By Hsin-Wei Shen, VP of Investment & Clinical Strategy, BE Health Ventures

The most common pitfall for AI SaMD (Software as a Medical Device) startups isn’t having a weak algorithm, but choosing the wrong regulatory pathway or the wrong Predicate device. Once either of these goes wrong, the consequence usually isn’t just having to submit a few extra documents. It can delay your time-to-market by over six months, severely damage your burn rate, and even exhaust your runway before your next fundraising round—forcing you down a De Novo path you never intended to take.

Below, we break down 510(k) and De Novo into an actionable internal decision-making framework using FDA review language. We will systematically cover a classification decision tree, Predicate risks, the cost of making the wrong choice, and mitigation tactics. Finally, we provide a ready-to-use Predicate screening memo template for your team.

I. Let’s Clarify the Rules: The Fundamental Differences Between 510(k) and De Novo

1) 510(k): You are establishing Substantial Equivalence (SE)

The premise of a 510(k) is that you can find a legally marketed, cleared Predicate device, and prove that your product has the same Intended Use, similar or comparable technological characteristics, and that its overall safety and effectiveness are not inferior to the Predicate.

The core value of this pathway lies in its controllable pace. The official target review time is 90 days, but in practice, it averages 4 to 6 months, depending on the rounds of Additional Information (AI) requests and product complexity. Currently, the vast majority of AI/ML SaMDs still utilize 510(k) as their mainstream pathway, with common statistics reaching 97% to 97.5%.

2) De Novo: You are establishing a new classification and Special Controls

When you cannot find a suitable Predicate, but the product still poses a low to moderate risk (Class I or Class II), the De Novo pathway is used to establish a new classification, a new product code, and propose Special Controls acceptable to the FDA.

The core cost of this pathway is higher evidence density and a longer timeline. The official target is 150 to 180 days, but practically averages 7 to 9 months, and can extend to 12 months if disputes arise. Operationally, you must act as your own Predicate, independently proving safety and effectiveness, providing clinical data or Real-World Evidence (RWE), performing a risk-benefit analysis, and proposing control measures for governable updates (e.g., a Predetermined Change Control Plan, or PCCP).

II. A Classification Decision Tree: Which Path Should You Actually Take?

Distill the pathway choice into three decision points. Following this flow will keep you from going off course.

• Point 1: Can the Intended Use be aligned word-for-word with an existing cleared product?

  The biggest fear in a 510(k) isn’t that your features are superior, but that your Intended Use looks similar but is fundamentally different. If the Intended Use doesn’t align perfectly, no matter how much you compare performance, it will be extremely difficult for the FDA to clear your 510(k) under the SE logic.

• Point 2: Is there a defensible Predicate?

  “Defensible” does not mean just pulling a K-number from a database. A defensible Predicate means your candidate meets several criteria simultaneously: the Intended Use aligns, technological characteristics are similar or equivalence can be bridged through testing, there are public data or literature to establish a performance baseline, and there are no known safety concerns or major design baggage.

• Point 3: Will technological differences raise new questions of safety and effectiveness?

  If the FDA believes that differences in output formats (e.g., heatmaps), use cases, or incomparable performance metrics lead to new questions, you are highly likely to receive an Additional Information (AI) request, or even be designated Not Substantially Equivalent (NSE), ultimately forcing a change of course.

III. How to Find a Defensible Predicate: The 4 Best Practices the FDA Cares About Most

The Predicate search should be broken down into two steps: first find candidates, then screen them.

Step 1: Find candidates (product code + keywords)

Start with the FDA 510(k) database (including AccessGUDID) and the official AI/ML device list. Narrow your scope using product codes and keywords. In practice, prioritize products cleared in the US within the last 2 to 3 years to align closer with current standards, avoiding wasting time on overly legacy cleared devices.

Step 2: Screen candidates (Based on the FDA’s 2023 4 Best Practices)

• Recent clearance: Prioritize products cleared within the last 2 years.

• No known safety issues: Check public records like MAUDE, recalls, and warning letters to ensure there are no major red flags.

• Demonstrated performance: Ensure there is public literature or clinical data sufficient to establish a baseline for safety and effectiveness.

• No major design issues: Avoid products with known design flaws or highly controversial designs.

Two bonus but often misunderstood points:

• A Predicate does not necessarily have to be an AI. You can use traditional, non-AI software as a Predicate, as long as you can prove the AI’s output is clinically equivalent or can clearly bridge the comparability through testing.

• Avoid “split predicates.” Piecing together equivalence using multiple Predicates makes your argument fragile. The FDA does not favor this approach, and in practice, AI/ML SaMDs primarily rely on a single primary Predicate.

IV. The Cost of Choosing the Wrong Predicate: Not Just Rejection, but Rerouting and Collapsed Timelines

In the AI SaMD space, Predicate disputes are high-frequency events. Data suggests that the rate of AI SaMDs facing an NSE determination by the FDA can be as high as 25%, and choosing the wrong Predicate is one of the most typical and fatal reasons.

The typical losses occur on three levels:

1. Review Delays: You will first receive an AI letter. The back-and-forth tug-of-war will delay your market launch by several months.

2. Rework Costs: Re-testing, rewriting documents, and gathering additional clinical data often burn through a significant amount of cash—commonly reaching the $50,000+ level or higher.

3. The Worst-Case Scenario: An NSE forces you to withdraw and resubmit, or pivot directly to De Novo. This means treating your initial 510(k) investments as sunk costs (tuition fees) while having to rebuild your evidence strategy and product narrative from scratch.

Common lessons fall into two categories:

• Medical imaging AI choosing older, non-ML software as a Predicate. Due to differences in deep learning outputs and presentation formats, they are deemed non-equivalent (NSE) and forced into a De Novo pathway, wasting months.

• Forcibly mapping Predicates across diseases or anatomical regions. For example, a cardiac diagnostic AI wrongly using a stroke AI as a Predicate. The vast differences in anatomical sites and performance metrics lead to heavy AI requests and ultimately a forced withdrawal and resubmission.

V. Three-Stage Mitigation Strategies: Use the Tactic That Matches Where You Are Stuck

Stage A: Before Submission (Lowest Cost, Highest Return)

It is strongly recommended to submit a Q-Submission (Q-Sub) 1 to 2 months before your formal submission. Show the FDA your Intended Use word-for-word alignment strategy, your reasons for selecting the Predicate (using the 4 Best Practices framework), and your plan to bridge differences with testing. Securing written feedback significantly reduces NSE risks and prevents you from realizing you picked the wrong path halfway through the review.

Stage B: Under Review, FDA Questions the Predicate, but No NSE Yet

The goal here is to stop the bleeding and stabilize the SE argument.

1. Activate a backup Predicate: If you prepared a second-choice candidate in advance, pivot quickly to avoid digging a deeper hole with the wrong Predicate.

2. Reinforce the hierarchy of evidence: Use FDA guidances, consensus standards, and peer-reviewed literature to establish a baseline, then use gold-standard comparative testing to turn your equivalence argument into quantifiable conclusions.

3. Reframe the differences: Rewrite the differences into testable claims, bringing the review focus back to endpoints you can control.

Stage C: Already Ruled NSE, or You Realize You Can’t Hold the Line

At this point, only two options remain:

1. File a new 510(k): Resubmit using a completely different Predicate, firmly nailing down the SE argument from Intended Use through to the testing plan.

2. Pivot decisively to De Novo: When no truly equivalent product exists on the market, forcing an SE argument will only drain your burn rate. Pivoting to De Novo to establish an exclusive classification is often a much better allocation of resources.

VI. De Novo is Not a “Harder 510(k)”; It’s a Different Kind of Evidence Battle

Having no Predicate means you must independently prove safety and effectiveness using comprehensive data and control measures, and propose Special Controls. A De Novo evidence package typically contains more robust clinical data or multi-center RWE, an independent and quantified risk-benefit analysis, and Special Controls proposals. For AI/ML products, a PCCP (Predetermined Change Control Plan) is often crucial because it proves that future model changes can be managed safely, ensuring your product update strategy doesn’t become a regulatory black hole.

VII. A Ready-to-Copy Predicate Screening Memo Template

Handing the following Memo template to your team will forge consensus faster than verbal discussions and is easier to reuse in Q-Sub and 510(k) documentation.

• Part 1: Device Overview. Explain in three sentences: Product Name, Draft Indication or Intended Use, and Clinical Workflow. Clearly map out the relationship between input data, algorithm output, and the physician’s decision-making point.

• Part 2: Candidate Predicate List. List at least three candidates and evaluate them one by one.

  • Consistency Check: Can the Intended Use be aligned sentence by sentence? Are the anatomical sites, populations, and users consistent?

  • Technological Comparability Check: Is the output format consistent (e.g., heatmaps, triage alerts)? Are performance metrics comparable (e.g., sensitivity, specificity, AUC)?

  • 4 Best Practices Quick Check: Cleared within the last 2 years? Clean MAUDE records? Public data supporting performance? No major design issues?

  • Risk Notes: If a second Predicate is needed to bridge the argument, flag it directly as high risk. If a non-AI Predicate is a candidate, clearly outline the equivalence evidence strategy and comparative testing design.

• Part 3: Backup Strategy. Specify the Primary Predicate and Backup Predicate. List the delta list (differences) and the corresponding testing matrix. Add the hierarchy of evidence plan, explaining what literature, consensus standards, or existing data you plan to cite to establish a baseline if the Predicate’s public data is insufficient.

• Part 4: Q-Sub Plan. Fill in the expected submission date (recommend 1 to 2 months prior to formal submission) and list the specific conclusions you want from the FDA (e.g., acceptance of Predicate, acceptance of classification, acceptance of testing plan).

• Part 5: Resources and Timeline Early Warning. Estimate preparation costs (additional testing, data acquisition, external clinical/statistical support). Explicitly state the impact of the worst-case scenario (e.g., if ruled NSE and forced to De Novo, time-to-market increases by 6 to 9 months) and its specific impact on burn rate and fundraising pacing.

Conclusion

From an investment and operational pacing perspective, the choice between 510(k) and De Novo is fundamentally a milestone and cash flow management problem. The speed dividend of a 510(k) is built on two premises: the Intended Use must align word-for-word, and the SE must be solidly backed by quantitative evidence.

As soon as the Predicate falters, the rounds of Additional Information will drag on, and timeline uncertainty will immediately reflect in your burn rate, business negotiations, and fundraising valuation—potentially exhausting your runway before the next round.

Conversely, the innovation dividend of a De Novo isn’t that it’s “harder” or “cooler,” but that you have the opportunity to establish a product code and Special Controls, setting the rules for yourself and future competitors. However, the cost is higher evidence density and stricter risk governance—especially ensuring that your PCCP and continuous monitoring mechanisms are executable and auditable.

The most pragmatic strategy is to front-load the uncertainty. Use a Q-Sub to lock in your classification and Predicate before formal submission. This allows the company to focus its limited resources on the evidence most likely to translate into clearance and revenue, preventing a mid-review forced pivot that sends both your product timeline and capital pacing into a tailspin.