FDA AI-Enabled Device Predicate Mining for 510(k) and De Novo

What This Article Covers / Does Not Cover

This article covers one specific task: how to use FDA's public AI-enabled medical device list, the 510(k) database, and publicly available summary documents to identify, evaluate, and defend a predicate device for an AI-enabled medical device seeking 510(k) clearance or De Novo classification. It includes a predicate mining workflow, a technological characteristic extraction table, a predicate strength scoring matrix, and common failure modes specific to AI-enabled devices.

This article does not cover the general 510(k) submission process, how to write a substantial equivalence argument from scratch, or how to prepare performance testing data. For the general 510(k) framework, see 510(k) Submission Guide. For predicate device selection fundamentals, see 510(k) Predicate Device Guide. For the broader AI/ML regulatory landscape, see AI/ML Medical Device Regulatory Guide.

Why AI-Enabled Device Predicate Mining Is Different

Approximately 97% of FDA-authorized AI-enabled medical devices have been cleared through the 510(k) pathway. Through the end of 2025, FDA's list contains over 1,450 cumulative authorized AI-enabled devices. In 2025 alone, 295 AI/ML-enabled devices received clearance, with 211 (71.5%) in radiology, 26 (8.8%) in cardiovascular, and 14 (4.7%) in neurology.

But AI-enabled devices present unique predicate challenges:

1. Software opacity: Public 510(k) summary statements rarely describe AI architecture, model type, training data characteristics, or inference pipeline details.

2. Rapid evolution: A predicate cleared two years ago may use fundamentally different technology than what is standard today.

3. Mixed technology comparisons: FDA permits AI-enabled devices to be found substantially equivalent to non-AI-enabled predicates if no new questions of safety and effectiveness are raised (per FDA's draft AI-enabled device software functions guidance, January 2025).

4. Sparse non-radiology predicates: Outside radiology, the predicate pool shrinks dramatically, forcing sponsors toward De Novo or toward weaker predicate arguments.

Step 1: Define Your Technological Characteristic Profile

Before searching for predicates, document your own device's technological characteristics in a structured format. FDA evaluates substantial equivalence along two axes: intended use and technological characteristics. The table below provides the characteristic categories most relevant to AI-enabled devices.

Technological Characteristic Extraction Table

Step 2: Mine the FDA AI-Enabled Device List

2.1 Access the List

FDA publishes the "Artificial Intelligence-Enabled Medical Devices" list on its website at fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices. The list is downloadable in CSV, Excel, and XML formats. It contains:

• Date of final decision

• Submission number (e.g., K252379)

• Device name

• Company name

• Panel (lead review division)

• Primary product code

2.2 Filter Strategically

Apply filters in this order:

2.3 Record Candidate Predicates

For each candidate identified, create a predicate candidate card:

Step 3: Extract Technological Characteristics from Public Summaries

3.1 What Public Summaries Actually Contain

FDA's publicly available 510(k) summary statements (found in the 510(k) database) typically include:

• Device description (sometimes detailed, sometimes vague)

• Intended use statement

• Comparison to predicate(s): intended use, technological characteristics

• Summary of performance testing

• Brief statement on software validation

• Conclusion on substantial equivalence

3.2 What Public Summaries Almost Never Contain

3.3 Technological Characteristic Comparison Matrix

Build this matrix for every candidate predicate. Use "Same," "Different," or "Unknown" for each row:

Step 4: Predicate Strength Scoring

Score each candidate predicate on a 1-5 scale across the following dimensions. A score below the threshold means the predicate is too weak to rely on as a primary predicate.

Predicate Strength Scoring Matrix

Weighted score threshold: A weighted score below 3.0 indicates a weak predicate. Below 2.5, consider De Novo or a different predicate strategy.

Decision Tree: Primary Predicate vs. Multiple Predicates vs. De Novo

START: Is there a candidate with weighted score >= 3.5?
├── YES → Use as primary predicate
│   └── Are there gaps in technological characteristics?
│       ├── NO → Proceed with single-predicate SE argument
│       └── YES → Can gaps be addressed with additional reference devices?
│           ├── YES → Add secondary predicate(s) for specific features
│           └── NO → Address gaps with performance testing data
└── NO → Is there a candidate with weighted score 2.5-3.5?
    ├── YES → Can additional performance data bridge the gap?
    │   ├── YES → Strengthen SE argument with robust testing
    │   └── NO → Consider Pre-Submission meeting with FDA
    └── NO → Is there ANY candidate with score >= 2.0?
        ├── YES → De Novo may be more appropriate
        └── NO → De Novo pathway is likely required

Step 5: Avoid Weak Predicate Arguments

Common Failure Modes for AI-Enabled Device Predicates

Reviewer Objection Table

Step 6: Use Pre-Submission to De-Risk Predicate Strategy

FDA's Pre-Submission program (Q-Submission) allows you to get feedback on your predicate strategy before committing to a full 510(k) or De Novo submission. For AI-enabled devices, this is especially valuable when:

1. Your best predicate scores between 2.5 and 3.5 on the strength matrix

2. You are considering a non-AI to AI comparison

3. You are outside radiology (where predicates are sparse)

4. You plan to use multiple predicates

Pre-Submission Predicate Package Contents

Illustrative Pre-Submission question format:

"We have identified [K-number] as our primary predicate device. Our device shares the same intended use (detection of [condition] from [modality]), same input data type, and same clinical workflow position. The primary technological difference is [describe]. We plan to address this difference with [proposed testing]. Does FDA agree that this predicate and testing approach are appropriate for a 510(k) submission?"

Step 7: Document the Predicate Strategy in Your Submission

Evidence Traceability Table

RACI for Predicate Mining Process

Common Failure Modes and Remediation

Sources

• FDA. "Artificial Intelligence-Enabled Medical Devices." Updated through December 2025. fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices

• FDA. "Artificial Intelligence-Enabled Device Software Functions: Lifecycle Considerations and Premarket Review Recommendations." Draft Guidance, January 2025.

• Innolitics. "2025 Year in Review: AI/ML Medical Device 510(k) Clearances." innolitics.com/articles/year-in-review-ai-ml-medical-device-k-clearances/

• IntuitionLabs. "FDA's AI Medical Device List: Stats, Trends & Regulation." Updated March 2026.

• FDA. "The 510(k) Program: Evaluating Substantial Equivalence in Premarket Notifications." Guidance.

• Complizen. "What Is Substantial Equivalence (SE) in FDA 510(k)? Definition & Criteria." 2025.