Summary
FDA submissions for AI/ML-enabled medical devices rarely fail because of a single mistake – they unravel due to a series of early assumptions that compound over time. In Part I, Leon Doorn, MedQAIR CEO & Co-Founder, sets the foundation by focusing on what is arguably the most critical (and often underestimated) step: defining the initial regulatory strategy, starting with product classification and product code identification.
The article opens with a candid reflection on early AI/ML submission experiences, when regulatory guidance was limited and expectations were unclear. What follows is not just a story, but a hard-earned lesson: even when teams believe they are aligned with the FDA, interpretations can shift, reviewers can change, and pre-submissions are ultimately non-binding. This uncertainty reinforces a central theme – experience and strategic clarity matter more than assumptions.
Leon then brings structure to what often feels like a fragmented process. Unlike EU pathways, where compliance is anchored around general safety and performance requirements, the FDA framework is classification-driven. For most AI/ML devices, this means navigating the 510(k) pathway through substantial equivalence – or, where no predicate exists, considering De Novo or even PMA routes. But none of these decisions can be made in isolation.
Everything begins with intended use and indications for use. These are not just descriptive elements; they define the clinical context, risk profile, and ultimately the regulatory pathway itself. From here, identifying the correct product code becomes a pivotal step, shaping expectations around evidence, validation, and applicable “special controls.”
The blog also highlights a practical but often overlooked challenge: navigating the FDA’s classification ecosystem. While modern AI tools can accelerate the search for predicates and product codes, they are not immune to inaccuracies. Leon emphasises the importance of cross-verifying outputs against official FDA databases, especially when working with emerging AI-driven technologies where classifications may not be straightforward.
Another key layer is predicate selection. Understanding how previously cleared devices were positioned: their intended populations, performance claims, and validation approaches, provides essential context for shaping your own submission. This comparative lens helps teams anticipate regulatory expectations rather than react to them.
The takeaway is clear: early decisions are not just foundational; they are determinative. Misalignment at this stage does not surface immediately; it manifests later as delays, rework, and costly course corrections.
This summary captures the starting point, but not the full depth, of Leon’s experience-driven insights. For a detailed walkthrough of how to define your regulatory pathway and avoid early-stage pitfalls, read the full article: AI/ML Systems, preparing an FDA submission – Part I: ‘Initial strategy’
