Skip to content

Latest commit

 

History

17 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 

Repository files navigation

AI Governance Framework for a High-Risk Clinical Decision Support System

EU AI Act · EU MDR · GDPR

A practical governance framework for a clinical AI triage system, mapping regulatory obligations from three overlapping EU frameworks into a single, integrated compliance architecture.


What This Is

This is a self-initiated project I built as an MBA student to understand what responsible AI deployment in healthcare actually requires — not at a theoretical level, but operationally.

The system in question is a Clinical Triage Decision Support System (AI-CTDSS): an AI that analyses patient vitals and symptoms at emergency department intake and recommends a triage priority score to the attending nurse.

The framework works through:

  • How this system is classified under EU law (and why it is simultaneously regulated by three different regulations)
  • What each of those regulations actually requires — and how those requirements interact with each other
  • How specific design decisions (like how the UI presents the AI's recommendation) can satisfy multiple legal obligations at once
  • What happens when something goes wrong — incident classification, reporting timelines, and who is accountable

Regulatory Scope

Regulation Classification Key Trigger
EU AI Act (2024/1689) High-Risk AI System Art. 6(1) — safety component in an MDR-regulated medical device
EU MDR (2017/745) Class IIa Medical Device Annex VIII Rule 11 — software influencing decisions with diagnostic intent
GDPR (2016/679) Special Category Data Processing Art. 9 — health data; Art. 22 — automated decision-making

All three classifications apply simultaneously. This framework resolves the obligations from all three in one integrated architecture rather than treating them in isolation.


Framework Structure

Section Title
0 Executive Summary
1 System Definition and Scope
2 Risk Classification Analysis
3 GDPR Compliance Architecture
4 Human Oversight Design
5 Technical Documentation Requirements
6 Audit Trail and Logging
7 Transparency and User Information
8 Incident Response and Post-Market Monitoring
9 Governance Ownership Map
10 Limitations and Implementation Roadmap

Three Things Worth Reading First

If you don't want to read all 50+ pages, these are the sections with the most interesting design decisions:

Section 4 — Human Oversight

The system uses a "Delayed Reveal" mechanism — the nurse must enter their own initial clinical impression before the AI's recommendation is displayed. This single design decision simultaneously satisfies:

  • GDPR Article 22 — the triage decision is never "based solely on automated processing"
  • EU AI Act Article 14 — effective human oversight is built into the workflow, not bolted on

Two legal requirements, one design solution. It also protects against automation bias — the documented tendency for clinicians to simply agree with whatever the algorithm outputs.

Section 3.7 — Algorithmic Fairness

SpO2 sensors (the small finger clip measuring blood oxygen) are clinically documented to over-read in patients with darker skin pigmentation — producing falsely elevated readings and masking hypoxia. The framework treats this as a regulatory compliance risk under EU AI Act Article 10 (data governance), not just an ethical concern.

A second bias risk: atypical MI presentations. Women and elderly patients frequently present with heart attacks using non-classic symptoms (fatigue, nausea, indigestion rather than chest pain). An AI trained primarily on male demographic data will systematically under-triage these patients.

Section 8 — Incident Response

Maps the difference between a "Serious Incident" under MDR Article 87 and a "Serious Incident involving a High-Risk AI System" under EU AI Act Article 73, and builds a unified incident classification matrix with the correct reporting timeline for each scenario (15-day maximum, 10-day escalation, and a strict 72-hour internal operational SLA).


Compliance Timeline

Full EU AI Act enforcement for AI systems classified as medical devices: August 2028 (per Digital Omnibus Reg. 2026/1744).


Limitations

This is a speculative project built by an MBA student, not a practising regulatory lawyer or clinical engineer. It is based on primary sources — the regulations themselves, EU Commission guidance, and Notified Body standards — but should not be treated as legal advice or used as an actual regulatory submission.


About Me

MBA student interested in AI governance, healthcare regulation, and the intersection of technology and policy.

Currently exploring roles in HealthTech AI governance, regulatory strategy, and GRC.

LinkedIn

About

An operational AI governance framework for a high-risk clinical triage system. Unifies EU AI Act, GDPR, and MDR compliance into a single architecture. Self-directed MBA portfolio project.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages