Korea's food and drug safety evaluation institute approved seven digital therapeutic clinical trial protocol guides on 10 September 2026, and the operative text of every one of them says the minimum clinically important difference is set in advance with its basis stated in the protocol, in documents that disclaim their own legal effect on page three

Seven documents, not one guideline

The announcement speaks of a guideline in the singular. The board holds seven separate civil applicant guides, one per indication, each with its own registration number and its own page count. We downloaded and opened all seven: alcohol use disorder (1177-02), nicotine use disorder (1178-02), depressive disorder (1258-02), panic disorder (1259-02), attention deficit hyperactivity disorder (1333-02), eating disorder (1334-02) and mild cognitive impairment (1525-01). Every one of them carries an approval date of 2026. 9. 10. and a signature block of the same date.

Six of the seven are second revisions of an existing document. The predecessors were approved on three separate dates: 2021. 12. 8. for alcohol and nicotine, 2023. 12. 7. for panic, eating and ADHD, and 2024. 12. 30. for depression. That is the actual shape of the exercise, a consolidation pass over a family that had accumulated over five years.

The mild cognitive impairment document is the exception and it is worth stating precisely. Its registration number is 1525-01, a fresh number at revision 01, not 02 like the rest. Its revision history table nevertheless lists a predecessor as 1258-01, approved 2024. 12. 30., which is the identical number and date the depressive disorder document gives for its own predecessor. We report both tables as they read. On their face one of them is a clerical carry-over, and we do not know which.

There is one more per-document variation. Six of the seven describe their second change item as adding efficacy evaluation variables and follow-up observation settings. The depressive disorder document says only follow-up observation settings, omitting the efficacy variable clause its six siblings carry.

What the MCID clauses actually say

Section 14 of each document covers efficacy evaluation criteria, methods and interpretation. It separates statistical significance from clinical significance and gives the second its own line. In the ADHD document the clause reads, in Korean: "임상적 유의성: 임상적으로 중요한 최소 변화량은 사전에 설정하고 그 근거를 임상시험계획서에 명시하며, 통계적 유의성과 함께 종합적으로 고려하여 유효성을 해석한다." Our translation, labelled as a translation and not as the text: clinical significance: the minimally clinically important change is set in advance and its basis is stated in the clinical trial protocol, and efficacy is interpreted by considering it together with statistical significance.

Section 8, on subject numbers and selection criteria, carries the operational half. Within its numbered list of the information a sample size calculation needs, the last item adds, in the same ADHD document: "특히 확증임상시험 및 초기 유효성 평가를 위한 탐색임상시험의 경우에는 설정된 평가변수에 대한 임상적으로 중요한 최소 변화량(Minimally Clinically Important Difference, MCID)기반의 군간 효과크기(Effect size)를 토대로 대상자 수를 확정하여야 한다." Our translation: in particular, for confirmatory trials and for exploratory trials evaluating early efficacy, the number of subjects is to be fixed on the basis of the between-group effect size derived from the minimally clinically important difference for the chosen endpoint. That is the clause that gives the threshold teeth in practice. It sits upstream of the enrolment number, so a change in the MCID changes the size of the study.

Then the documents decline to supply a number. Each worked example of a success criterion leaves the threshold as a blank. The ADHD example reads that the trial is judged successful where the change in the ADHD rating scale total from baseline to post-treatment is at or above the MCID of two open circles, is in the direction superior to control, and has a p-value below the 0.05 significance level. The two circles are literal placeholder characters in the PDF. The depressive disorder example places the same blank against the Korean Beck Depression Inventory second edition, and the mild cognitive impairment example against the Clinical Dementia Rating sum of boxes.

The note under each of those examples says the MCID that has been set, and its basis, are to be submitted and stated in the protocol. Two variants of that sentence are in circulation across the seven: the ADHD, nicotine, panic and eating documents use one particle, the alcohol and mild cognitive impairment documents use another, and the meaning is the same. So the posture is consistent across the family: the regulator fixes the procedure and leaves the number to the sponsor, against a justification it expects to see in the protocol rather than at the analysis stage.

One indication goes further. The eating disorder document defines the responder rate using the same threshold, as the proportion of subjects whose questionnaire score falls by at least the pre-defined MCID from baseline, or by at least a set proportion. There the threshold is not only a success criterion but a definition inside a secondary endpoint.

The documents disclaim their own imperative voice

This is the part a reader should not have to take on trust, because the guides are written throughout in the shall-form and would otherwise read as rules. Page three of each document says, in Korean: "본 안내서는 대외적으로 법적 효력을 가지는 것이 아니므로 본문의 기술방식('∼하여야 한다' 등)에도 불구하고 참고로만 활용하시기 바랍니다." Our translation: this guide does not have external legal effect, so notwithstanding the drafting style used in the body (expressions such as shall), please use it for reference only.

The same front matter carries a completed compliance checklist. One line asks whether the document contains anything that departs from superior legislation to create or strengthen new regulation, or that binds the civil applicant. In all seven the box marked no is the one that is ticked. Another line asks whether the document explains law or administrative rules in plain terms or states the agency's external position for the benefit of civil applicants; the yes box is ticked, which is what classifies it as a guide rather than an internal instruction for officials.

So the accurate reading is narrow and useful at the same time. No sponsor is in breach of anything by setting an MCID differently, or by not setting one. What the sponsor faces instead is a review division working from a document that assumes the number exists and is justified in the protocol, and a design that departs from it has to carry the argument.

Nothing in the seven documents states a transitional provision, a deadline, a comment period, or any position on protocols already submitted or trials already running. We looked for all four.

Where the algorithm actually appears in the text

We can now be precise about the AI question rather than gesture at it. Across the documents we read, the term artificial intelligence appears three times in each, and every occurrence is the name of the issuing office, the Medical Artificial Intelligence Software Division, in the header, the contact block and the enquiries line. The body does not discuss artificial intelligence, and we are not going to pretend otherwise.

The word algorithm appears twice in the body of each document, and both places are substantive. The first is in the mechanism of action section, which asks for the scientific and clinical basis for the product, expressly including the treatment mechanism and algorithm, described together with the detailed functions of the program and the specific characteristics of the digital technology applied. The listed acceptable evidence types are clinical practice guidelines recognised by the Korean Academy of Medical Sciences, clinical papers in peer-reviewed journals, and exploratory or investigator-initiated trial data.

The second is the definition of the sham comparator, and it is the sharpest line in these documents for anyone building software as a medical device: "대조군(Sham) 애플리케이션은 시험군과 유사한 외형, 사용 환경 및 조작 방식을 제공하되, 치료효과를 유발하는 핵심 치료 알고리즘 또는 중재 요소는 포함하지 않도록 설계" Our translation: the sham application is designed to provide an appearance, usage environment and operating method similar to the test group, while not containing the core therapeutic algorithm or intervention element that produces the treatment effect. The control arm is defined by subtracting the algorithm. That is the evidentiary consequence of shipping a therapy as software, stated in one line.

The ADHD document adds that a no-treatment or wait-list control is possible depending on design but requires scientific justification, and that for a double-blind design a sham device control is recommended in principle, with a separate account of blinding maintenance and bias minimisation required if a no-treatment control is used.

What we did not verify

Quotations from the guides are reproduced with ASCII apostrophes in place of the typographic single quotation marks the documents use around the phrase inside the disclaimer. That is a house typesetting convention and not a change to any word, character or particle of the Korean.

What we opened, all this session: the MFDS press materials board entry at seq 50336 and its attached PDF press reference note; the MFDS civil applicant guides board listing; and all seven guideline PDFs, downloaded from that board and read as text. The seven are 67 to 76 pages each. Every quotation in this article is from one of those PDFs and is identified by document.

The material limitation, stated plainly: we did not obtain the superseded editions of any of the seven. So we cannot say which of the MCID clauses is new in this revision and which was carried over from the predecessor. None of the seven revision history tables itemises MCID. MFDS's press note does place the setting of efficacy evaluation variables and of the minimum clinically important difference among the current review criteria the revision reflects, and that sentence is the authority's own description, which we have not tested against an earlier text.

We did not open the HWPX versions, only the PDFs, so we cannot say whether the two differ. We did not open the Digital Medical Products Act, its enforcement decree or its enforcement rules, and we describe them only as the guides cite them. We have not verified the mild cognitive impairment numbering discrepancy with MFDS; we report the two tables as they read and draw no conclusion about which is in error.

What we refuse to claim: we do not say the guides impose a requirement, because they say in terms that they have no legal effect. We do not state any numeric MCID for any indication, because the documents state none and leave a literal blank. We do not say the guides regulate artificial intelligence, because the only occurrences of the term are the issuing division's name. We give no count of affected trials, sponsors or products, because none of the documents gives one.

On language: every Korean passage above is reproduced in Hangul from the PDF and each is followed by an English rendering explicitly labelled as our translation. Where the source sets a phrase with wide inter-character spacing for emphasis, we normalise that spacing to single spaces, which changes no character and no word.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• Korea's MFDS on generative AI medical devices
• Australia's TGA on AI software as a medical device
• Asia Pacific AI regulation tracker
• AI healthcare compliance tracker
• AI Regulation News hub