Description
This draft guidance document provides recommendations regarding the contents of marketing submissions for devices that include artificial intelligence (AI)-enabled device software functions including documentation and information that will support FDA’s evaluation of safety and effectiveness. The recommendations reflect a comprehensive approach to the management of risk throughout the device total product life cycle (TPLC). To support the development of appropriate documentation for FDA’s assessment of the device, this draft guidance also proposes recommendations for the design, development, and implementation of AI-enabled devices that manufacturers may wish to consider using throughout the TPLC.
Scope & Applicability
Product Classes
5Devices that include one or more AI-enabled device software functions; medical devices utilizing artificial intelligence; Devices incorporating artificial intelligence software functions.; The model is part of the mechanism of action for an AI-enabled device.; validation includes ensuring that the device will perform its intended use safely and effectively; devices particularly susceptible to unexpected differences in performance due to data reliance; Ongoing performance monitoring is important
Model Name: Disease X Screening Model
devices administered by a licensed practitioner
The guidance provides recommendations on documentation for marketing submissions for AI-enabled devices.
A product comprised of two or more regulated components.
Stakeholders
7Entity responsible for submitting applications under section 524B
Assist sponsors in the nonclinical evaluation
Users whose characteristics impact user needs and communication format.; Intended users of the AI-enabled device; Intended User: Healthcare professionals
Way questions are framed is critical to collecting unbiased patient input
Provides clinical judgment and conducts assessments for ClinROs.
It may be a hardship for patients and/or caregivers to attend in-person; Individuals who may assist patients or provide observations when patients cannot self-report.; conducted with caregivers to obtain similar feedback
Entity responsible for submitting NDINs
Regulatory Context
Attributes
10Strategy to address transparency and bias throughout the TPLC
sponsors should include confidence intervals on all reported results
Performance metric achieved by the model
Characteristic of AI models that impacts user understanding
Ability to detect intended mechanism of action without interference; Performance characteristic to be validated
Analysis by sex of clinical performance measures such as sensitivity
Metric describing predictions and uncertainty in the model.
If OUS data are used during validation, an explanation regarding how the data compares to the U.S. population
Consideration for study sampling in research protocols
Characteristics of the development data population.
Identified Hazards
Hazards
9Potential tendency to produce incorrect results in a systematic way.; AI bias is a potential tendency to produce incorrect results in a systematic way due to training data limitations.; Understanding the methods used to develop the model also helps FDA identify potential limitations, sources of AI bias; spurious learnings could impact performance differentially across patient demographics; Risk associated with patient populations not well represented in training
Inclusion of cybersecurity risks as part of informed consent form
manipulation of training data to introduce or accentuate biases; incorrect follow-up due to a false positive or false negative output, which can occur because of model bias
Cyber attackers could modify input samples to deceive models
Cyber threats could intentionally use forged data to replicate models
Cyber threats could lead to data poisoning
Data leakage between validation and development datasets can create uncertainty regarding true performance.; exploit vulnerabilities to access sensitive training data
Risks resulting from mistakes made by the operator.
Occurs when systems that produce inputs for AI-enabled devices change over time.; consider the impact of factors (e.g., data drift) on the safety and effectiveness
Related CFR Sections (14)
- 21CFR801.5§ 801.5 Medical devices; adequate directions for use.
Adequate directions for use means directions under which the layman can use a device safely and for the purposes for which it is intended. Section 801.4 defines intended use. Directions for use may be inadequate because, among other reasons, of omission, in whole or in part, or incorrect specificatiRead full regulation →
- 21CFR814.9§ 814.9 Confidentiality of data and information in a premarket approval application (PMA) file.
(a) A “PMA file” includes all data and information submitted with or incorporated by reference in the PMA, any IDE incorporated into the PMA, any PMA supplement, any report under § 814.82 , any master file, or any other related submission. Any record in the PMA file will be available for public discRead full regulation →
- 21CFR860.220§ 860.220 De Novo request content.
(a) Unless the requester justifies an omission in accordance with paragraph (c) of this section, a De Novo request must include:Read full regulation →
- 21CFR860.200§ 860.200 Purpose and applicability.
(a) The purpose of this part is to establish an efficient, transparent, and thorough process to facilitate De Novo classification into class I or class II for devices for which there is no legally marketed device on which to base a review of substantial equivalence and which meet the definition of cRead full regulation →
- 21CFR807.81§ 807.81 When a premarket notification submission is required.
(a) Except as provided in paragraph (b) of this section, each person who is required to register his establishment pursuant to § 807.20 must submit a premarket notification submission to the Food and Drug Administration at least 90 days before he proposes to begin the introduction or delivery for inRead full regulation →
- 21CFR807.93§ 807.93 Content and format of a 510(k) statement.
- 21CFR807.92§ 807.92 Content and format of a 510(k) summary.
(a) A 510(k) summary shall be in sufficient detail to provide an understanding of the basis for a determination of substantial equivalence. FDA will accept summaries as well as amendments thereto until such time as FDA issues a determination of substantial equivalence. All 510(k) summaries shall conRead full regulation →
- 21CFR814.82§ 814.82 Postapproval requirements.
(a) FDA may impose postapproval requirements in a PMA approval order or by regulation at the time of approval of the PMA or by regulation subsequent to approval. Postapproval requirements may include as a condition to approval of the device:Read full regulation →
- 21CFR814.44§ 814.44 Procedures for review of a PMA.
(a) FDA will begin substantive review of a PMA after the PMA is accepted for filing under § 814.42 . FDA may refer the PMA to a panel on its own initiative, and will do so upon request of an applicant, unless FDA determines that the application substantially duplicates information previously revieweRead full regulation →
- 21CFR830.50§ 830.50 Changes that require use of a new device identifier.
(a) Whenever you make a change to a device that is required to bear a unique device identifier (UDI) on its label, and the change results in a new version or model, you must assign a new device identifier to the new version or model.Read full regulation →
- 21CFR801.20§ 801.20 Label to bear a unique device identifier.
(a) In general.Read full regulation →
- 21CFR4.2§ 4.2 How does FDA define key terms and phrases in this subpart?
The terms listed in this section have the following meanings for purposes of this subpart:Read full regulation →
- 21CFR3.2§ 3.2 Definitions.
For the purpose of this part:Read full regulation →
- 21CFR820.3§ 820.3 Definitions.
The definitions in ISO 13485 and in Clause 3 of ISO 9000 (incorporated by reference, see § 820.7 ) apply to this part, except as specified in paragraph (b) of this section, and do not affect the meaning of similar terms defined in this title.Read full regulation →
Related Warning Letters (10)
- 2025-12-09
CGMP/QSR/Medical Devices/Adulterated
Envoy Medical Inc.
- 2025-11-25
CGMP/QSR/Medical Devices/Adulterated
Hong Qiangxing Shenzhen Electronics Limited
- 2025-10-28
CGMP/QSR/Medical Devices/Adulterated
Contec Medical Systems Co., Ltd.
- 2025-10-28
CGMP/QSR/Medical Devices/Adulterated/Misbranded
Royal Philips
- 2025-10-28
CGMP/QSR/Medical Devices/Adulterated/Misbranded
Qianjiang Kingphar Medical Material Co Ltd.
- 2025-10-21
CGMP/QSR/Medical Devices/Adulterated
LEVO AG
- 2025-10-07
CGMP/QSR/Medical Devices/Adulterated
Technological Medical Advancements LLC
- 2025-09-30
Medical Device Reporting/Misbranded
Insung Medical Co. Ltd.
- 2025-09-23
Medical Device/Adulterated/Misbranded/Lacks PMA and/or 510(k)
The Richline Group, Inc.
- 2025-09-16
Investigational Device Exemptions (IDE)/Premarket Approval Application (PMA) Adulterated Device
SeniorLife Technologies, Inc.
See Also (8)
- CPG Sec. 396.400 Policy on Warned on Sunlamp Products (Status: Final)
- CPG Sec. 300.300 Ineffective Devices - 502(f)(I) Labeling Requirements (Status: Final)
- Policy on Warning Label Required on Sunlamp Products (Status: Final)
- Labeling - Regulatory Requirements for Medical Devices (FDA 89-4203) (Status: Final)
- Guidance Document For Nonprescription Sunglasses - Guidance for Industry (Status: Final)
- Content and Format of Premarket Notification [510(k)] Submissions for Liquid Chemical Sterilants/High Level Disinfectants - Guidance for Industry and FDA Reviewers (Status: Final)
- Labeling Recommendations for Single-Use Devices Reprocessed by Third Parties and Hospitals: Final Guidance for Industry and FDA (Status: Final)
- Distinguishing Medical Device Recalls from Medical Device Enhancements: Guidance for Industry and Food and Drug Administration Staff (Status: Final)