Privacy, KVKK and Data Security: Protecting Pregnant Data

Gains:

• Ability to distinguish the data that makes the pregnant woman directly or indirectly identifiable and de-identify it before giving it to artificial intelligence
• Ability to apply data minimization and acquire the habit of using only institution-approved tools within the framework of KVKK
• 'Can this person be found from this text?' Ability to manage the risk of re-identification by applying the test

Midwifery is a profession that touches people's most private information: how a pregnancy occurs, past pregnancies, miscarriages, losses, relationships, mental state, sexual health. When this information falls into the wrong hands, it causes irreversible harm — a domestic crisis, a social stigma, a loss of trust. Therefore, confidentiality is not an "additional rule" in midwifery, but the foundation of the profession. Once you introduce AI into your business, a new frontier of privacy opens up: anything you type into a chat assistant could be out of your control. This unit teaches you end-to-end how to protect pregnant data when using AI — especially within the framework of KVKK (Personal Data Protection Law; the law regulating the processing of personal data in Türkiye).

The basic principle is in one sentence: Any information that makes a pregnant woman recognizable should never enter an artificial intelligence tool. Learning this rule takes precedence over all other skills in the module; Because once leaked data cannot be retrieved.

Attention: The text you type into an AI tool goes to the servers of that service; can be stored, processed, and in some cases used for model training. Saying "I just wanted a summary" does not return the data. The only safe way is to never type the credential in the first place.

What is personal data, what information makes it identifiable?

Personal data is any information that makes a person directly or indirectly identifiable. In midwifery, these are especially in the special category (such as health data, which receives the highest protection in the law). Consider the information that provides recognition in two groups:

Direct identifiers: Name-surname, TR ID number, file/protocol number, telephone, address, e-mail, date of birth, social media account, photo.

Indirect descriptors (combined to make it recognizable): "The only triplet pregnancy in that neighborhood", "the only 44-year-old in the institution with her first pregnancy", a rare disease + a small settlement, workplace + gestational age. A few details that seem innocent on their own can point to a person when brought together. This is called reidentification risk.

De-identification is the process of removing these identifiers from the text. A well-done de-identification clears both direct and indirect identifiers.

Step by step: De-identification before giving data to AI

1. Delete direct identifiers: Name, ID, phone number, file number, address, date — remove them all. Replace it with generalizations such as "Pregnant A", "around age 38", "second trimester".
2. Blur indirect clues: Turn exact age into range (e.g. "early 40s"), precise location into region, rare combinations into general.
3. Discard unnecessary detail: Give the minimum information needed for the AI ​​to do its job (data minimization). A pregnant woman's history is not required to write a breastfeeding card.
4. Check institution policy: Some institutions define approved/institutional AI tools; may prohibit entering patient data into personal/free tools. Follow your institution's rule.
5. Final reading: Before sending the prompt "can this person be found from this text?" Read it once more.

The most difficult of these five steps is the second—blurring indirect clues—because deleting direct identifiers is easy and comes to mind first, but most of the recognition actually comes from the combination of indirect details. The combination of a profession, an age, a place of residence, and a rare condition can reduce a person to a single person, even if no name is mentioned. Therefore, de-identification should not be thought of as "erasing names" but as "ensuring that no one reading this text can guess the person." This mental framework radically reduces the risk of re-identification.

Hint: A good test: If you show the de-identified text to a colleague who knows that pregnant woman, would she say "this is so-and-so"? If he can say that, it means you haven't disidentified him enough.

three mini cases

Case 1 — Indirect recognition: When asking the AI for a history summary, a midwife omitted the ID but left the phrase “the only quadruplet pregnancy in our town.” This single sentence directly points to the person in that small place. His colleague notices and warns; The midwife generalizes the expression as "multiple pregnancy". Lesson: indirect identifiers are just as dangerous as direct identifiers.

Case 2 — Data minimization: A second midwife habitually pastes the pregnant woman's entire file into the AI ​​instead of just printing out a "pregnancy nutrition" brochure. It stands: no personal data is required for the brochure. He takes the brochure without handing over the file, just writing the subject. The most secure data is data that is never shared.

Case 3 — Wrong tool: In the third example, a midwife is about to type patient information into a random application on her personal phone when she remembers the institutional policy: patient data can only be processed de-identified and using institution-approved tools. Not using the personal app. A data leak often arises from such a moment of "convenience" and not from a technical error.

Copiable templates

Role: You are the de-identification control assistant.Task: Mark each expression in the text below that could DIRECTLY or INDIRECTLY identify a person:- Direct: name, ID, telephone, address, file number, exact date, photo.- Indirect: rare combination, small place + special situation, exact age, workplace, unique/unique quality.Suggest a safer GENERALIZATION for each.Text: [text]

Role: You are a data minimization consultant.Task: I want to do this job: [job]. What is the minimum information I MUST give YOU in order to do this? Mark unnecessary patient information as "[DO NOT PROVIDE]".

Role: You are the secure summary assistant.Task: Summarize the DIDIDENTIFIED story below. If you see identification information in the text, do not summarize; instead write "I detected credentials, please clear and resend."Story: [story]

Institutional policy reminder checklist (ask yourself):- Is this tool approved by my institution?- Is there any personally identifiable information in the data I entered?- Is this information really needed for this job?- Can someone I know find the person if they read the text?- Will I add identification information when saving the output?

Weak prompt / Strong prompt

Weak: "Ayşe Y., 39, third pregnancy, in the following hospital, file number 4412; summarize her history."

This prompt is full of direct identifiers; Private data was leaked as soon as it was sent.

Strong:

Role: secure summary assistant. Summarize the following anonymous history. "Pregnant A, late 30s, third pregnancy, [related medical details anonymous]." If you see identifying information in the text, do not summarize, warn.

Difference: There are no direct/indirect identifiers in the strong version; there's just the identityless context needed to do the job. Security starts before the prompt is written.

What to share and what not to share

Information type

Does it fall into AI?

alternative

Name, TR ID, telephone, address, file number

never

"Pregnant A"

Exact age / exact date

no

Age range/trimester

Rare + small place combination

no

generalize

General medical context (anonymous)

Yes, if necessary

Apply data minimization

Photo/image

No (face/sign)

not necessary

Vehicle not approved by the institution

No patient data

Approved vehicle

Common mistakes

• Just deleting the name and leaving indirect clues: The risk of re-identification remains.
• Pasting the entire file for "convenience": Violates the principle of data minimization.
• Entering patient data into personal/unapproved devices: It may be against institutional policy and KVKK.
• Sharing a photo/screenshot: The face, name and file number in the image can remain confidential.
• Adding an ID while saving the output: Adding an ID to the output that comes without an ID brings the risk back.
• Assuming "it gets deleted anyway": You don't know what the data is; The only guarantee is not to send anything at all.

In summary

When using artificial intelligence, privacy is a prerequisite, not a technicality. Any information that makes a pregnant woman directly (name, ID, phone) or indirectly (rare combination + small location) recognizable should not enter an AI tool. Practice data minimization — provide the minimum information needed to do the job — and use only tools approved by your organization. Without sending the prompt "can this person be found from this text?" do the test. Once leaked, data cannot be retrieved; Therefore, the security starts before the prompt is written.

Application task

Take a real pregnancy history (on paper) and purge the text step by step using the deidentification template: first direct, then indirect identifiers. Show the cleared text to a colleague and ask "who could this be?" ask; If you can recognize it, generalize it further. Also find your own institution's AI/data policy, summarize it in one sentence, and note which tools are approved.

checklist

• [ ] I removed all direct identifiers (name, ID, phone, address, file number).
• [ ] I generalized indirect descriptors (rare combination, small location, exact age).
• [ ] I have provided only the minimum information needed to do the job.
• [ ] I used only a tool approved by my institution.
• [ ] "Can this person be found from this text?" I did the test.
• [ ] I did not add any credentials to the output afterwards.