Cedars-Sinai researchers published what might be the strongest evidence yet of AI’s ability to improve echocardiography workflows and accuracy.
The researchers randomized 3,495 transthoracic echo LVEF studies to be assessed by either a Cedars-Sinai-developed AI tool or expert sonographers (within the standard cardiology PACS workflow), before final assessments by blinded cardiologists.
This was the first blinded RCT evaluating an echo AI solution, and one of the first for any imaging AI solutions.
Echo AI nailed the study’s primary endpoint, as the final cardiologist assessments required substantial LVEF changes (>5% change) far less often with AI-based studies than those from the sonographer group (16.8% vs. 27.2%).
The echo AI solution AI further supported its advantage over sonographer assessments, producing…
- A smaller average LVEF difference between the initial and final cardiologist assessments (2.79% vs. 3.77%)
- A smaller average LVEF difference between the final cardiologist assessments and patients’ previous/historical cardiologist assessments (6.29% vs. 7.23%)
- Faster average initial LVEF assessments (AI 131 seconds faster)
- Faster median cardiologist LVEF assessments (54 vs. 64 seconds)
The blinded cardiologists also often couldn’t tell whether AI or the human sonographers produced the initial assessments (43%), while predicting the correct assessment source only slightly more often than the incorrect source (32% & 24%).
This study shows that AI-based LVEF assessments can be “non-inferior and even superior” to initial assessment by sonographers, while also reducing labor and variability throughout the echo workflow.
Even though previous echo AI research has produced similar results and this study only used data from a single center, the fact that it was echo AI’s first blinded RCT is very notable, and bolsters the growing body of evidence supporting AI’s ability to improve echo workflows in the real world.