US FDA finalizes guidance on machine learning for medical devices

The US Food and Drug Administration has finalized guidance on the regulation of software as a medical device (SaMD) that utilizes artificial intelligence or machine learning (AI/ML) in components for radiological medical devices, such as x-ray machines or magnetic resonance imaging equipment (MRI).

In its Technical Performance Assessment of Quantitative Imaging in Radiological Device Premarket Submissions, the agency outlines the information manufacturers must include in premarket submission applications for radiological devices with quantitative imaging functions.

Required information in premarket submissions of AI/ML-based SaMDs

The new regulations for medical devices that use quantitative imaging were first proposed in 2019 and cover what to include when describing how the quantitative imaging functions on the SaMD you are submitting for FDA approval works.

FDA’s AI/ML device guidance follows the publication of the agency’s in-depth AI/ML SaMD action plan in early 2021. As we previously reported, transparency of AI/ML technologies in support of patient-centered approaches was identified as a key goal in the plan report, and the publication of the registered AI/ML device list last fall was a step toward that goal.

The new guidance gives background on what to consider including in your quantitative imaging algorithms that use AI/ML to produce results. Most imaging diagnostics are interpreted by experienced doctors qualitatively, but increasingly, imaging equipment utilizes quantitative imaging results produced via machine learning. FDA now requires manufacturers submitting a device for market approval to include a technical description of the quantitative imaging functions of the device with enough detail for regulators to understand how it works.

FDA guidance to manufacturers:

• Shows you ways to describe how your device’s quantitative imaging function(s) work
• Explains how to demonstrate that its quantitative imaging functions meet the predefined performance specifications
• Tells you how to ensure you include sufficient details in labeling for end-users to take into account in clinical decision-making, including a description of the substance being measured, a description of the algorithm inputs, performance expectations and instructions for quality assurance and image acceptance activities performed by the user

The document also identifies a list of sources of errors that FDA believes are important to understand to better characterize the performance of the quantitative imaging functions.

Software documentation should be included in AI/ML-based SaMD premarket submissions

In this new regulation, FDA asks sponsors to include documentation that explains how the software used in the algorithm works, especially for the quantitative imaging functions.

Submissions should also include software details such as:

• Input data
• Image acceptance procedures
• Instructions provided to users
• Interaction required by users in order for device to be used as intended

FDA cautions that quantitative imaging values are subject to errors

FDA guidelines caution that quantitative imaging values from medical images and imaging data can be affected by errors from a host of sources and to consider that fact when designing algorithms. These inaccuracies can be caused by both systemic errors as well as random variation. Understanding the source of these errors can help manufacturers measure quantitative imaging performance.

Quantitative imaging errors can be caused by:

• How the image is acquired
• Patient profile
• Image processing algorithms

Keeping track of potential errors is key for manufacturers to provide quality quantitative imaging results. Manufacturers should consider conducting a sensitivity analysis to determine all the potential areas where data could be problematic, including image acquisition, image processing and patient features.

Emergo by UL will post updates to FDA’s AI/ML-based SaMD regulations as they become available.

Learn more about FDA’s AI/ML-based SaMD regulations at Emergo by UL:

• SaMD secure development lifecycle management support
• US FDA 510(k) consulting for medical device, IVD and software companies
• Cyber regulatory support for medical devices and software
• Webinar: Mapping cybersecurity standards to FDA guidance

Original Article: (https://www.emergobyul.com/blog/2022/06/us-fda-finalizes-guidance-machine-learning-medical-devices)