JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) holds promise for enhancing diagnostic accuracy in cardiology. However, the extent of diagnostic concordance between AI tools and clinicians, and how clinicians respond to discordant assessments, remains poorly understood, particularly in stress echocardiography for coronary artery disease.

Objective:

This study examines the diagnostic alignment between an AI-driven stress echocardiography (SE) analysis system (EchoGo Pro; EGP) and conventional clinical judgment by expert cardiologists. It also explores clinicians’ perceived risks of following EGP recommendations and their responses to diagnostic disagreement.

Methods:

A mixed-methods approach was used. The quantitative study assessed concordance in coronary artery disease (CAD) diagnosis between EGP and cardiologists using data from 854 participants with suspected CAD from the PROTEUS trial. The qualitative study analysed open-text responses from a 2023 survey of UK consultant cardiologists, where AI tools were considered in general rather than focusing solely on EGP.

Results:

EGP rejected 26.0% of SE scans, with males’ scans more likely to be rejected [odds ratio (OR)=1.38 (1.01-1.88), P<.05]. The alignment between EGP and cardiologists were observed in 60.0% of cases. Cardiovascular risk factors were associated with reduced concordance, including hypertension [OR = 0.58, 95% CI: 0.38–0.89], diabetes [OR = 0.56, 95% CI: 0.35–0.90], and pre-existing CAD [OR = 0.48, 95% CI: 0.30–0.77]. Cardiologists in the qualitative study supported this observation, citing EGP’s limited capacity to incorporate broader clinical context, which may lead to diagnostic discordance. The most frequently reported response by clinicians in cases of disagreement with AI was to conduct additional diagnostic testing.

Conclusions:

EGP shows promise as a diagnostic support tool, but its limitations, particularly the exclusion of patient history and comorbidities, could lead to clinically meaningful diagnostic discordance. Cardiologists responded to AI disagreement by seeking further validation rather than accepting AI recommendations, indicating a preference for confirmation over relying on automation.