Skip to main content

AI/ML Medical Device Guidance

Korea has been at the forefront of AI/ML medical device regulation, with MFDS issuing initial guidance in 2017 and updated guidance in 2021 and 2024.

Core requirements (2024 guidance)โ€‹

Pre-market requirementsโ€‹

RequirementDetails
Algorithm descriptionFull description of the model architecture, training methodology
Training dataCharacteristics, size, representativeness, labelling quality
Validation dataIndependent test dataset; performance metrics (sensitivity, specificity, AUC, etc.)
Clinical validationReal-world performance in the intended use population
ExplainabilityExplainability and interpretability: For Grade IIIโ€“IV AI/ML devices, manufacturers must provide documentation explaining how the model makes diagnostic or treatment decisions, suitable for review by clinicians and regulators. This may include feature importance analysis, decision rules, or other interpretability techniques proportionate to the model's complexity.
Intended use statementPrecise statement of what the AI is intended to do and in which setting

Post-market requirementsโ€‹

RequirementDetails
Performance monitoring planOngoing monitoring of real-world AI performance
Drift detectionMonitoring for algorithm performance degradation over time
Re-validation triggerDefine when re-validation is required (performance threshold breach)
ReportingIntegrate AI performance data into PMS and periodic safety reports

Predetermined Change Control Plan (PCCP)โ€‹

MFDS is developing PCCP guidance allowing manufacturers to pre-define the scope of future algorithm changes that do not require a new ํ’ˆ๋ชฉํ—ˆ๊ฐ€ submission.

Risk-based explainability: MFDS requires explainability documentation proportionate to the device's risk classification. For Grade IIIโ€“IV AI/ML devices or those used in high-stakes clinical decisions (e.g., cancer diagnosis), full model explainability is mandatory. Lower-risk applications may require performance metrics documentation instead of detailed model architecture explanation.