Key Points:
- In a randomized crossover study, authors assessed whether artificial intelligence (AI) can improve accuracy of cardiologists when evaluating echocardiograms for severe aortic stenosis (AS).
- While diagnostic sensitivity did not statistically improve, AI demonstrated perfect classification of severe AS and significantly reduced inter-reader variability and time to complete interpretations.
- The study highlights opportunities and barriers in integrating AI into clinical workflows.
Early and accurate recognition of severe aortic stenosis (AS) is essential to prevent heart failure, hospitalization, and death, but the diagnosis can be missed or delayed, even on routine echocardiograms. With the rise of artificial intelligence (AI)-enabled tools in cardiovascular imaging, there remains an unexplored resource to meaningfully augment clinical interpretation and improve patient outcomes.
In a featured Clinical and Investigative Horizons session at the American College of Cardiology Scientific Sessions on March 31, 2025, Dr. David A. Playford presented results from a randomized, case-control, fully-crossed, paired-reader, paired-case study to evaluate the performance of the EchoSolv AS technology in assisting cardiologists in rapidly and uniformly reporting aortic stenosis severity on resting trans-thoracic echocardiograms.
The study involved 2 board-certified cardiologists reviewed 200 echocardiograms (100 with severe AS and 100 controls), first unassisted and then assisted by AI. A third adjudicating cardiologist was used if there was disagreement. Finally, five independent cardiologists interpreted all 200 echocardiograms independently, totaling 2,000 reads. The AI did not analyze images directly but instead used key measurement inputs including peak AV velocity and LVEF to classify disease severity.
Cardiologists had a baseline error rate of 20.6% for diagnosing severe AS. After AI introduction, no cardiologist consistently accepted AI recommendation despite AI being 100% accurate in identifying severe AS. There was no improvement in Area under the Receiver Operating Characteristic (AUROC) between assisted and unassisted reads (p = 0.064). However, echo review times were significantly decreased by 24.1% (p < 0.001) with AI assistance and cardiologist concordance was improved with assistance.
In the ACC presentation, lead author Dr. Playford concluded that “AI correctly identified all cases of severe AS.” However, “cardiologists error rates persisted despite AI assistance, suggesting a lack of trust in AI systems.” Overall, this study highlighted the potential impact on AI in clinical workflow but that significant barriers exist in the uptake from providers.