Summary
By the time teams reach the submission phase, the assumption is often that the hardest decisions are behind them. In Part III, Leon challenges that notion, highlighting how the final stage of an FDA submission is not just about compiling documentation but about execution precision, where small gaps can lead to significant delays.
Building on the foundations of Parts I and II, this article shifts focus to the structure and expectations of the submission itself. A 510(k) or De Novo filing is not a single document, but a highly structured package, guided by frameworks such as the FDA’s eSTAR template and governed by Refuse-to-Accept (RTA) criteria. Missing or incomplete elements at this stage don’t just slow things down; they can halt the process entirely, adding weeks or even months to timelines.
Leon outlines the key components that define a robust submission. These include the indications for use, detailed device descriptions (with specific attention to AI model architecture, training data, and outputs), substantial equivalence comparisons, and performance validation across bench, standalone, and clinical contexts. Increasingly, software and cybersecurity documentation have become central, not peripheral, elements of the submission.
One of the more forward-looking aspects discussed is the Predetermined Change Control Plan (PCCP). As AI/ML systems evolve post-deployment, regulators are adapting to allow controlled updates without requiring entirely new submissions. However, this flexibility comes with expectations: clear documentation, defined boundaries for change, and robust lifecycle management strategies.
A recurring theme throughout the article is where teams tend to lose time: not in major technical failures, but in incomplete alignment between sections. For example, inconsistencies between intended use, validation datasets, and labelling can trigger additional queries or review cycles. Similarly, insufficient detail in software documentation or risk management can raise concerns that stall progress.
What stands out is Leon’s emphasis on discipline. At this stage, success is less about innovation and more about coherence, completeness, and clarity. Every section of the submission must tell the same story: aligned, justified, and defensible.
The blog ultimately reframes the submission phase as more than a final step; it is the point where strategy, validation, and documentation converge. And if the earlier stages were about making the right decisions, this stage is about proving them convincingly.
This summary offers a high-level view, but the real value lies in the detailed guidance Leon provides. To understand how to structure your submission effectively and avoid last-mile delays, read the full article: AI/ML Systems, preparing an FDA submission Part III: ‘The submission’
