FDA AI Device TPLC: Submissions — Description, UI & Risk

This article summarises FDA's draft recommendations for three interconnected marketing submission sections: device description, user interface and labeling, and risk assessment. For Hong Kong healthcare professionals, these sections define much of what should appear in instructions for use (IFU), model cards, and risk management summaries from U.S.-regulated AI devices.

Device description

The device description helps FDA understand intended use, clinical workflow, use environment, model features, and overall design. Sponsors should generally include:

• A statement that AI is used in the device
• Inputs and outputs — manual vs automatic entry, compatible input devices and acquisition protocols
• How AI achieves intended use — including interactions between multiple AI-DSFs and non-AI functions
• Intended users — qualifications, training, and roles (clinicians interpreting output, technicians, patients, caregivers, administrators)
• Intended use environment — hospital, clinic, home, etc.
• Intended clinical workflow — degree of automation vs standard of care, clinical circumstances leading to use, how outputs support decisions
• Installation, maintenance, calibration, and configuration procedures

For user-configurable devices, sponsors should describe configurable elements (visualisations, thresholds, operating points), who configures them, configuration level (patient, site, network), and potential impact on decision-making.

Hong Kong takeaway

When procuring an AI tool, ask whether vendor documentation clearly states who may use it, where it was validated, and how outputs fit your workflow — not only headline accuracy.

User interface and labeling

The user interface includes all points of interaction: displays, controls, alarms, packaging, labeling, and training. FDA recommends integrating important information throughout the interface so users receive context at the right time — not only in a static manual.

Submissions should include graphical representations, workflow overviews, example outputs across expected outcome ranges, and optionally video demonstrations.

Labeling must satisfy U.S. labeling regulations and, for AI devices, should address:

• How AI is used and model inputs/outputs
• Compatible devices, acquisition protocols, and impact of lost inputs
• Intended degree of automation
• High-level model architecture and development/validation data characteristics (sources, sites, sample sizes, demographics, reference standards)
• Performance metrics with confidence intervals, subgroup performance, and operating points
• Performance monitoring tools and instructions when ongoing user monitoring is necessary
• Known limitations — including under-represented populations or rare disease presentations
• Installation, workflow integration, and customisation instructions
• Patient/caregiver materials at appropriate reading levels when applicable

FDA encourages transparency design (Appendix B) and optional model cards (Appendix E) to organise this information consistently.

For clinicians

Labeling is where you should find subgroup performance, dataset limitations, and how to interpret uncertainty. If these are absent or vague, that is a governance red flag — even if the algorithm performs well in a developer's benchmark.

Risk assessment

A comprehensive risk management file — plan, assessment, and report — should address risks across the TPLC. FDA recommends ISO 14971 and AAMI CR34971 (AI/ML-specific application of ISO 14971).

AI-enabled devices introduce or amplify risks related to understanding information:

• Performance in disease subtypes may not be apparent
• Model logic may be non-intuitive
• Users may not recognise when the device underperforms
• Unclear instructions may lead to misuse

Risk assessment should cover the full continuum of use: installation, maintenance, interpretation of results, and performance over time. When labeling or UI elements serve as risk controls, sponsors should explain how those controls address identified risks.

Usability evaluation (Appendix D) may support assessment of information-related risk controls.

Quality system linkage

Much submission content may originate in Quality System documentation — design controls, design validation under simulated use conditions, design change control, CAPA, and management review. FDA explains how QS outputs can demonstrate performance-related evidence in premarket review.

Source: U.S. FDA — Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations (January 2025, draft guidance)