October 18, 2024NEJM AI17 citationsOpen Access
View Full PaperPROTEUS: A Prospective RCT Evaluating Use of AI in Stress Echocardiography
Key Result
AI-augmented decision-making in stress echocardiography did not meet non-inferiority for appropriate angiography referral compared to standard care (AUROC difference 0.09; 95% CI -0.22 to 0.39).
Study Design
Type
RCT (n=2,341)
Randomization
parallel-group
Multicenter
Yes
Structured PICO
Does decision-making augmented by AI improve appropriate referral for coronary angiography in participants undergoing a stress echocardiogram?
P
Population
2,341 participants undergoing a stress echocardiogram across 20 centers in the UK, followed for 6 months.
I
Intervention
Decision-making augmented by artificial intelligence (AI)
C
Comparator
Standard clinical decision-making
O
Outcome
Appropriate referral for coronary angiography (true positives defined as severe coronary disease requiring revascularization in participants referred for invasive angiography and false negatives defined as an acute coronary event within 6 months)surrogate
AI-augmented decision-making in stress echocardiography did not meet the non-inferiority endpoint for appropriate referral for coronary angiography compared to standard decision-making.
Main Result
Mean Difference: 0.09 (95% CI -0.22–0.39)
Absolute Event Rate: 0.63% vs 0.55%
Abstract
BackgroundUse of artificial intelligence (AI) in cardiovascular imaging may potentially augment clinical decision-making in disease management, but no prospective randomized controlled trials have assessed the impact on cardiovascular outcomes. This study evaluates whether AI–augmented decision-making is non-inferior to standard decision-making when selecting participants for invasive coronary angiography following stress echocardiography.MethodsPROTEUS was a multicenter, parallel-group randomized controlled trial. We enrolled participants undergoing a stress echocardiogram at 20 centers across the United Kingdom between November 2021 and June 2023. Participants were randomly assigned to standard clinical decision-making (control) or decision-making augmented by AI (intervention). The primary end point was appropriate referral for coronary angiography, with true positives defined as severe coronary disease requiring revascularization in participants referred for invasive angiography and false negatives defined as an acute coronary event within 6 months. Secondary analysis examined intervention versus control in prespecified subgroups where interpretation is known to be more challenging.ResultsOut of 2341 randomly assigned participants, 2213 (94.53%) completed 6 months’ follow-up. Eighty-five participants were referred for angiography, 61 of whom had significant coronary disease. Of the participants not referred, 41 participants had acute coronary syndrome or died within 6 months. The difference between the area under the receiver operating characteristic curve (AUROC) for the intervention (0.63; 95% confidence interval (CI), 0.43 to 0.83) and control (0.55; 95% CI, 0.33 to 0.80), did not meet the prespecified non-inferiority margin of −0.05 (difference, 0.09; 95% CI, −0.22 to 0. 39). The sensitivity in the intervention (64.2%; 95% CI, 33.3 to 80.0%) and control (55.1%; 95% CI, 43.7 to 84.2%) was similar (difference, 9.1%; 95% CI, −21.8 to 39.6%). Likewise, the specificity in the intervention (98.6%; 95% CI, 98.1 to 99.8%) and control (99.2%; 95% CI, 97.2 to 99.5%) was similar (difference, 0.6%; 95% CI, −2.1 to 0.9%). Subgroup analyses suggest potential benefit of AI–augmentation in low-volume stress echocardiography centers.ConclusionsAI–augmented decision-making in stress echocardiography did not meet the non-inferiority end point when evaluated in a large, prospective randomized controlled trial, but may be beneficial in low-volume centers. (Funded by the Accelerated Access Collaborative and others; ClinicalTrials.gov number, NCT05028179; ISRCTN number, ISRCTN15113915.)