Korea SaMD Approval Pathway for AI/ML Medical Devices

Korea Is Not the FDA, But Not the EU Either

Korea sits in an interesting middle ground for AI/ML-based Software as a Medical Device (SaMD). The MFDS framework borrows clinical evidence expectations from the FDA's risk-based approach, but the classification rules are closer to the EU MDR's Rule 11. The result is a hybrid pathway that does not perfectly map to either FDA or EU experience. Manufacturers extending an existing AI/ML product into Korea consistently underestimate the classification step and overestimate the clinical step.

Classification: Higher Than You Think

Korea's SaMD classification framework, codified in the 2022 SaMD Guideline and refined in 2025, classifies software based on:

1. Significance of information — Does the output inform a treatment decision, screen, drive treatment, or diagnose?
2. Healthcare situation — Is the condition critical, serious, or non-serious?

The two dimensions form a matrix that pushes most clinically meaningful AI software into Class II or Class III. An FDA Class II AI imaging triage device often classifies as Class II in Korea — that matches. An FDA-cleared SaMD that diagnoses (rather than triages) often classifies as Class III in Korea, requiring a clinical trial rather than a comparator study.

The classification opinion is the highest-leverage early deliverable. A 30-minute mis-mapping of the intended use phrase can push a product from a 6-month notification track to a 24-month approval track.

Clinical Evidence: Korean Patient Data Is Often Required

For Class II and III SaMD, MFDS expects clinical performance data on a Korean-representative dataset. This is not necessarily a Korean clinical trial — retrospective validation on a Korean dataset can suffice — but the reviewer will reject a submission supported only by US or European cohort data unless the manufacturer demonstrates that:

• The training dataset includes Korean samples, or
• A retrospective validation on Korean data shows performance parity with the original validation set, or
• A prospective clinical study at a Korean site is planned and registered.

This is where US-developed AI products run into delays. The training dataset rarely includes statistically meaningful Korean samples, so the validation route is the default. Building a Korean validation cohort takes 6–12 months when the manufacturer does not already have a Korean clinical partner.

The Hybrid Submission Format

The MFDS SaMD submission combines elements of the FDA 510(k) and the EU MDR technical file:

• Algorithm description — architecture, training methodology, training/validation/test split.
• Performance characterization — accuracy, sensitivity, specificity, AUC, with confidence intervals.
• Clinical performance — on Korean-representative data as described above.
• Cybersecurity packet — per the 2026 cybersecurity guidance (SBOM, threat model, vulnerability management).
• Quality system evidence — KGMP certification covering software development.
• Risk management file — ISO 14971 plus AI-specific risk considerations (data drift, retraining, edge-case failure modes).
• Labeling — including intended use scope, performance claims, and known limitations.

The full dossier for a Class II SaMD typically runs 400–600 pages. Class III dossiers run 800+ pages with the clinical trial protocol and report inline.

PCCP-Equivalent: Korea Has Not Adopted It Yet

The FDA's Predetermined Change Control Plan (PCCP) — pre-authorizing specific algorithm changes without resubmission — has no formal equivalent in Korea as of mid-2026. MFDS has indicated openness to a comparable framework and has issued a consultation, but until adopted, any material algorithm change requires a change notification that triggers re-review.

Practical implication: design your post-market update cadence assuming each material model retrain is a 60–90 day MFDS review. The architecture of the change control plan still matters — it sets the threshold for what counts as "material" — but you cannot rely on US-style pre-authorization.

Three Failure Modes

1. Predicate-thinking for AI. A 510(k) AI device with FDA clearance is not a sufficient comparator for an MFDS notification of the same algorithm. The Korean reviewer rebuilds the case from technical first principles.
2. English-language clinical evidence assumption. Performance reports must be translated. Native English papers in JAMA or Nature Medicine are accepted as supporting evidence but not as primary submission data.
3. Underbudgeting cybersecurity for SaMD. Pure software products are still subject to the full cybersecurity submission. There is no SaMD carve-out.

A Realistic Timeline

For an FDA-cleared Class II AI imaging device extending into Korea:

Phase	Duration
Classification opinion + intended use phrasing lock	2 weeks
Korean clinical validation dataset assembly	3–6 months
Document translation and reformat	2–3 months
MFDS submission and substantive review	5–8 months
Total	10–17 months

For a Class III diagnostic AI, add 6–12 months for the prospective clinical study.

Where Leanabl Plugs In

The Korea SaMD Approval solution runs classification, clinical strategy, and submission as one workstream — including Korean clinical site partnership for validation cohorts. The Submissions service coordinates parallel FDA and MFDS tracks for manufacturers running both. For software architecture and design control work that anchors the technical file, Discovery & Design plugs in upstream.