Despite major advances in the implementation of artificial intelligence (AI) in Cardiology, from both a clinical and research perspective, no clinical trials to data have proven its benefit or efficacy in this space. The investigators of EchoNet-RCT sought to rectify this, by conducting and publishing the first randomized clinical trial studying the implementation of AI for the practice of clinical cardiology. In a Hot Line Session at the 2022 European Society of Cardiology Congress, Dr. David Ouyang (Smidt Heart Institute, Cedars-Sinai, Los Angeles) presented the results of their trial, testing the fidelity and accuracy of echocardiograms read by artificial intelligence.
Echonet-Dynamic is the proprietary AI software studied in the trial. The programming of the deep-learning algorithm had previously been published in NATURE. The software was trained on echocardiogram videos to assess cardiac function and determine left ventricular ejection fraction. Prior results showed the AI to accurately read ejection fractions with a mean absolute error of 4.1 – 6%. In order to reproduce the results and minimize error, the algorithm employs multiple cardiac cycles in its determination of ejection fraction.
In this prospective randomized control trial assessing Echonet-Dynamic in clinical use, the investigators sought to determine how often AI-generated LVEF was changed by the reading cardiologist. These changes would then be compared against changes the cardiologists made to the sonographer-generated LVEFs. To achieve this, 3495 echocardiograms obtained by cardiac sonographers were randomized in a 1:1 fashion to undergo LVEF tracings by the sonographer or by the AI. These scans were then sent to the reading cardiologist, who was blinded to the source of the EF assessment, for final interpretation. To assess the adequacy of blinding, the cardiologists were asked to choose whether the thought the EF assessment was AI-guided or sonographer-guided at the end of each scan, and chose correctly only 32.3% of the time.
The primary efficacy outcome was the change in initial EF (AI or sonographer-guided) vs. final EF assessment by the cardiologist. The AI-guided EF required a substantial change 16.8% of the time, compared to 27.2% in the sonographer-guided arm (mean difference -10.5% (95% CI -13.2% – 7.7%), p<0.001) meeting significance for both non-inferiority and superiority. The mean absolute difference of EF percentage points was also significant, favoring AI-generated assessments (2.79 vs 3.77, p<0.001). A pertinent secondary outcome was the time it took for reading cardiologists to overread each scan, and as expected, more time was spent reading the sonographer-guided scans (64 seconds vs 54 seconds, p<0.001. Subgroup analyses by patient characteristics and echo characteristics demonstrated consistent results in favor of AI-generated assessments.
The authors concluded that LVEF assessment by a proprietary AI algorithm was both noninferior and superior to an initial sonographer assessment, as cardiologists were less likely to change these assessments overall. As this was a single-center study, it is unclear how generalizable these results would be, and whether they can be externally validate. Still, this trial is promising for the future of AI in cardiology, and as Dr. Ouyang stated, the technology could “be very effective at not only improving the quality of echo reading output but also increasing efficiencies in time and effort spent by sonographers and cardiologists by simplifying otherwise tedious but important tasks”.