IMDRF finalizes good machine learning practice, software risk documents
The International Medical Device Regulators Forum (IMDRF) on Wednesday finalized two highly anticipated technical documents detailing guiding principles for good machine learning practices (GMLP) and risk characterization for medical device software.GMPL principles
The final IMDRF document lists ten guiding principles for good machine learning practices. They are practically identical to principles presented by the US, UK, and Canada, which are IMDRF members, in 2021. (RELATED: Regulators release 10 principles for good machine learning practice, Regulatory Focus 27 October 2021)
The principles include understanding the product's intended use; implementing good software, design, and security practices; using population-representative datasets during clinical evaluation; and training the datasets independently of test sets. An overarching theme in the principles is to take a total product lifecycle (TPLC) approach to the products so that they not only meet premarket regulatory needs but are monitored after they're allowed on the are allowed on the market. (RELATED: IMDRF publishes AI/ML guiding principles echoing US, UK, Canadian regulators, Regulatory Focus, 1 July 2024)
The final version has minor changes compared to the draft document, which was published in July 2024. One notable change is that when discussing using good software engineering, medical device design, and security practices throughout the lifespan of the product, manufacturers should also ensure their software is able to reproduce results. The group also noted that they should consider the infrastructure needed for model deployment, monitoring, and maintenance of the product.
"These practices help support the rights, safety, and welfare of patients, including through the ethical use of patient data," said the final document.
While the draft guidance focused on the human-AI team in the intended use environment when assessing performance, the final guidance also emphasizes that performance isn't just about assessing the device but needs to factor in its interactions with healthcare providers, patients, and caregivers.
Software risk characterization
IMDRF’s final document on considerations for device and risk characterization for medical device software contains numerous revisions to the draft version released in January 2024.
One of the changes to the final document is a reminder that software may change throughout a product’s service life, and that risk characterization should be revisited accordingly as software is updated.
The document also notes that the considerations listed within “are not intended to be used by stakeholders as a checklist or prescriptive means of medical device software characterization or determining software-specific device risks.”
Additionally, IMDRF stresses that the document is not intended to replace its 2014 document, “Software as a Medical Device”: Possible Framework for Risk Categorization and Corresponding Considerations, but rather to serve as a supplement to the framework presented in that document.
IMDRF GMLP, IMDRF Software Risk Characterization
