Held to the standard of a clinical instrument.
Verityn is validated against retrospective, de-identified imaging datasets in controlled reader studies. Every model version is paired with its evaluation report. Performance figures on this site are illustrative; real performance data is available in our clinical evaluation and validation summary under agreement.
Illustrative figures notice
All performance figures on this page are illustrative placeholders. They show the types of metric a clinical audience expects in a formal validation report and the general magnitude of results seen in published second-reader AI studies. They are not claims about Verityn's measured performance in any released version. The clinical evaluation and validation summary is available to qualified clinical and procurement teams under a non-disclosure agreement. Contact our clinical team to request access.
The performance a clinical audience needs to evaluate.
A second-reader validation reports both standalone model performance and the more important question: how does a radiologist read differently with it? Both are shown here as illustrative examples of the kinds of figures that appear in a full validation report.
Nodule detection on chest radiograph, retrospective de-identified dataset. Illustrative placeholder only.
Same chest radiograph validation cohort, per-image specificity. Illustrative placeholder only.
Area under the ROC curve, nodule detection task. Higher is better; 1.0 is perfect. Illustrative placeholder only.
Reader performance with Verityn flags versus unaided. The assisted AUC is the primary outcome of interest in a second-reader study. Illustrative placeholder only.
Relative reduction in missed findings, assisted versus unaided read arms. Illustrative placeholder only.
Percentage-point improvement in sensitivity for sub-centimetre lesions, assisted over unaided read. Sub-group analysis. Illustrative placeholder only.
All figures are illustrative placeholders and do not constitute a claim of regulatory clearance or measured product performance.
How we run a reader study.
A rigorous reader study design is the foundation of any credible second-reader AI validation. Verityn follows a five-stage process, from protocol registration through to full analysis and reporting.
Protocol registration
Each study is defined in a written protocol before data access begins. The primary endpoint, cohort size, reader selection criteria, and statistical analysis plan are fixed in advance and do not change once data access opens.
Dataset curation
Cases are drawn from retrospective, de-identified imaging archives. Sampling targets a realistic case mix, including difficult and subtle findings, not a cherry-picked or enriched subset designed to favour any outcome.
Reader recruitment
A panel of radiologists with relevant subspecialty and general reading experience is recruited. Readers are independent of Verityn Ltd and blinded to each other's reports throughout both read arms.
Read design
Readers complete an unaided arm and then, after a controlled washout period, an assisted arm with Verityn flags visible. Order and washout length are pre-specified. A ground-truth reference standard is adjudicated by a separate expert panel.
Analysis and reporting
Primary metrics are sensitivity, specificity, and AUC across both arms. Pre-specified sub-group analyses cover finding size, modality, and reader experience level. Results are reported in full whether or not they favour Verityn.
Every model paired with its evaluation.
Verityn does not release a model update without a corresponding evaluation report. Each version carries a unique identifier, the validation dataset reference, the metrics recorded, and the date of lock. Previous versions and their reports are retained and remain accessible.
Unique version identifiers
Every released model version is assigned a unique alphanumeric identifier that travels with it into the audit trail and into any structured report it generates for a study.
Locked evaluation reports
Evaluation data is locked at the time of release. Subsequent model changes open a new evaluation cycle rather than overwriting or amending existing records.
Change impact assessment
Any modification to the model, training data, or post-processing pipeline triggers a documented impact assessment before deployment proceeds.
ISO 13485 aligned quality system and post-market surveillance.
Verityn is developed under a quality management system aligned with ISO 13485, the international standard for medical device manufacturers. It governs design controls, risk management under EN ISO 14971, supplier qualification, and the post-market surveillance system that operates after every deployment.
Design controls
Requirements, design outputs, and verification and validation activities are documented and traced through the design history file. Nothing ships without a signed-off design review.
Risk management
Clinical and software risk assessments follow EN ISO 14971. Residual risk is reviewed against clinical benefit before any release decision is taken. The risk file is a living document.
Post-market surveillance
A formal post-market surveillance system collects structured feedback from deployed sites, monitors complaint trends, and feeds findings back into risk management and product development.
Vigilance and reporting
Serious incidents involving a Verityn-assisted read are reported to the relevant competent authority within the timeline required by each jurisdiction's vigilance rules.
Want the full validation summary?
The clinical evaluation and validation summary is available to qualified clinical and procurement teams under a non-disclosure agreement. Our clinical team will walk you through the methodology and results. You can also read our regulatory posture to understand deployment context by region.
Give every read the evidence it deserves.
The clinical evaluation and validation summary is available under agreement. Our clinical team will take you through methodology, metrics, and what the results mean in your deployment context.