Guidance on the Review and Approval of Artificial Intelligence (AI)-based Medical Devices in South Korea

1. What this guidance is and why it matters

South Korea’s Ministry of Food and Drug Safety (MFDS) published a dedicated guidance to support consistent review and approval of AI-based medical devices, reflecting the fact that machine learning–driven performance can be sensitive to data, workflow, and software changes. Source: MFDS (English) – Guidance on the Review and Approval of Artificial Intelligence (AI)-based Medical Devices ([mfds.go.kr][1]) Source: MFDS Guidance (attached PDF)

2. Scope and core review logic

The guidance focuses on AI-based medical devices that use machine learning and sets expectations for what reviewers need to confirm:

• the product is a medical device (vs. non-medical software),
• the intended use and clinical context are unambiguous, and
• evidence shows safety and performance are appropriate for the intended users and environment. Source: MFDS Guidance (attached PDF)

3. Determining whether an AI software product is a “medical device”

A practical early step is to determine whether the software qualifies as a medical device. The guidance highlights that some AI software may be non-medical, and therefore outside medical-device approval pathways, even if it is used in healthcare settings (e.g., administrative or general wellness functions). Source: MFDS Guidance (attached PDF)

What you should document for MFDS review

• Intended medical purpose (diagnosis, triage, treatment support, etc.) and target patient population
• User profile (e.g., radiologist, general practitioner) and use environment
• Input data type(s), outputs, and how outputs should be interpreted (including limitations) Source: MFDS Guidance (attached PDF)

4. Technical file expectations: algorithm, data, and verification/validation

MFDS expects AI-based submissions to include enough detail to evaluate model development, verification, and validation, such as:

• device description, operating principle, and software overview
• dataset description (source, representativeness, labeling/ground truth approach)
• verification and validation (V&V) evidence demonstrating the product meets specified requirements Source: MFDS Guidance (attached PDF)

Where AI output quality depends on clinical workflow, MFDS expects performance evaluation aligned to the intended clinical task, with appropriate performance metrics and clearly described test procedures. Source: MFDS Guidance (attached PDF)

5. Clinical performance evidence and comparators

A key review question is whether the device delivers clinically meaningful performance in the intended population and setting. The guidance discusses performance evaluation approaches, including how to use clinical datasets and how to justify the evaluation design. Source: MFDS Guidance (attached PDF)

The guidance also describes an approach using comparison/equivalence to a pre-approved product in limited scenarios, provided that differences in intended use, algorithm characteristics, and data handling are appropriately justified and do not introduce new risks. Source: MFDS Guidance (attached PDF)

6. Lifecycle controls: versioning and change management

Because AI devices are software-reliant, MFDS places emphasis on lifecycle governance, including:

• version control of the AI model and software, and
• clear policies for managing changes that could affect performance or safety. Source: MFDS Guidance (attached PDF)

The guidance notes that certain changes may be subject to change approval or change certification, while some changes can be exempt depending on their nature and impact (as described in the guidance). Source: MFDS Guidance (attached PDF)

For training-data governance, the guidance highlights the need for a training dataset management policy, reflecting the risk that dataset shifts can change clinical performance. Source: MFDS Guidance (attached PDF)

7. Cybersecurity and international alignment

AI-based medical devices commonly involve connectivity, cloud deployment, or frequent software updates, making cybersecurity part of a credible safety case. While MFDS has jurisdiction-specific expectations, many developers also align their evidence packages with international best-practice resources. Source: IMDRF – Principles and Practices for Medical Device Cybersecurity (N60, 2020) ([imdrf.org][2])

For broader “good ML practice” concepts (data quality, evaluation, monitoring, and transparency), the IMDRF consultation materials can also be used as a reference point when structuring documentation and lifecycle controls. Source: IMDRF Consultation – Good Machine Learning Practice (GMLP) ([imdrf.org][3]) Source: IMDRF GMLP Draft (N73)

8. Practical checklist for an MFDS-ready submission

1. Define intended use with clear clinical workflow boundaries and limitations.
2. Describe the AI (model type, inputs/outputs, operating principle) and explain interpretability where relevant.
3. Document datasets (training/validation/test), labeling/ground truth, representativeness, and bias considerations.
4. Provide V&V evidence tied to requirements and clinical task performance.
5. Justify clinical performance evaluation (population, setting, metrics, acceptance criteria).
6. Implement versioning, change control, and dataset governance suitable for AI performance sensitivity. Source: MFDS Guidance (attached PDF)