AI in medical devices · Expert · 13 min lesson

Post-market monitoring of AI: performance drift, data collection and triggers

MDR Articles 83 to 88 and Annex III, EU AI Act Article 72, ISO 14971 clause 10, FDA PCCP guidance

What you will learn

Explain why a complaint driven system cannot see silent performance drift, and describe the proactive monitoring that MDR Articles 83 and 84 and AI Act Article 72 expect.
Define data sources, a validated baseline, sampling and ground truth checks that let an AI device show whether it still performs as claimed.
Write graded triggers with defined actions and show in a record how a monitoring signal reaches the risk file, CAPA and clinical evaluation.

Key takeaways

Silent drift leaves no complaint, so MDR Articles 83 and 84 and AI Act Article 72 call for proactive, documented monitoring.
The monitoring plan fixes a baseline from validation, a sample design with a reason for its size, ground truth checks and the subgroups to watch.
Input signals such as data quality and output rates usually move before outcome metrics, so they earn their own watch level.
Graded triggers need owners, deadlines and a stated statistical method, and baselines change only through controlled design change.
Each signal must reach the risk file under ISO 14971 clause 10, corrective action under ISO 13485 and the clinical evaluation under MDR Article 61.

Free practice question

Halden monitors Halden Rhythm only through the complaints inbox. In a year it receives four complaints, all about login problems. A new customer reports privately that the software seems quieter than before, but nobody records the remark.

What is the correct assessment of this surveillance system?

Adequate, because no performance complaint exists.

Not quite. An absence of complaints is not evidence of performance, and MDR Article 83 requires a proactive and systematic system instead.

Inadequate, because it cannot detect silent drift and has no proactive data collection.

Correct. MDR Article 83 and AI Act Article 72(1) require proactive monitoring proportionate to the AI risks, which a complaints inbox cannot give.

Adequate if the inbox is checked daily.

Not quite. Frequency of checking does not make a reactive channel proactive, so Article 83 of the MDR is still not met.

Inadequate only because login complaints were not fixed.

Not quite. Login problems are a separate matter; the defect against Annex III of the MDR is the missing performance monitoring.

Key terms

performance drift: A change over time in how well a model performs, caused by changes in data, population, practice or the environment.
data drift: A change in the statistical features of the inputs a model receives, compared with those seen in development and validation.
concept drift: A change in the relationship between input and the true outcome, so the same input now deserves a different answer.
baseline: The validated performance values against which monitoring results are compared, controlled through design change control.
ground truth: The best available determination of the true outcome, often by expert review blind to the model output.
trigger: A predefined threshold in a monitoring plan that obliges a named person to act within a stated time.
post market surveillance plan: The plan required by MDR Article 84 and Annex III, describing how post market data is collected and assessed.

Official sources

Go further

The full lesson has narrated slides, more scenario questions with explained answers and a simulation.

Open the Academy Watch on YouTube

Illustrative example. Educational only, not legal advice.