Alignment Between Cardiologists and AI-Driven Diagnostic Systems: Mixed Methods Study.

The research team employed a mixed methods design to evaluate diagnostic alignment between clinicians and automated software. A quantitative analysis examined data from 854 participants enrolled in the PROTEUS randomized controlled trial. Logistic regression models identified predictors of agreement, disagreement, and scan rejection while adjusting for various cardiovascular risk factors. A qualitative approach involved distributing surveys to 61 UK consultant cardiologists recruited via the Qualtrics platform. This survey explored professional perceptions regarding the risks and benefits of following algorithmic recommendations. The review approach synthesized these findings to understand decision-making strategies during instances of discordance. Statistical adjustments accounted for age, sex, smoking status, and body mass index to ensure robust comparisons. This dual-pronged strategy provided both broad performance metrics and deep insights into the human factors influencing clinical adoption.

The AI system and cardiologists achieved a 60% agreement rate across the 854 analyzed cases. Agreement was significantly lower for patients with hypertension, diabetes, or pre-existing coronary artery disease. The software rejected 26.1% of scans due to insufficient image quality, with higher rejection rates observed in male patients. Combining human and machine diagnoses increased the identification of positive cases from 17.9% to 22.1%. Surveyed cardiologists required between 65% and 69% confidence in their own initial diagnosis to disregard contradictory AI recommendations. Respondents consistently treated algorithmic outputs as advisory rather than definitive, preferring to seek second opinions or additional testing. A paradoxical finding showed that cardiologists with higher trust in AI tools required even greater personal confidence to ignore the system. The primary cause of discordance was identified as the inability of the software to incorporate patient history and clinical context.

The authors propose that automated systems currently function best as advisory aids rather than definitive diagnostic authorities. Clinicians prioritize their own expertise and seek further verification when algorithmic outputs contradict their initial assessments. This synthesis suggests that the inability of software to incorporate broader patient history remains a primary driver of diagnostic discordance. The researchers note that high rejection rates for image quality may limit the utility of these tools in diverse patient populations. Implications for future development include the necessity of integrating comprehensive clinical data to improve diagnostic accuracy. The findings imply that clinicians use these tools primarily to prompt deeper scrutiny of their own findings. The authors indicate that current models may inadvertently exacerbate existing healthcare inequities if they are not trained on representative datasets. These results highlight the need for systems that better align with the complex, multimorbid nature of real-world clinical practice.