Cardiac imaging sits at a crossroads: AI tools are fast, scalable, and increasingly accurate, yet how they perform alongside — not instead of — experienced clinicians in real-world cardiology is far less understood than headline accuracy benchmarks suggest. The answer has direct implications for which patients get the right diagnosis on the first try.
This mixed-methods study drew on data from 854 participants with suspected coronary artery disease enrolled in the multicenter PROTEUS randomized controlled trial to evaluate diagnostic concordance between EchoGo Pro, an AI-driven stress echocardiography platform, and consultant cardiologist interpretations. Logistic regression modeled predictors of agreement, disagreement, and AI scan rejection, adjusting for age, sex, BMI, smoking, and a range of cardiovascular risk factors including hypertension, hypercholesterolemia, diabetes, and prior CAD events. A parallel qualitative arm surveyed 61 UK consultant cardiologists to probe how they navigate clinical decisions when AI outputs conflict with their own assessments — a dimension rarely captured in purely quantitative validation studies.
What makes this work analytically valuable is its dual-track design. Most AI cardiology studies report accuracy against a ground truth; this one foregrounds the human-AI interface and the social dynamics of disagreement. The cardiologist survey reveals that clinician response to AI conflict is neither automatic override nor automatic deference — it appears to depend on factors like patient risk profile, confidence in image quality, and institutional culture around AI adoption. This matters because scan rejection rates (cases where cardiologists decline to use the AI output) could systematically disadvantage certain patient subgroups if rejections cluster around particular demographics or imaging conditions. The study is limited by its single-device focus, UK-centric clinical context, and the fact that qualitative survey responses may not reflect actual bedside behavior. Still, as AI cardiac tools move toward regulatory approval and broader deployment, this type of concordance mapping offers a more realistic picture of clinical integration than benchmark accuracy figures alone.