For healthcare professionals and procurement teams. This is editorial analysis, not patient advice, a product recommendation or a statement of South African availability.
A cardiac image can be the starting point for a risk model without automatically becoming a clinical decision tool. That is the distinction at the centre of our review of CARDINAL, a recently posted research framework using non-contrast cardiac CT. The topic is relevant to imaging, prevention and service planning, but the evidence reported here remains preliminary. [1]
What the researchers reported
Gabriel and colleagues evaluated CARDINAL in 17,659 patients for prediction over one-, three-, five- and ten-year horizons. The framework learns compact imaging representations rather than relying only on a small set of predefined measurements. For the joint model at ten years, the abstract reports an AUROC of 0.866 ± 0.020 and an AUPRC of 0.890 ± 0.015. [1]
Those are retrospective model-evaluation results. The plus-or-minus quantities are reproduced as reported, not relabelled as confidence intervals. This commentary does not establish their statistical definition. Nor does it report a trial in which using the model improved care or outcomes. [1]
Start with the intended decision
Our editorial question is deliberately practical: if a service had this output, which decision would it be intended to inform? A risk model proposed for one purpose should not quietly acquire several others in a review article. We would ask a future evaluation to state its user, setting, prediction horizon and proposed consequence before debating a headline performance metric.
That exercise would also help separate a research question from a purchase proposal. A model may be scientifically interesting while many details of a usable implementation remain unexamined. The appropriate response is to document those details as questions, not to fill the gaps with assumptions about the software, its support or its local availability.
Discrimination and the number a clinician sees
In the full appraisal, we would examine how reported performance relates to the output presented to a user. Does the interface present a continuous estimate, a category or an alert? What threshold, if any, is proposed? How was that threshold selected? We would want answers from the relevant study materials rather than infer them from a single summary number.
The review should also ask how the model's estimated risks align with observed outcomes in each evaluation setting. A complete appraisal of calibration should identify the population, horizon and uncertainty involved. This commentary does not turn the abstract's brief reference to calibration into a claim that calibration has been demonstrated for South African practice.
What a South African review could examine
Our suggested local assessment would begin with the intended population and the available imaging inputs. A committee could ask which examinations would be eligible, which would be excluded and how incomplete data would be handled. It could then ask whether an independent evaluation reflects the same proposed use. These are planning questions, not a recommendation to adopt CARDINAL.
We would separately ask who is accountable for interpreting an output and documenting its use. The research team, reporting clinician, information-technology team and governance function may each need different information. A future evaluation proposal could describe those responsibilities explicitly and include a method for recognising when the system is not operating as intended.
For procurement, a detailed documentation request would be more useful than a broad statement that an algorithm is advanced. Our proposed request would cover the exact version, intended-use scope, supported inputs, deployment requirements, update policy, maintenance arrangements and applicable local authorisation. This article establishes none of those commercial details.
Keep non-contrast imaging claims narrow
We would not use the fact that this study analyses non-contrast CT to advertise a renal-protection benefit. That would be a different clinical claim. A future cardiorenal discussion would need evidence directed at that question, including the relevant population and outcomes, rather than borrow certainty from the imaging method's name.
We would also avoid suggesting that a research risk estimate substitutes for independent clinical judgement. The purpose of this journal is to help professional readers organise evidence and questions. It is not to provide a screening recommendation, individual risk assessment or management plan for a patient reading the website.
What an expert interview should add
Our preferred expert question is not whether AI will transform imaging. It is which external-validation result would change the expert's assessment of this particular framework for a particular proposed use. A precise answer could identify a population, outcome or failure mode that deserves closer examination.
This article contains no independent expert interview. It is commentary on the publicly indexed preprint abstract, not a complete critical appraisal of the manuscript. Uncertainty definitions, follow-up and subgroup methods require separate examination. The reported figures are presented as preliminary research findings; they are not evidence of prospective clinical utility or local validation.
What to watch next
We would follow independent replication, complete calibration reporting, version-specific testing and a prospectively stated clinical-use evaluation. Our standard for future coverage is that each new claim should be attached to the evidence that directly tests it. A research model's potential and its demonstrated utility are separate editorial questions.
A risk estimate needs an intended decision before it has an intended workflow. — Aperture Science editorial perspective.
Source material
