Retinal Screening Assistant

Support ophthalmology screening services by detecting referable diabetic retinopathy in colour fundus images and prioritising cases for specialist review. Software as a medical device deployed in screening clinics. The system reads fundus image pairs, produces a referral recommendation and a confidence value, and returns both to the reviewing clinician. The clinician retains the decision. Deployed configuration: MV-002 2.1.0.

Drawn from AIS-001 Retinal Screening Assistant.

PM-001 Periodic retraining on extended screening archive Retrain the referable classifier on a newer Dataset Version of the screening archive without changing architecture, inputs, outputs, or intended purpose. PM-002 Site expansion retraining Retrain on a Dataset Version that adds two northern screening sites, in response to the observed drift signal. PM-003 Confidence threshold adjustment Adjust the referral confidence threshold within a declared range in response to service capacity.

Drawn from PM-001 Periodic retraining on extended screening archive, PM-002 Site expansion retraining, PM-003 Confidence threshold adjustment. Open controlled evolution

PM-001 Periodic retraining on extended screening archive Training data drawn only from released Dataset Versions of the registered screening archive. The evaluation set DV-002 is held out and never enters training. Site, device and demographic composition are recorded for each version. PM-002 Site expansion retraining Adds two northern screening sites to the archive under the same consent and provenance rules. A northern-site subset is carved out of the evaluation set and excluded from training.

Drawn from PM-001 Periodic retraining on extended screening archive, PM-002 Site expansion retraining. Open controlled evolution

PM-001 Periodic retraining on extended screening archive Retraining is triggered by the annual cycle or by a breached monitoring threshold. Frozen training protocol TP-002 with identical hyperparameters, identical preprocessing and a fixed random seed. Frozen training protocol TP-002. Identical hyperparameters. Evaluation on the independent evaluation set DV-002. PM-002 Site expansion retraining Frozen training protocol TP-002. Site-stratified sampling so no single site exceeds 30 percent of the training records. Frozen training protocol TP-002. Site-stratified evaluation on DV-002 plus a northern-site subset. PM-003 Confidence threshold adjustment Threshold sweep on DV-002 with fixed operating-point reporting.

Drawn from PM-001 Periodic retraining on extended screening archive, PM-002 Site expansion retraining, PM-003 Confidence threshold adjustment. Open controlled evolution

PM-001 Periodic retraining on extended screening archive Sensitivity and specificity with 95 percent confidence intervals on DV-002, reported overall and stratified by site, device and age band, against the acceptance criteria below. Sensitivity not below 0.92 and specificity not below 0.86 on DV-002, with no subgroup falling more than 0.04 below the overall sensitivity. PM-002 Site expansion retraining Site-stratified sensitivity and specificity on DV-002 plus the northern-site subset, against the acceptance criteria below. Northern-site sensitivity not below 0.90 and overall specificity not below 0.86. PM-003 Confidence threshold adjustment Sensitivity not below 0.92 at any permitted threshold.

Drawn from PM-001 Periodic retraining on extended screening archive, PM-002 Site expansion retraining, PM-003 Confidence threshold adjustment. Open controlled evolution

PM-001 Periodic retraining on extended screening archive Release is authorised by a Change authorisation review with an approver independent of the author. Users are notified through the release note before deployment. The previous AI Model Version stays deployable for 12 months.

Drawn from PM-001 Periodic retraining on extended screening archive. Open controlled evolution

PM-001 Periodic retraining on extended screening archive No change to intended purpose or clinical workflow. Risk profile unchanged. Labelling unchanged. Recovers sensitivity lost to population drift and reduces the subgroup performance spread. A retrained model can shift the operating point and change the referral rate. Fixed acceptance criteria, site-stratified reporting, and a monitored referral rate for the first 90 days after release. Second retraining under this plan. Cumulative sensitivity change since authorisation is within 0.01 and the intended purpose is unchanged. PM-002 Site expansion retraining No change to intended purpose. Population coverage widened. Requires an updated data governance statement. Extends validated performance to the northern screening population. Population coverage widens without an updated data governance statement.

Drawn from PM-001 Periodic retraining on extended screening archive, PM-002 Site expansion retraining. Open controlled evolution

AIS-001 evolves through periodic retraining and bounded threshold adjustment. Architecture, intended purpose, inputs, outputs, and clinical workflow are outside the declared boundaries and require a new authorisation. Permitted: retraining on released Dataset Versions of the registered datasets; threshold adjustment within 0.30 to 0.55. Not permitted: architecture change, new input modality, new clinical claim, autonomy change. Constraints . No change to the intended purpose or the clinical claim . No change to model architecture or input modality . No change to the level of human oversight . Evaluation always on an independent Dataset Version not used in training Regulatory reassessment triggers . Any change outside the declared boundary . Sustained subgroup sensitivity below 0.90 across two monitoring periods . A new user group or a new market . A serious incident attributable to the model output

Drawn from AIS-001 Controlled evolution boundary. Open controlled evolution

Monthly subgroup sensitivity and specificity by site. Monthly image-statistics drift metric. Quarterly review of override rate. Previous approved AI Model Version remains deployable for 12 months. Rollback is authorised by the AI System Owner and recorded as a Change authorisation review.

Drawn from AIS-001 Controlled evolution boundary. Open controlled evolution

EU AI Act Article 9, Risk management system: AI Risk records maintain hazard, control, and residual acceptance state across the life cycle. EU AI Act Article 10, Data and data governance: Dataset Versions are immutable and their suitability for each AI Model Version is recorded. ISO/IEC 42001 8.3, AI system impact assessment: Impact assessment is recorded per Planned Modification. EN 18286 6.2, Data quality: Dataset quality assessment is required before release. ISO 14971 7, Risk control: Risk controls are verified before system approval.