FDA Requests Feedback on Regulating Artificial Intelligence and Machine Learning in Drug Development and Manufacturing

Artificial intelligence (AI) and machine learning (ML) have become ubiquitous across industries, and the pharmaceutical industry is no exception. AI and ML are already influencing drug development and manufacturing, but these innovations present unique regulatory challenges. For example, if a ML algorithm can change a CGMP-compliant manufacturing process on its own to increase efficiency, how does the drug manufacturer ensure that the updated, machine-created process is CGMP-compliant? If researchers use AI to identify ideal candidates to participate in a drug trial, how do they account for biases in the data underlying the AI’s decision-making? FDA recently released a discussion paper^[1] requesting input from pharmaceutical industry stakeholders on how to tackle such issues. Stakeholders should seize this opportunity to provide input to FDA as the agency develops the applicable regulatory landscape.

Below, we discuss real-world examples of AI/ML in drug development and manufacturing, explain the overarching principles that guide FDA’s current thinking on the AI/ML regulatory landscape, and highlight key questions posed by the agency to industry stakeholders.

What are AI and ML?

FDA describes AI as “a branch of computer science, statistics, and engineering that uses algorithms or models to perform tasks and exhibit behaviors such as learning, making decisions, and making predictions.” ML is a subset of AI that allows “models to be developed by ML training algorithms through analysis of data, without models being explicitly programmed.”^[2]

Applications of AI and ML in Drug Development and Manufacturing

FDA’s discussion paper highlights potential benefits and applications of AI and ML in drug development and manufacturing. For example:

• AI/ML can create “digital twins” of individuals that allows predictive modeling of that individual’s reaction to a given drug before use.
• AI/ML can optimize existing drug manufacturing processes to maximize efficiency and minimize waste. Continuous, real-time sensor data enables manufacturers to detect changes or deviations during the manufacturing process that signal the need for equipment maintenance.
• AI can monitor product quality. For example, AI can perform a quality control inspection of product packaging by analyzing a product for visible deviations from the images pre-programed into AI-based software.
• AI/ML can advance logistics and prevent supply chain disruptions by forecasting product demand and optimizing inventory.
• Post-manufacture, AI can collect and monitor consumer complaints and adverse events, and identify trends in reported product issues, which may expedite the identification of root causes, “as solely manual review of deviation trends can be very time-consuming.”

Overarching Principles For Using AI/ML In Drug Development And Manufacturing

While the FDA is still in the early stages of developing its regulatory approach to AI and ML in drug development and manufacturing, the agency outlines three overarching principles guiding its thinking on these issues: “(1) human-led governance, accountability, and transparency; (2) quality, reliability, and representativeness of data; and (3) model development, performance, monitoring, and validation.”

First, FDA considers human-led governance a priority for creating trustworthy AI/ML and any future regulatory framework for AI/ML will likely require transparency and documentation with respect to all decision-making or process changes implemented by AI/ML in the drug development and manufacturing space. This approach is consistent with FDA’s approach to AI and ML in the medical device context, where FDA has proposed requiring manufacturers to monitor software changes implemented by AI/ML and provide periodic updates of these changes to FDA.^[3]

Second, FDA is concerned with the quality and reliability of data underlying AI/ML processes, particularly because of the potential for AI/ML to “amplify preexisting biases that exist in the underlying input data.” FDA’s focus on bias identification and management indicates that a future regulatory framework is likely to require documentation and explanation of how biases in data AI/ML underlying were managed in the drug development process.

Finally, while FDA has specified that regular monitoring and documentation are critical to ensure that AI/ML models are explainable, reliable, and verifiable, the agency also notes that “a risk-based approach may guide the level of evidence and record keeping needed for the verification and validation of AI/ML models for a specific context of use.” FDA’s future regulatory framework is likely to impose less stringent requirements on simpler, more transparent modeling, and require more documentation and auditing for complex models (e.g., artificial neural networks).

FDA Requests Input From Pharmaceutical Industry Stakeholders

In light of the three overarching principles discussed above, FDA identifies specific areas requiring feedback and discussion with stakeholders to help inform their regulatory activities. These include:

• How can the pharmaceutical industry ensure accountability, transparency, and trustworthiness of AI/ML systems that may not be readily explainable or understandable due to their complexity?
• How can drug developers using AI/ML prevent amplification of errors and biases in underlying data sources, and ensure the privacy of patient data?
• How can parties using cloud applications (particularly from third-party hosts) to store product manufacturing data ensure data integrity, quality, and security, particularly when these issues affect a manufacturer’s CGMP obligations and oversight requirements?
• How can drug manufacturers store data generated for regulatory compliance (e.g., data that supports future quality decisions such as product recalls), in a manner that enables retrieval and analysis to support their decision-making?
• How can parties ensure that processes controlled by ML algorithms comply with regulatory obligations, when those algorithms change and adapt processes based on real-time data?

Conclusion

Interested parties should strongly consider taking advantage of FDA’s invitation to participate in the discussion as it develops its approach to AI/ML in drug development and manufacturing. Industry participants are in the best position to help educate FDA about these evolving issues. FDA has requested electronic or written comments by August 9, 2023. You can submit your feedback here. We will continue to monitor and report on regulatory developments in this space as they become available.