Compliance & Regulation

Regulatory Shift for Machine Learning in Software as a Medical Device (SaMD)

By |

Machine Learning in Software as a Medical Device

Machine Learning in Software as a Medical Device

Machine learning (ML), a subset of artificial intelligence, has become integral across software in all industries, and the medical and life science spaces are no exceptions. ML can help medical systems improve the identification and diagnosis of disease, create personalized medicine, and help with drug discovery and manufacturing (just to name a few areas). Current guidance and regulations require validation of the final version of the software prior to a release—so what does this mean for Software as a Medical Device (SaMD) systems that incorporate ML and ‘continual’ algorithms designed to accumulate and improve knowledge after a system’s release in the market?

The Current Situation

The FDA and other regulatory agencies currently lack the guidance needed for medical device manufacturers to include continual algorithms that adjust and improve post-market submissions. A changed algorithm would require a premarket review for each minor adjustment due to the potential impact on patient care.

There are many medical ML products on the market today, which currently use the DeNovo and 510(k) process, but the process is locked. Otherwise, the ML product would violate regulatory control. With this limitation, the FDA has been listening to input from the industry to help inform potential guidance and regulatory change.

Future FDA Regulatory Process

In April 2019, the FDA released a paper called “Proposed Regulatory Framework for Modifications to Artificial Intelligence/ Machine Learning (AI/ML) Based Software as a Medical Device (SaMD).” The paper describes and outlines a possible approach to a premarket review for artificial intelligence and machine learning-driven software modifications.

The proposed framework includes elements from the FDA’s current premarket programs including:

  • IMDRF’s risk categorizing principles
  • FDA’s benefit-risks framework
  • Risk management approaches
  • IEC 62304 principles
  • Change management approaches

One of the key proposed expectations from the FDA would be a commitment from manufacturers to be transparent with the constant monitoring of artificial intelligence and machine-learning-based software providing periodic updates, including any changes to algorithm protocols.

The proposed framework received positive feedback from the industry, and the FDA has now decided to implement the discussion paper with guidance. It is expected to be released sometime in 2021. The intended premarket approval process will allow a trusted manufacturer (see below for definition) to make pre-approved post-release changes only if the manufacturer follows a predetermined change control plan.

What is a Trusted Manufacturer?

So what does the FDA mean by a trusted manufacturer?

Firstly, a trusted manufacturer should ensure they have an efficient and proactive quality system in place. Some of the key quality-related areas to focus on are:

  • Design controls and required documentation to prove to the FDA exactly how the manufacturer has provided for the safety and efficacy of a system or device.
  • Design verification and validation establishing that what the manufacturer has built works for the end user as intended.
  • Risk management, ensuring that any applicable hazards have been identified and that mitigations have been implemented.
  • Iterative design reviews, allowing risks and omissions to be seen quicker, reducing total in-field corrective actions or bugs.

Several other considerations are critical in becoming a ‘trusted manufacturer’ under the anticipated guidance, some of which include:

  • The ability to follow good machine learning practices during all design stages.
  • Ensuring that all algorithm changes that are implemented are done according to pre-specified objectives and any applicable change protocols.
  • Trusted manufacturers are expected to document how the system will learn both pre- and post-release.
  • Ensuring the integrity of reference data used by continual algorithms.

This is an exciting time for medical and life science companies embarking on or continuing their ML product’s journey. Connect with us to find out how Jama can help!




Steven Meadows
Latest posts by Steven Meadows (see all)