AI in medical devices · Medium · 13 min lesson
AI risk management with ISO 14971 and ISO/IEC 23894: hazards unique to AI
ISO 14971:2019 clauses 4 to 10, ISO/IEC 23894:2023, AAMI TIR34971:2023 and EU AI Act Article 9
What you will learn
- Apply the ISO 14971:2019 process, clauses 4 to 10, to a device that contains a machine learning model.
- Identify hazards unique to AI, such as data, distribution shift, opacity, overtrust and model updates, and write them as hazard, sequence and harm.
- Choose risk controls in the order of clause 7.1, verify them under clause 7.2, and connect field signals to the risk file under clause 10.
Key takeaways
- ISO 14971 is technology neutral: the same plan, analysis, evaluation, control, review and monitoring apply to AI-enabled devices.
- AI hazards often start with data, distribution shift, opacity, overtrust or retraining, so they must be added to the clause 5.4 hazard list.
- Intended use sets the boundary of the evidence; a vague intended use puts untested conditions inside the claim.
- Acceptability criteria use severity and probability of harm, not an accuracy target; controls follow design, protection, information (clause 7.1).
- ISO/IEC 23894 is guidance, AAMI TIR34971 bridges to ISO 14971, and Article 9 of the AI Act requires a lifecycle risk system.
Free practice question
Corvale's risk file lists hazards for PulmoAssist: crash, wrong patient record and slow response. The model was trained on images from fixed scanners only, and clinics now ask to connect portable units.
What should the team do first?
Add a data and distribution-shift hazard, with sequence and harm, to the clause 5.4 analysis and review the intended use.
Correct. ISO 14971 clause 5.4 requires all hazards to be identified, and unfamiliar inputs are a hazard source for a model.
Leave the file unchanged because the software has no defects.
Not quite. Clause 5.4 covers any source of harm, so hazards without a software defect still belong in the analysis.
Rely on the overall accuracy figure as proof of safety.
Not quite. Accuracy is a performance result and does not replace identifying hazards under ISO 14971 clause 5.4.
Ask clinics to avoid portable units in a verbal briefing.
Not quite. A verbal request is not an identified hazard or a verified control under ISO 14971 clauses 5 and 7.
Key terms
- hazard: A potential source of harm (ISO 14971:2019 clause 3).
- hazardous situation: A circumstance in which people, property or the environment are exposed to one or more hazards.
- residual risk: The risk remaining after risk control measures have been implemented.
- distribution shift: A difference between the data a model was trained on and the data it meets in use, which can degrade performance.
- out-of-distribution detection: A check that flags inputs unlike the training data so the model output is not trusted blindly.
- automation bias: The tendency of users to rely too much on an automated output (named in AI Act Article 14(4)(b)).
- risk control: A measure that reduces risk, chosen in the order design, protective measure, information for safety (clause 7.1).
Official sources
- ISO 14971:2019 Medical devices - Application of risk management to medical devices
- ISO/IEC 23894:2023 Information technology - Artificial intelligence - Guidance on risk management (IEC webstore)
- AAMI TIR34971:2023 Application of ISO 14971 to machine learning in artificial intelligence - Guide (announcement)
- Article 9: Risk management system, AI Act explorer
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal
Go further
The full lesson has narrated slides, more scenario questions with explained answers and a simulation.
Open the AcademyWatch on YouTubeIllustrative example. Educational only, not legal advice.