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
| Requirement | Details |
|---|---|
| Algorithm description | Full description of the model architecture, training methodology |
| Training data | Characteristics, size, representativeness, labelling quality |
| Validation data | Independent test dataset; performance metrics (sensitivity, specificity, AUC, etc.) |
| Clinical validation | Real-world performance in the intended use population |
| Explainability | Explainability and interpretability: For Grade IIIโIV 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 statement | Precise statement of what the AI is intended to do and in which setting |
Post-market requirementsโ
| Requirement | Details |
|---|---|
| Performance monitoring plan | Ongoing monitoring of real-world AI performance |
| Drift detection | Monitoring for algorithm performance degradation over time |
| Re-validation trigger | Define when re-validation is required (performance threshold breach) |
| Reporting | Integrate 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 ํ๋ชฉํ๊ฐ submission.
Related pagesโ
Risk-based explainability: MFDS requires explainability documentation proportionate to the device's risk classification. For Grade IIIโIV 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.