Cardiologists and Patients See Value in an AI Tool for Excluding Coronary Artery Disease
A qualitative study of prospective end users found broad interest in an AI decision support system designed to rule out coronary artery disease, but the enthusiasm came with clear ethical and clinical reservations. Cardiologists and patients generally viewed the ARGUS algorithm as a useful supplement to clinical judgment, especially when it promised better accuracy and faster decision-making. The appeal was not limited to the technical performance of the model. For many respondents, the real draw was the possibility of reducing uncertainty in a crowded clinical pathway where decisions about imaging, follow-up, and reassurance often carry substantial weight.
The study centered on semi-structured interviews with 11 cardiologists and 20 cardiology patients. Those conversations were transcribed and analyzed thematically, allowing recurring concerns to emerge from the viewpoints of both clinicians and patients rather than from a fixed questionnaire. Participants described an AI-DSS that could help exclude disease as attractive in principle, but only if it fit the realities of clinical work and patient expectations. A low-risk result on a screen does not automatically translate into confidence at the bedside.
Among the clearest themes was a preference for usefulness over novelty. Respondents repeatedly returned to accuracy and efficiency as the main expected gains, while the idea that the tool might lessen the burden on patients appeared less prominent. In practice, that means the value proposition was framed less as a humane shortcut and more as a way to sharpen triage and support decision-making. For cardiologists, this kind of assistance could reduce ambiguity when determining who truly needs additional testing. For patients, it could offer a clearer explanation of why further workup is or is not necessary.
One sentence from the interviews stands out in clinical terms: a low likelihood of disease may still not feel like enough reassurance.
Reassurance Remains a Difficult Endpoint in Disease-Exclusion AI
The strongest hesitation involved the psychology of exclusion. Many cardiologists and patients questioned whether a negative or low-risk AI result would actually calm concern, particularly when the system is used to exclude disease rather than confirm it. If a patient remains worried about symptoms, family history, or prior experiences, an algorithmic estimate may not end the conversation. Some may still want a scan regardless of the model’s recommendation, which creates a direct tension between probabilistic output and human fear.
This issue matters because exclusion tools are not only judged by whether they are statistically sound. They are also judged by whether they can function as a credible answer in a real consultation. A model that says “unlikely” may be technically appropriate and still fail to persuade. That gap between algorithmic confidence and patient acceptance is not a minor communication problem; it shapes whether the tool can actually reduce downstream testing or merely add another layer of interpretation.
The interviews also pointed toward a broader shift in how responsibility may be distributed. Participants recognized that introducing AI into routine cardiology changes who is expected to interpret, defend, and ultimately act on the result. If the system suggests one course and the clinician chooses another, responsibility cannot be left vague. The study found that questions about accountability need to be addressed before deployment, not after a difficult case has already occurred.
Shared Responsibility and the Limits of Automation in Cardiology
The role of the clinician did not disappear in these interviews. It became more central. Respondents described a setting in which AI may inform decisions, but the final judgment still belongs in the clinical encounter, especially when symptoms, history, or clinical intuition conflict with the model. Human-AI disagreement was not treated as an edge case; it was presented as an expected part of use in real practice.
That makes overruling the algorithm a necessary feature of implementation, not a breach of protocol. The study’s findings suggest that cardiologists want room to reject an AI recommendation when the broader clinical picture does not fit. At the same time, patients need to understand why that override happens. Without that explanation, trust in the system may erode quickly. If an algorithm is introduced as authoritative but can be set aside without clear reasoning, both confidence and accountability become unstable.
Responsibility therefore appears to be shared, but not in an abstract way. The burden falls on systems, clinicians, and patients to understand what the model can and cannot do. An AI-DSS for excluding CAD has a narrower task than a diagnostic tool that identifies disease, yet the ethical demands may be just as substantial. A negative result can be persuasive only if the surrounding workflow makes room for explanation, uncertainty, and clinical judgment.
Sex-Specific Design Emerged as a Concrete Ethical Requirement
One of the most specific findings was the call for sex-specific AI-DSSs in cardiology. Participants identified this as an important ethical and practical issue rather than a theoretical preference. Cardiovascular disease presentation, risk patterns, and referral decisions can vary by sex, and respondents wanted those differences reflected in how the algorithm is built and applied. For a disease-exclusion tool, the concern is obvious: if the model performs unevenly across groups, it could falsely reassure one subgroup while overtesting another.
This point gives the study a broader relevance beyond one algorithm. It shows that end users are not only asking whether AI works; they are asking for whom it works, under what conditions, and with what consequences. Fairness, in this context, is tied directly to clinical credibility. A system that ignores sex-specific variation may be technically elegant and ethically brittle.
Patients also appeared to judge the system through the lens of personal relevance. A tool that speaks in probabilities can feel detached unless it is clearly connected to the individual sitting in front of the clinician. The interviews suggest that acceptance will depend on more than model performance metrics. It will depend on whether the AI fits the consultation, respects uncertainty, and preserves the patient’s sense that the final decision is still anchored in a human relationship.
Ethics Will Shape Whether Disease-Exclusion AI Enters Routine Care
The study offers a practical warning for developers of medical AI: disease exclusion may be the next frontier, but trust will not come from accuracy alone. Clinicians and patients in this cohort saw the potential value of the ARGUS AI-DSS, yet they also identified the points where implementation could fail. Reassurance, shared responsibility, and the right to overrule the algorithm all emerged as central to acceptance.
These findings have immediate relevance for hospitals and developers planning real-world adoption. A model that excludes coronary artery disease must be introduced with clear guidance on how to handle disagreement, how to explain low-risk outputs, and how to account for sex-based differences in performance and interpretation. Without that structure, the tool risks becoming another source of uncertainty rather than a support for care.
What remains unresolved is how these concerns will hold up once the algorithm is used in actual clinical pathways rather than interview discussions. The study establishes the concerns clearly, but the next step will be to see whether workflow design, clinician communication, and patient-facing explanations can close the gap between algorithmic exclusion and meaningful reassurance.