FDA Digital Health Guidance: Regulatory Clarity Accelerates Innovation | LeverVenture
The FDA's predetermined change control plan lets an AI-enabled device keep learning after clearance without refiling for every update. That changes what's investable.
By Peleg Chevion, Managing Partner
6 min read · Life Sciences and Healthcare
In this note05 · 6 min
Regulatory clarity is an investment thesis, not a compliance footnote. For most of the last decade, software and AI-enabled medical devices were reviewed under a framework built for a scalpel or a stent: a static product, cleared once, unchanged until the next filing. That framework never fit software that is designed to keep learning after it ships, and the mismatch showed up as underwriting risk — nobody could say with confidence how long clearance would take or what would trigger a new one. The FDA's recent guidance on lifecycle management for AI-enabled devices narrows that uncertainty, and narrower uncertainty is what makes a category investable at scale.
This is a regulatory piece, not a clinical one. The question is not whether a given digital health product works. It is whether the path to, and through, market authorization is now predictable enough that a growth equity investor can underwrite a company's regulatory strategy with the same confidence as its commercial strategy.
01Three Pathways, One New Mechanism
Digital health products still move through the same three device pathways as any other medical device: 510(k) clearance for devices substantially equivalent to something already on the market, De Novo classification for genuinely novel device types with no existing predicate, and premarket approval for the highest-risk products. None of that changed. What changed is a mechanism that sits on top of all three: the predetermined change control plan, which lets a manufacturer describe in advance how an AI model is allowed to change after clearance — the data it will retrain on, the performance bounds it must stay within, how a modification will be validated — and then implement changes inside that plan without a fresh submission for each one.
That is a meaningful shift. Under the old model, a materially updated algorithm could require the company to refile, effectively freezing the product's learning in place between submissions. A cleared change control plan converts an AI-enabled device from a single frozen artifact into a governed, evolving product — much closer to how the underlying technology actually behaves.
02Good Machine Learning Practice as the New Diligence Checklist
Alongside the change control mechanism, the agency has been building out a parallel set of expectations under the banner of Good Machine Learning Practice: guidance on how training and testing data should be selected and documented, how models should be validated for the populations they will actually see in the field, and how much of that reasoning should be disclosed to a clinician using the tool. None of this is unique to digital health investing, but it is unusually useful as a diligence checklist, because it gives an investor a structured way to ask a founder specific, falsifiable questions: What does your training population look like relative to your target population? What is your retraining cadence, and does it fall inside your cleared change control plan or outside it? What happens to model performance when the input data drifts?
A company that can answer these cleanly has effectively been pressure-tested by the same framework the FDA uses. A company that cannot is carrying regulatory risk that will eventually surface as either a delayed submission or a post-market enforcement action.
This guidance did not appear all at once. It has accumulated in stages over several years, with earlier documents establishing the guiding principles for machine learning development practice and for change control planning specifically, and more recent guidance layering on lifecycle management expectations for the full submission. The direction of travel across that sequence is consistent: less case-by-case negotiation, more standardized documentation a company can prepare well ahead of any filing. For an investor, that sequence itself is useful diligence context — a founding team that has tracked and built toward this progression looks different from one encountering the framework for the first time during a submission.
A company's regulatory strategy is no longer separable from its product roadmap. The change control plan is the roadmap, filed in advance and held to.
03Where the Clarity Still Runs Out
The framework above works cleanly for AI models performing a bounded task: flagging a scan, triaging a risk score, adjusting a dosing recommendation within defined limits. It works far less cleanly for generative AI systems that produce open-ended text or conversation, which is precisely the category the agency's Digital Health Advisory Committee has been convened to think through, particularly in mental health applications where a model's output is not a single bounded prediction but an ongoing interaction. There is no settled change control framework yet for a product whose behavior space is that open, and any company operating in that segment should be underwritten with wider regulatory uncertainty than one building a narrower clinical decision support tool.
This is the honest caveat to the thesis: regulatory clarity has advanced furthest exactly where the underlying task is most bounded. The more open-ended the AI system, the more the investor is still underwriting genuine regulatory ambiguity, not settled process.
The agency's own advisory process reflects this. Its digital health advisory committee has convened specifically to weigh the benefits, risks, and appropriate premarket evidence and postmarket monitoring for generative AI-enabled mental health products — a sign that the framework for bounded, task-specific AI is settled enough to build a company on, while the framework for open-ended conversational systems is still being written in public, one advisory meeting at a time. A company operating in the latter category is not wrong to pursue it, but its timeline to market authorization should be modeled with a wider range than a company building a narrower, bounded tool.
04What Changed for Underwriting
| Pathway or mechanism | Fits | What to diligence |
|---|---|---|
| 510(k) | A device substantially equivalent to an existing predicate | Strength of the predicate comparison, not just the algorithm |
| De Novo | A genuinely novel device type, no predicate available | Whether the agency has classified anything comparable since |
| Premarket approval | Highest-risk devices, most extensive clinical evidence required | Trial design and endpoint selection, well ahead of filing |
| Predetermined change control plan | Any of the above, for a model expected to keep learning | Whether the plan's bounds match how the company actually intends to iterate |
05How to Use This in Diligence
Growth equity investors underwriting a company in this category should treat the regulatory strategy as a first-class diligence workstream, not a legal sign-off that happens after the investment thesis is set. Ask whether the company has filed or intends to file a predetermined change control plan, and read the plan's bounds against the company's actual retraining and deployment cadence — a plan written too conservatively will force refilings the company has not budgeted for, and one written too loosely may not survive review. Ask how the company's training data compares with its intended-use population, and treat a vague answer as a real finding, not an administrative gap to close later.
The companies best positioned in this window are the ones that have made regulatory strategy part of the product decision from the start, rather than a constraint discovered at filing. For how this connects to the broader diagnostics and clinical-decision-support stack, see our note on the precision medicine investment landscape, and for the AI layer specifically, see AI in digital health diagnostics.
Nothing in this piece is investment, legal, tax or accounting advice, and nothing in it is an offer to sell or a solicitation of an offer to buy any security.
Three decades of operational leadership across agriculture, pharmaceuticals, biotechnology, climate, food and defense. He has spent that career doing the work this seat requires: strategic partnerships, licensing and business development, executed transaction by transaction. He sits with a company as an operating partner and a growth strategist rather than as a name on its board. A corporate finance and intellectual property attorney, an INSEAD MBA, and a former Air Force officer in aerospace. He also brings strong private equity experience.