Support ophthalmology screening services by detecting referable diabetic retinopathy in colour fundus images and prioritising cases for specialist review.
100, less 12 for each open blocker and 4 for each warning. It measures the record, not the performance of the AI System.
The system is in operation while its record stands below the floor of 70. That contradiction is raised as blocker STATE-01.
A control is not effective until its effectiveness is verified and the verification is linked to evidence.
Where the risks of this AI System stand. A risk moves one stage at a time, and every move is written to the audit trail.
Risks arising from the AI System, its use cases, its models or its data. Open risks are listed first; closed risks are kept below.
Draft progression use case is used clinically before authorisation. Harm: Unvalidated recall decisions.
ControlsNo control recorded.
Model under-performs on populations under-represented in training data. Harm: Delayed referral and avoidable vision loss.
ControlsSubgroup performance reported per release.; Monthly monitoring by site with a defined action threshold.; Clinician retains the referral decision.
Reviewer accepts the recommendation without independent assessment. Harm: Incorrect referral decision.
ControlsConfidence value and image evidence shown with every recommendation.; Training record required before access.
New capture hardware shifts image characteristics away from training data. Harm: Reduced sensitivity without detection.
ControlsDrift metric on image statistics, monthly.
Retained training images allow re-identification of a screening participant. Harm: Loss of confidentiality.
ControlsPseudonymisation at ingest.; Access restricted to the data platform role.
You are acting as AI System Owner; the audit trail will say so.