W7 · Regulation, data protection and a security incident
Workshop W7, teaching units 25–28: two intended purposes under the AI Act and the MDR, data flows for care and research under the GDPR, a threat model with tested controls, and a tabletop exercise on an incident.
Guiding question
When a research dashboard is to be used in care, receives an update and leaks case text, who has to decide what, on which legal basis, and by when?
Learning objectives
By the end of this workshop you can
- compare two intended purposes of the same software and derive roles and questions under the AI Act, the MDR and, where relevant, the IVDR, citing the primary source and marking what is a case assumption and what is open;
- draw separate data flows for care and for later research use and check purpose, legal basis under Article 6 and exception under Article 9 GDPR, processor contract, retention, rights of the people affected and the DPIA threshold;
- complete a threat model and propose for confidentiality, integrity and availability one control each, a test that shows it works and a recovery step;
- run an incident from containment to restart, deciding notification to the authority (Article 33 GDPR) and communication to the people affected (Article 34 GDPR) separately;
- write a short, honest message to affected patients with a reachable contact.
Preparation
- Required reading (about 20 minutes): B22 · Regulating research software and clinical AI – core reading and B23 · Data protection along the entire data flow – core reading.
- On the table during the exercise: B24 · Information security and recovery and B25 · Security risks of learning and generative systems; B21 · Ethics for the patient perspective and B20 · Pilot operation, impact and monitoring for the restart.
- Guests (proposal): contributions from data protection, medical device regulation and information security. The roles are suggestions; any of them can be played by course participants.
Teaching units
| Unit | Task | Minutes | Observable evidence |
|---|---|---|---|
| 25 | Compare two intended purposes and determine roles and questions under the EU AI Act, the MDR or, where relevant, the IVDR. Name one missing manufacturer document (starting situation). | 45 | A dated regulatory note with primary source, case assumption and open next step. |
| 26 | Design one data flow for care and one for later research use. Check legal basis, health data, processing on behalf, retention, rights of the people affected and the DPIA threshold (card 1). | 45 | A data flow with recipient, purpose, access roles, deletion point and a reasoned DPIA pre-check. |
| 27 | Complete a threat model. For data leaving the system, a manipulated model version and an outage, design one protection, one detection test and one recovery measure each (card 2). | 45 | A control matrix for confidentiality, integrity and availability and a fall-back option tested in the exercise. |
| 28 | Tabletop: synthetic case text is visible in a support log. Contain, secure evidence, alert roles, assess consequences and decide on the restart (card 3). | 45 | An incident log, a reasoned assessment of notification and of communication, the patient perspective and a release criterion. |
Legal status, November 2026
Materials
- Case sheets W7: starting situation, two intended purposes, data flow, role sheets, three case cards, threat model, empty control matrix, event log and legal anchors.
- One printed set per team; the course team keeps cards 1–3 until the start of units 26, 27 and 28.
Four perspectives in this workshop
All four are central here and are assessed separately in the project rubric.
- Ethics. Who is affected by a wrong alert, a silent update and a leak, and what do they need to know? The patient representative’s text is part of the result.
- Regulation. The change from purpose A to purpose B changes the regulatory question, not the code.
- Data protection. Care and research are different purposes with different legal bases; notification and communication are different decisions.
- Security. Each control comes with a test that someone can watch: a refused access, a refused model version, a working fall-back.
Transition to W8
Your regulatory note, data flow and control matrix complete the decision sheet of your project. In W8 you defend the whole package.
Further reading
- B24 · Information security and recovery and B26 · Secondary use and responsible open science – recovery tests and research use without exporting raw data.
- CTTIR Courses: TRIPOD-AI, fairness auditing, reproducibility at scale – documentation that a regulator or reviewer can check.
- Methods in R: Shared reviewer checklist – questions a reviewer asks before trusting a model.
- Strategy in R: Ethical Frameworks for Strategic AI – competing ethical frameworks made explicit.
- Scientometrics in R: Ethics and Responsible Metrics – responsible use of indicators in another domain.
Sources
- Regulation (EU) 2024/1689 (AI Act): EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus on AI): EUR-Lex
- Regulation (EU) 2017/745 (MDR): EUR-Lex; Regulation (EU) 2017/746 (IVDR): EUR-Lex
- MDCG 2019-11 Rev. 1 (June 2025): PDF
- Regulation (EU) 2016/679 (GDPR): EUR-Lex
- Federal Data Protection Act (Bundesdatenschutzgesetz, BDSG), § 27: gesetze-im-internet.de
- Regulation (EU) 2025/327 (European Health Data Space): EUR-Lex
- World Health Organization. Ethics and governance of artificial intelligence for health: guidance on large multi-modal models. 2024. who.int
- Federal Office for Information Security (BSI). Generative KI-Modelle, version 2.0, January 2025 (German): PDF
- NIST AI 100-2e2025. Adversarial machine learning: a taxonomy and terminology of attacks and mitigations. doi:10.6028/NIST.AI.100-2e2025