Key considerations in 510(k) submissions for AI-enabled products

When you are developing an AI-enabled medical device, one thing is clear: getting your 510(k) submission right is crucial. It can be a daunting task, but with the right approach, you can streamline the process and increase your chances of success.

At Enhatch, we recently received U.S. Food and Drug Administration (FDA) 510(k) clearance for our patient-specific AI-driven instrumentation system for total knee arthroplasty. Our system utilizes AI to transform patient X-rays or CT images into 3D anatomical models.

Having recently completed the 510(k) submission process, I am sharing some experiences and suggestions that could be useful to companies when navigating the process.

1. Understand FDA Guidance on AI/ML Products

First things first: Get familiar with the FDA's guidelines specific to AI and machine learning (ML) in medical devices. The FDA has issued several documents to help companies understand how AI products are regulated. While traditional MedTech products have well-established submission paths, AI-enabled devices add an extra layer of complexity because of their dynamic nature.

For example, the U.S. FDA, Health Canada, and the U.K.'s Medicines and Healthcare Products Regulatory Agency (MHRA) have collaboratively outlined ten key principles to guide the creation of Good Machine Learning Practice (GMLP). This document can be a good starting point.

FDA documents that you may want to review:

• Good Machine Learning Practice for Medical Device Development: Guiding Principles.
• Artificial Intelligence and Medical Products: How CBER, CDER, CDRH, and OCP are Working Together (2024).
• Content of Premarket Submissions for Device Software Functions.
• Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan (2021).

The FDA emphasizes transparency, performance testing, and risk management for these products. Ensure your AI system meets these criteria and is consistent with the agency's guidance.

2. Demonstrate Clinical Validity with Real-World Evidence

For AI-enabled devices, clinical validity is a major point of focus for the FDA. It's not enough to show that your algorithm works in a controlled environment - you need to demonstrate that it performs just as well in real-world settings.

One way could be gathering real-world evidence (RWE) from clinical scenarios. The data needed to ensure safety and effectiveness will depend on the AI/ML function, risk level, and intended use.

You must assess whether and how your algorithms fall under FDA regulation and how the agency will evaluate real-world data (RWD) used as evidence. Ensure your evidence shows that your AI solution consistently provides accurate and reliable outcomes.

Some points to consider:

• It is essential that the data set represents the intended patient population. For example, at Enhatch, the data collection protocol was very specific to the intended use, i.e., patient imaging undergoing total knee arthroplasty at varying levels of disease state through multiple imaging facilities.
• Maintain data integrity and quality without biasing the data.
• Make sure that your training data set is independent from the validation data set.
• Ensure the amount of data collected results in a validation data set that is sufficiently powered with representative subgroup analyses.

FDA guidance document that you may want to review:

• Use of Real-World Evidence to Support Regulatory Decision-Making for Medical Devices.

3. Highlight Risk Management and Mitigation Strategies

Risk management is critical when it comes to AI-enabled medtech products. Since AI can be prone to bias or errors, you must identify and address any risks associated with your device. The FDA wants to see that you have a robust risk management plan and that you are actively working to minimize potential harm.

As part of the submission, include information on:

• Software engineering and data governance practices that have been followed.
• How your AI model detects and mitigates risks, such as using fail-safes, alerts, or human-in-the-loop (HITL) mechanisms to double-check the systems' recommendations. Some processes are semi-automated, with a technician reviewing software outputs as a risk mitigation strategy.
• Device's risk mitigation strategies, like statistically sound test plans, monitoring its performance in real-time, or allowing for continuous updates based on new data.
• Robust cybersecurity practices with measures like encryption, secure access protocols, and regular software updates to prevent breaches. When preparing your 510(k) submission, know that documentation that shows your cybersecurity efforts and outlines how you protect patient data is required.

FDA guidance document to reference:

• Content of Premarket Submissions for Device Software Functions.

4. Communicate Algorithm Transparency and Explainability

AI's "black box" nature can be a sticking point in 510(k) submissions. The FDA needs to understand how your AI makes its decisions. Algorithm transparency and explainability are crucial components of the submission process for AI-enabled medtech products.

• Explain how your AI algorithm processes input data and reaches its conclusions. For example, you can show how you validated the model and ensured its consistency across different patient types.
• If possible, use visualizations or simplified flowcharts to make it easier to understand.
• Engage regularly with your FDA reviewer and involve them early in the process via the pre-submission process, especially if modifications to the software device are intended, as well as throughout premarket reviews.

In addition, Instructions for Use (IFUs) should be clear and straightforward, making it easy for end users to understand how to safely and effectively operate the AI-powered medtech device. These proactive measures can help build trust in your device and demonstrate to the FDA that you're committed to patient safety.

The more you demystify your AI, the better your chances are of getting a green light from the FDA.

FDA documents that you may want to review:

• Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles.
• Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD).

Navigating the 510(k) submission process for AI-enabled medtech products is no easy task. But with a strategic approach that focuses on understanding FDA guidelines, providing solid evidence, and ensuring transparency, you can set your product up for success.