In early January, FDA released a draft guidance document titled Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (Draft AI Guidance). FDA has issued discussion papers (see our posts here and here ) and guidance on use of Pre-determined Change Control Plans (PCCPs) for devices enabled by artificial intelligence (AI) (see post here and stay tuned for a post on the newest final guidance). However, up to now the agency has said little about the content sponsors should include in an original marketing application for an AI-enabled medical device. As a consequence, sponsors of submissions for such devices have been surprised by lengthy requests for additional information from FDA during premarket review.
By way of background, the Draft AI Guidance defines AI-enabled devices as devices that include one or more AI-enabled device software functions (AI-DSF). An AI-DSF is a device software function that implements one or more “AI models” to achieve its intended purpose. An AI model is a mathematical construct that generates an inference or prediction based on new input data. The Draft Guidance acknowledges that there are some differences between the terminology used by FDA and the broader AI community, so it is important that all stakeholders review definitions and make sure they are speaking FDA’s preferred language when preparing device documentation. While the Draft AI Guidance does not include a glossary of AI-related terms, such a glossary is available on FDA’s website.
The Draft AI Guidance is lengthy, containing 64 pages of information and examples to help sponsors improve the quality of the device documentation. The Draft AI Guidance provides information on both recommendations on the documentation and information that should be included in marketing submissions as well as recommendations for the design, development, deployment, and maintenance of AI-enabled devices that may be generated as part of following quality system procedures and useful to support premarket authorization.
Appendix A of the Draft AI Guidance (copied here) provides a helpful overview of documentation that should be submitted in a marketing application for an AI-enabled medical device. More detailed descriptions are included in the body of the Draft AI Guidance, and other Appendices include detailed discussion of transparency design, performance validation, and usability evaluation considerations, and also provide examples of a Model Card and 510(k) Summary with Model Card.
Copy of Appendix A in Draft AI Guidance: Table of Recommended Documentation
| Guidance Section and Recommended Information | Recommended Section in Sponsor’s Marketing Submission |
| Section V Device Description | Device Description |
| Section VI.A User Interface | Software Description |
| Section VI.B Labeling | Labeling |
| Section VII Risk Assessment | Risk Management File of Software Documentation |
| Section VIII Data Management | Data for development: Software Description of Software Documentation
Data for testing: Performance Testing |
| Section IX Model Description and Development | Software Description |
| Section X.A Performance Validation | Clinical and non-clinical testing: Performance Testing
Software verification and software validation: Software testing as part of verification and validation of Software Documentation |
| Section XI Device Performance Monitoring | Risk Management File of Software Documentation |
| Section XII Cybersecurity | Cybersecurity |
| Section XIII Public Submission Summary | Administration Information |
Ensuring transparency and reducing bias over the total product lifecycle are recurring themes of the Draft AI Guidance. AI-enabled devices may pose challenges to user understanding due to the opacity of many models. Transparency of AI-DSF information to device users ensures that important information is both accessible and functionally comprehensible. Bias, in the context of AI, refers to the potential for an AI-enabled device to produce incorrect results in a systematic, but sometimes unforeseeable way, such as when an AI-model relies on data correlations that do not map to biologically plausible mechanisms of action. This can affect safety and effectiveness of the AI-enabled device in all or a subset of the intended use population.
The Draft AI Guidance discusses data management practices at length, which is not surprising as data management practices are an important means for identifying and mitigating bias and the performance and behavior of AI-enabled devices rely heavily on the quality, quantity and diversity of data used to train and tune the AI-DSF. For both training and test data, the sponsor should provide information on data collection, the reference standard (representative truth), data annotation, data storage, management and independence of data, and representativeness. For training data only, information on data cleaning and processing should also be provided (i.e., test data should not be cleaned). Details of model development should also be provided in a marketing submission.
The AI Draft Guidance includes a section on validation of AI-enabled devices, stating that validation should demonstrate both the ability of the device to meet performance specifications and the ability of users to interact with and understand the device. Performance validation should use data collected from different sites than were used for collection of training data. A statistical analysis plan, including a plan for subgroup analysis, is also recommended to pre-specify plans to analyze validation results. Appropriate subgroups for analysis will vary based on the intended use of the device but should generally include patient sex, gender, age, race, ethnicity, disease variables, clinical data site, data acquisition equipment, and, if applicable, conditions of use (including skill level of the user when relevant), device configurations, and other relevant confounding factors that may impact device performance. Validation of an AI-enabled device should also include, when feasible and appropriate, an evaluation of its repeatability and reproducibility, which may include testing using phantom, simulated, contrived, or clinical data.
The Draft AI Guidance includes recommendations for postmarket device performance monitoring. FDA notes that performance of AI-enabled devices may change over time in the real-world environment and states that sponsors may include a monitoring plan in the marketing application to support FDA’s evaluation of risk controls. While acknowledging that information on a sponsor’s quality system regulation (QSR) compliance is not generally included in a 510(k) submission, the Draft AI Guidance notes that it may be appropriate for the Agency to review information about the sponsor’s quality system during review of a 510(k) submission. The Draft AI Guidance states that inclusion of a postmarket device performance plan in a marketing application is an option for providing “reasonable assurance of the device’s safety and effectiveness” and “to ensure adequate ongoing performance . . . [to] support a determination of substantial equivalence.” That being said, the Draft AI Guidance also acknowledges that for a 510(k) submission, a postmarket performance plan would not be required absent a special control calling for the plan. For a De Novo classification request, a postmarket performance plan may be required by FDA and included as a special control for the device type going forward. Finally, a performance monitoring plan may also be a condition of approval for devices subject to premarket approval. The Agency is encouraging sponsors to include a postmarket performance plan even when not strictly required and recommends sponsors discuss these plans in a pre-submission.
To support transparency of the AI-enabled device, the Draft AI Guidance discusses content of device labeling and the public submission summary. A Model Card is noted as a means of providing appropriate information in both places. A Model Card is a short document that provides key information about an AI-model. Model cards are used in the broader AI-industry and may be helpful to communicate information about an AI-enabled device. FDA recommends that an AI model card include device identification information, device regulatory status, a description of the device’s use, a description of the device’s performance and limitations, a discussion of potential risks, and a description of data used to develop the device.
FDA has been reviewing AI-enabled devices for years and sponsors of many of these submissions have received lengthy pre-submission feedback and requests for additional information related to the topics covered in the Draft AI Guidance. Given this, having FDA’s expectations laid out in the Draft AI Guidance is a good step, though the level of detail required for a submission for an AI-enabled device is extensive and burdensome. FDA’s expectations for data management practices seem particularly challenging and will be discussed in a follow-up post.
FDA plans to host a webinar to discuss the Draft Guidance. It was originally scheduled for Tuesday, February 18, 2025, but has recently been postponed. Comments on the Draft AI Guidance are due by April 7, 2025.
