The FDA published final guidance on artificial intelligence and machine learning-based Software as a Medical Device (AI/ML SaMD) in October 2025, completing a regulatory development process that began with the agency's 2019 discussion paper and has included multiple draft guidance documents, public workshops, and international harmonization efforts through the International Medical Device Regulators Forum (IMDRF). The final guidance establishes a comprehensive framework for the lifecycle management of AI/ML SaMD, with particular emphasis on the Total Product Lifecycle (TPLC) approach and the role of Predetermined Change Control Plans (PCCPs) in managing iterative algorithm modifications.
Scope and Applicability
The guidance applies to software that meets the definition of a device under the FD&C Act and that uses AI or ML techniques - including machine learning, deep learning, neural networks, natural language processing, and related methods - to perform or inform a clinical function. Software that uses only rule-based algorithms, look-up tables, or simple statistical methods is generally outside the scope of the AI/ML SaMD framework. The guidance distinguishes between "locked" algorithms, which do not change after deployment without a new marketing submission, and "adaptive" algorithms, which can modify their behavior based on new data - and applies different regulatory expectations to each category.
The Total Product Lifecycle Framework
The TPLC approach requires manufacturers to think about AI/ML SaMD not as a product to be approved once but as a continuously evolving system whose safety and effectiveness must be monitored and maintained throughout its market life. The key TPLC elements FDA expects manufacturers to address include:
- Reference datasets: Transparent documentation of the training, tuning, and validation datasets used to develop the algorithm, including demographic characteristics and geographic provenance
- Algorithmic transparency: Documentation of model architecture, training methodology, and performance metrics in terms that support FDA review and post-market monitoring
- Real-world performance monitoring: Defined metrics and monitoring intervals for assessing whether the algorithm's real-world performance matches its validated performance envelope
- Feedback mechanisms: Processes for incorporating real-world performance data into decisions about algorithm updates and retraining
Good Machine Learning Practice (GMLP)
The final guidance formally adopts the Good Machine Learning Practice (GMLP) framework developed through collaboration between FDA, Health Canada, and the UK MHRA. GMLP consists of ten foundational principles covering multi-disciplinary team expertise, defined development and testing datasets, device performance tailored to clinically meaningful tasks, human-AI team performance considerations, and transparent device development processes. Manufacturers developing AI/ML SaMD should use the GMLP principles as a design control framework, documenting how each principle is addressed in their design history file.
Regulatory Submission Considerations for AI/ML SaMD
From a submission strategy perspective, the final guidance reinforces that the appropriate marketing submission pathway for AI/ML SaMD - whether 510(k), De Novo, or PMA - is determined by the device's risk classification and intended use, not by the fact that it uses AI or ML. What changes with AI/ML SaMD is the content expected within the submission, including algorithm description, dataset documentation, performance testing across relevant subgroups, and, where applicable, a PCCP. Manufacturers should engage FDA early through the Q-Sub program to align on submission content expectations before investing in a full submission package.
Frequently Asked Questions
What is the Total Product Lifecycle (TPLC) approach for AI/ML SaMD?
The TPLC approach requires manufacturers to treat AI/ML SaMD not as a product approved once, but as a continuously evolving system whose safety and effectiveness must be monitored and maintained throughout its market life. Key TPLC elements include transparent documentation of training, tuning, and validation datasets; defined real-world performance monitoring metrics and intervals; and processes for incorporating real-world performance data into decisions about algorithm updates and retraining.
What is Good Machine Learning Practice (GMLP) and how does it apply to device development?
GMLP is a framework of ten foundational principles developed through collaboration between FDA, Health Canada, and the UK MHRA, covering topics such as multi-disciplinary team expertise, defined development and testing datasets, device performance tailored to clinically meaningful tasks, and transparent development processes. The final FDA guidance formally adopts GMLP, and manufacturers should use these principles as a design control framework, documenting how each principle is addressed in their design history file.
Does using AI or ML change the marketing submission pathway for a SaMD device?
No. The appropriate marketing submission pathway - whether 510(k), De Novo, or PMA - is determined by the device's risk classification and intended use, not by the fact that it uses AI or ML. What changes with AI/ML SaMD is the content expected within the submission, including algorithm description, dataset documentation, performance testing across relevant subgroups, and, where applicable, a Predetermined Change Control Plan (PCCP). Early engagement with FDA through the Q-Sub program is recommended to align on submission content expectations.