PROTEUS

PROTEUS: a Prospective RCT Evaluating Use of AI in Stress Echocardiography

Patient / Population Intervention / Exposition Vergleich Endpunkt

In 2,341 patients undergoing stress echocardiography at 20 UK centres, AI-augmented clinical decision-making did not show non-inferiority to standard clinical decision-making for appropriate referral for coronary angiography.

N
2.341
Design
Multicentre parallel-group non-inferiority RCT, UK (20 centres)
Endpunkt
Appropriate referral for invasive coronary angiography (AUROC; true positive = severe disease requiring revascularisation, false negative = acute coronary event within 6 months; non-inferiority margin −0.05)
Relevanz
2
ErgebnisAUROC difference (AI-augmented vs standard care) 0.09 (95% CI −0.22 to 0.39), not meeting the −0.05 non-inferiority margin; sensitivity and specificity similar between groups; 2,213 (94.5%) completed 6-month follow-up; subgroup signal of benefit in low-volume centres.
Upton R, et al. NEJM AI. 2024;1(11):AIoa2400865. 10.1056/AIoa2400865
Diskussion & Kritik

The first multicentre RCT of AI decision support in stress echocardiography, and a negative one: the point estimate favoured AI but the confidence interval was far too wide to exclude the non-inferiority margin. It taught that AI-reader trials must be powered for their margin and that any benefit may be concentrated in low-volume centres where reader expertise is thinner. The facts here come from the University of Bath repository abstract summary rather than the primary abstract. The same first author's earlier model-development and reader study appears as the Upton AI stress echo entry.