pharmaceutical manufacturing.
EDSTP: Focused specifically on AI in pharmacovigilance for postmarketing activities.
Alignment with Industry Standards
The regulatory framework is increasingly aligned with industry best practices, such as the ISPE GAMP® Guide: Artificial Intelligence. This guide bridges established GAMP 5 risk-based concepts with AI-specific characteristics, focusing on quality by design, explainability, and the integration of specialized roles like data scientists into the quality assurance unit.
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit bioedgepartners.substack.com
