Retinal Screening Assistant
Support ophthalmology screening services by detecting referable diabetic retinopathy in colour fundus images and prioritising cases for specialist review.
- Status
- Approved for operation
- State
- Deployed
- Owner
- Cirocco Jones
- Readiness
- 0
- Blockers
- 12
Identity
What the system is, where it is placed, and how often it is reviewed.
Lifecycle
Where the system stands, and the act that moves it on. Every move is written to the audit trail under the acting persona.
The system exists as a record. Design work has not started.
Design and development are under way. Nothing is in use outside the project.
Released for use. The post-market duties apply from this point.
In use, but with a limitation placed on it after a signal or a finding.
Use is stopped while a matter is settled.
Withdrawn from use. Records are held for the retention period.
Linked records
Navigate from the AI System to every governed record.
Chronology
Every dated act on this AI System and its records, in order.
- 2024-03-04SystemAIS-001AI System registered
- 2024-03-04UDI(01)05412345000041UDI-DI not registeredThreshold recalibration for the northern sites. Performance changes, so registration is required before release.
- 2024-09-12DataDV-001Dataset version 1.0 frozen84,120 records
- 2024-09-30DataDV-002Dataset version 1.0 frozen12,480 records
- 2024-11-08ModelMV-001Model version 1.0.0 frozenSensitivity 0.92, specificity 0.88 on DV-002.
- 2025-01-22ModelMV-004Model version 1.3.0 frozenAgreement with grader consensus 0.96.
- 2025-02-04ReviewREV-003Model approval: ApprovedApproved. Performance and suitability records are complete.
- 2025-02-11ReviewREV-004System approval for operation: Approved with conditionsApproved with conditions. Subgroup monitoring is required from first deployment.
- 2025-02-20UDI(01)05412345000027UDI-DI supersededFirst market release of the screening classifier.
- 2025-10-28DataDV-003Dataset version 2.0 frozen103,660 records
- 2025-12-15ModelMV-002Model version 2.1.0 frozenSensitivity 0.94, specificity 0.89 on DV-002.
- 2026-01-19ReviewAIS-001Periodic review completed
- 2026-01-19ReviewREV-010Change authorisation: ApprovedApproved. Change stays inside the declared modification boundaries.
- 2026-01-19ReviewREV-011Periodic review: Approved with conditionsContinued operation approved. Two conditions remain open.
- 2026-01-19ChangePM-001Planned modification implementedRetrain the referable classifier on a newer Dataset Version of the screening archive without changing architecture, inputs, outputs, or intended purpose.
- 2026-01-26Release9f13c7aRelease v4.2.0Retrain on DV-003, recalibrate operating point
- 2026-02-04UDI(01)05412345000034UDI-DI registeredRetrained on an extended dataset version. Sensitivity and specificity both changed, so a new UDI-DI was assigned.
- 2026-05-11Release1a77e2bTag v4.2.1Fix DICOM loader crash on 12-bit input
- 2026-06-24DataDV-004Dataset version 3.0 frozen118,904 records
- 2026-07-01SignalPerformance: Sensitivity below threshold, northern sitesSensitivity 0.88 against a floor of 0.92 across four northern sites over the July window.
- 2026-07-01SignalDrift: Image statistics drift above tolerancePopulation drift metric 0.21 against a tolerance of 0.15, coincident with a camera fleet replacement at two sites.
- 2026-07-02ModelMV-003Model version 2.2.0-candidate frozenSensitivity 0.95, specificity 0.87 on DV-002. Not yet reviewed.
- 2026-07-06ReviewREV-012Model approval: In preparationApproval of the retrained referable classifier candidate MV-003.
- 2026-07-14ChangePM-003Planned modification draftAdjust the referral confidence threshold within a declared range in response to service capacity.
- 2026-07-18SignalComplaint: Clinician reports repeated ungradable classificationSite 11 reports the quality gate rejecting images a grader considers gradable. Under triage.
- 2026-07-20ChangePM-002Planned modification assessedRetrain on a Dataset Version that adds two northern screening sites, in response to the observed drift signal.
- 2026-07-22Release4de90c1Commit mainSite-level threshold recalibration, northern cohort
- 2026-07-27Release77b0aa3Commit #412Add second input modality: OCT volumes
Governance gates
A gate is held until it is proven. It passes when the records it depends on exist, an approved AI Review of the matching type is recorded, and no blocker in its scope is open.
| Gate | Purpose | Responsible | Outcome | What is missing |
|---|---|---|---|---|
| Development authorisation | Authorises development against a stated intended purpose. | Independent Reviewer | Held |
|
| Model approval | Approves an immutable AI Model Version for use. | Independent Reviewer | Held |
|
| System approval for operation | Approves the complete AI System for operation in a market. | Independent Reviewer | Held |
|
| Change authorisation | Authorises a change inside the Controlled Evolution boundaries. | Quality and Regulatory | Held |
|
| Resumption | Resumes operation after restriction or suspension. | Independent Reviewer | Held |
|
Posture summary
A written summary of the computed findings and the recommended order of work. The findings themselves are deterministic. The narrative never states that the AI System is compliant.
Demonstration fill
Scout drafts AI Use Cases, AI Models and versions, Datasets and versions, the dataset to model links, AI Risks, AI Reviews and Planned Modifications, all consistent with what this AI System already states. Records are added as drafts and every one carries an audit entry saying it was generated.
Open findings
12 blockers, 9 warnings.
- BlockerTRACE-01MV-004 has no training suitability record
No Dataset Model record states which Dataset Version was used for training or why it is suitable.
- Affects
- MV-004
- Effect
- Blocks continued operation and PCCP generation.
- BlockerDATA-01DV-004 has no quality assessment
A Dataset Version cannot be used for training or evaluation until its quality assessment is recorded.
- Affects
- DV-004
- Effect
- Blocks model approval and change authorisation for records that use it.
- BlockerRISK-01RSK-003 residual risk is not accepted
New capture hardware shifts image characteristics away from training data. No accepted residual position is recorded.
- Affects
- RSK-003
- Effect
- Blocks continued operation and change authorisation.
- BlockerRISK-01RSK-005 residual risk is not accepted
Draft progression use case is used clinically before authorisation. No accepted residual position is recorded.
- Affects
- RSK-005
- Effect
- Blocks continued operation and change authorisation.
- BlockerREV-01REV-011 condition is overdue
Close the drift risk RSK-003 with a verified control. Owner Louis Wu, due 2026-06-30.
- Affects
- REV-011
- Effect
- Blocks change authorisation until closed.
- BlockerMON-01Monitoring threshold breached: Sensitivity, northern sites
Observed 0.88 against threshold "Not below 0.92" on 2026-07-01.
- Affects
- AIS-001
- Effect
- Requires an AI Risk assessment and a change or restriction decision.
- BlockerMON-01Monitoring threshold breached: Image statistics drift
Observed 0.21 against threshold "Not above 0.15" on 2026-07-01.
- Affects
- AIS-001
- Effect
- Requires an AI Risk assessment and a change or restriction decision.
- BlockerCE-01PM-003 has no impact assessment
A Planned Modification cannot be authorised or included in a PCCP without one.
- Affects
- PM-003
- Effect
- Excludes the modification from the generated PCCP.
- BlockerREPRO-01MV-004 cannot be reproduced
No commit, environment, hyperparameters or dataset versions are recorded, so the result cannot be re-derived.
- Affects
- MV-004
- Effect
- Blocks continued operation. The deployed result is not evidenced.
- BlockerSIG-01Performance signal not acknowledged: Sensitivity below threshold, northern sites
Sensitivity 0.88 against a floor of 0.92 across four northern sites over the July window. Observed 2026-07-01.
- Affects
- AIS-001
- Effect
- Restriction is proposed. Blocks continued operation until answered.
- BlockerSIG-01Drift signal not acknowledged: Image statistics drift above tolerance
Population drift metric 0.21 against a tolerance of 0.15, coincident with a camera fleet replacement at two sites. Observed 2026-07-01.
- Affects
- AIS-001
- Effect
- Requires triage and a recorded response.
- BlockerUDI-02UDI-DI (01)05412345000041 is not registered
Threshold recalibration for the northern sites. Performance changes, so registration is required before release. A change that affects performance or safety carries a new UDI-DI, and that UDI-DI is registered before the version is placed on the market.
- Affects
- (01)05412345000041
- Effect
- Blocks release of this version.
- WarningSTRUCT-02UC-003 references no AI Model
Use case "Longitudinal progression flag" cannot be traced to a model, so its performance cannot be evidenced.
- Affects
- UC-003
- Effect
- Blocks approval of this use case.
- WarningTRACE-01MV-003 has no training suitability record
No Dataset Model record states which Dataset Version was used for training or why it is suitable.
- Affects
- MV-003
- Effect
- Blocks model approval.
- WarningTRACE-02MV-001 has no testing suitability record
Performance cannot be traced to a specific, immutable Dataset Version.
- Affects
- MV-001
- Effect
- Blocks model approval.
- WarningTRACE-02MV-003 has no testing suitability record
Performance cannot be traced to a specific, immutable Dataset Version.
- Affects
- MV-003
- Effect
- Blocks model approval.
- WarningREV-01REV-011 condition is open
Complete the quality assessment for DV-004 before use in training. Owner Louis Wu, due 2026-08-31.
- Affects
- REV-011
- Effect
- Must be closed before the next periodic review.
- WarningEVID-01DHF-2291-RMR-003 is stale
Approved 2025-02-06. Review interval is 12 months.
- Affects
- DHF-2291-RMR-003
- Effect
- Weakens the mapping that depends on it.
- WarningREPRO-01MV-003 cannot be reproduced
No commit, environment, hyperparameters or dataset versions are recorded, so the result cannot be re-derived.
- Affects
- MV-003
- Effect
- Blocks model approval.
- WarningREPO-01Repository event main has no governance response
Site-level threshold recalibration, northern cohort, Performance changes. Freeze a new AI Model Version, open a change authorisation review, and assign a new UDI-DI.
- Affects
- 4de90c1
- Effect
- Change is happening in the repository ahead of the record set.
- WarningREPO-01Repository event #412 has no governance response
Add second input modality: OCT volumes, Outside the declared Controlled Evolution boundary. Requires a new authorisation and a new Basic UDI-DI.
- Affects
- 77b0aa3
- Effect
- Change is happening in the repository ahead of the record set.
Monitoring
Signals observed on the deployed configuration.
| Metric | Observed | Threshold | State | Observed on |
|---|---|---|---|---|
| Sensitivity, all sites | 0.93 | Not below 0.92 | Within threshold | 2026-07-01 |
| Sensitivity, northern sites | 0.88 | Not below 0.92 | Breached | 2026-07-01 |
| Image statistics drift | 0.21 | Not above 0.15 | Breached | 2026-07-01 |
| Clinician override rate | 6.1% | Not above 10% | Within threshold | 2026-07-01 |
Framework mappings
A mapping records where a requirement is addressed. It does not assert compliance.
| Framework | Clause | How it is addressed | Evidence |
|---|---|---|---|
| EU AI Act | Article 9, Risk management system | AI Risk records maintain hazard, control, and residual acceptance state across the life cycle. | DHF-2291-RMR-003 |
| EU AI Act | Article 10, Data and data governance | Dataset Versions are immutable and their suitability for each AI Model Version is recorded. | DHF-2291-DGS-002 |
| ISO/IEC 42001 | 8.3, AI system impact assessment | Impact assessment is recorded per Planned Modification. | Not linked |
| EN 18286 | 6.2, Data quality | Dataset quality assessment is required before release. | DHF-2291-AP-001 |
| ISO 14971 | 7, Risk control | Risk controls are verified before system approval. | DHF-2291-RMR-003 |
Audit trail
- 2024-03-04Cirocco JonesRegistered AI System
- 2025-02-11Lilith IyapoApproved for operation, REV-004
- 2026-01-19Lilith IyapoCompleted periodic review, REV-011
- 2026-07-14Cirocco JonesOpened Controlled Evolution revision 3