Decision Support, Not Diagnosis

Every regulatory framework for AI in medical imaging eventually arrives at the same question: when does a software output become a diagnosis? The answer shapes the technical file, the clinical evaluation, the user interface, and the instructions for use. For Verityn, the answer has always been clear. The output is never a diagnosis. It is a flag, a location, a confidence, and a reason. The diagnosis belongs to the clinician.

A classification that matters

Under the UK Medical Devices Regulations 2002 and their post-Brexit amendments, software that influences clinical decisions is a medical device. Within that category, software as a medical device (SaMD) is further classified by the severity of the clinical decision it supports and the degree to which it replaces a human decision. The crucial axis is whether the software provides information that could cause serious harm if acted upon incorrectly, and whether that information is presented as a conclusion or as supporting evidence. Verityn is designed, tested, and labelled as decision support: the model provides a flag and the evidence for it, and the clinician makes the clinical judgement. This classification is not chosen for convenience. It reflects what the product actually does.

What the regulations require

The UKCA pathway for a class IIa medical device requires a technical file demonstrating compliance with general safety and performance requirements, a clinical evaluation, and a quality management system. Verityn is developing its technical file under a QMS aligned with ISO 13485:2016, with a clinical evaluation structured around reader studies comparing unaided reads to Verityn-assisted reads on retrospective, de-identified datasets. Parallel work is under way for the CE marking process under the EU Medical Device Regulation 2017/745, which applies in markets where Verityn intends to operate. Both pathways require the intended purpose to be stated precisely, and the intended purpose of Verityn leaves no ambiguity: it is a second reader, not a first opinion, and the clinician is responsible for the report.

The regulator is not asking whether the AI is accurate. It is asking whether the clinician remains in control of the decision.

The clinician-in-the-loop is the design

There is a version of AI-assisted reading in which the model's output arrives pre-integrated into the report, the clinician reviews it and approves or modifies, and the workflow pressure is such that most of the time the model's output becomes the report. This is not what Verityn is built for, and it is not what our integration design permits. Verityn flags are presented as an overlay on the study, visible before the primary read, and the radiologist forms their own view of the image before deciding what to do with any flag. The flag is an input to the clinician's reasoning, not a draft report.

What Verityn does and does not do

• Verityn does not generate a report text.
• Verityn does not assign a diagnosis code.
• Verityn does not communicate directly with the patient.
• Verityn does not modify the PACS worklist priority without a clinician acting on the flag.
• Verityn does not have a threshold below which it suppresses findings from the clinician's view.
• Every flag is visible and can be dismissed, overridden, or escalated by the clinician.

Liability runs through the clinician

Under English law, and under the clinical governance frameworks of the NHS and equivalent bodies, clinical responsibility for a diagnostic report lies with the clinician who signs it. A decision support tool does not dilute or share that responsibility. The clinician uses the tool, applies their professional judgement, and reports. If Verityn flags a finding the clinician reviews and discounts on clinical grounds, and that clinical judgement is subsequently challenged, the question before any professional body is whether the clinician's reasoning was defensible, not whether the algorithm agreed. Verityn's role is to make that reasoning better informed by ensuring that a second set of eyes has been over the image before the report is finalised.

Being precise about what Verityn is and is not has shaped everything from the model training objectives to the user interface copy. A clinician who understands they are receiving a flag, with a stated confidence and an indicated region, and who knows they are responsible for what they do with it, is in a better position than one who receives an opaque output with no stated uncertainty. Clarity is not just a regulatory obligation. It is the thing that makes the tool trustworthy.