AI in medical devices · Medium · 13 min lesson
Verification and validation of AI: performance metrics, test sets and acceptance criteria
ISO 13485:2016 clauses 7.3.6, 7.3.7 and 7.3.10, Regulation (EU) 2024/1689 Articles 9, 10 and 15, GMLP principles 3 to 8
What you will learn
- Choose performance metrics that match the intended use and risk, and read a confusion matrix, intervals and prevalence effects correctly.
- Specify an independent, locked and representative test set and size it with a stated statistical rationale.
- Write acceptance criteria before testing, recognise a weak record and answer the questions an auditor asks.
Key takeaways
- Verification checks the AI against written requirements; validation checks the intended use in realistic conditions (clauses 7.3.6, 7.3.7).
- Choose metrics from the risk file, never accuracy alone, and report confidence intervals and prevalence.
- The test set is independent of training, representative and locked before model decisions (GMLP 4, Article 10 paragraph 3).
- Acceptance criteria are approved before the run, written on interval bounds, with a rule for what happens on failure.
- Freeze and record data, code, model and threshold versions; keep signed and dated test logs for the technical file.
Free practice question
Corvane reports that its sleep-apnoea screening model reached 97 percent accuracy. The test set holds 970 normal recordings and 30 recordings with the condition. The model flags almost nothing. A reviewer asks whether the claim supports the intended use of finding patients with the condition.
What is the main weakness of the accuracy figure?
With only 3 percent real cases, a model that flags nothing scores high accuracy while missing nearly every patient.
Correct. Accuracy blends the classes, so GMLP principle 6 clinically meaningful goals require sensitivity and specificity for the intended use.
Accuracy is always wrong for classification.
Not quite. Accuracy is valid arithmetic but misleading under imbalance, which GMLP principle 6 tells teams to avoid relying on alone.
Ninety seven percent is too low to be useful.
Not quite. The number is not the issue; the metric cannot show whether patients with the condition are found, per Article 9 paragraph 8 metrics.
The set is too large to be representative.
Not quite. Size is not the defect here; the imbalance hides missed cases, which clause 7.3.6 verification must expose.
Key terms
- sensitivity: Share of real cases that the device correctly flags; true positives divided by all real cases.
- specificity: Share of normal cases that the device correctly clears; true negatives divided by all normal cases.
- positive predictive value: Share of flagged cases that are real; it changes with the prevalence of the condition.
- prevalence: How common the condition is in the population where the device is used.
- AUC: Area under the receiver operating characteristic curve; summarises how well scores rank real above normal cases.
- confidence interval: A range that expresses statistical uncertainty around an estimated metric.
- acceptance criterion: A pre-approved numeric rule that decides whether a test is passed.
- data leakage: Information from the test data reaching the training process, which inflates measured performance.
Official sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal
- Article 9: Risk management system, AI Act explorer
- Article 10: Data and data governance, AI Act explorer
- Good Machine Learning Practice for Medical Device Development: Guiding Principles (MHRA, with FDA and Health Canada)
- ISO 13485:2016 Medical devices - Quality management systems - Requirements for regulatory purposes
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.