Poulos C, Gebben D, Peay H, Saha A, Vaezy S, Pierce A, Pina I, Farb A, Moultrie R, Babalola O, Tarver ME. Benefit-risk preferences of patients for the use of artificial intelligence and ultrasound imaging in different settings in echocardiography. Poster presented at the ISPOR 2023 Conference; May 7, 2023. Boston, MA. [abstract] Value Health. 2023 Jun; 26(6 supplement):S349. doi: 10.1016/j.jval.2023.03.2378

OBJECTIVES: To measure the benefit-risk preferences of US adults with heart failure (HF) for artificial intelligence (AI)-assisted echocardiograms (echos) in non-conventional settings.

METHODS: Using an discrete-choice experiment (DCE), adults with HF chose between experimentally-designed pairs of hypothetical AI-assisted echos comprising four attributes with varying levels: site of administration (primary care office or home), risk of unusable images (5%, 20%, 45%, 60%), how results are interpreted (AI-based or cardiologist confirmation of AI-based), and risk of false-negative echo results (10%, 15%, 25%, 30%, 40%, 50%). Respondents could also prefer a standard echo (cardiologists’ office without AI). Choices were analyzed using random-parameters logit. Conditional relative attribute importance (CRAI) out of 100% and maximum acceptable risks were calculated.

RESULTS: The sample included 171 adults - 75 with physician-confirmed and 96 with self-reported HF diagnoses. On average, respondents had 2.7 echos during the prior 2 years. Most (89.5%) respondents were in New York Heart Association Class II or greater. On average, respondents preferred AI-assisted echos over standard echos. An AI-assisted echo’s risk of a false-negative had the greatest CRAI (64.1%), followed by its risk of unusable images (CRAI, 29.7%). Respondents preferred cardiologist confirmation of AI-based results to AI-based interpretation alone and preferred AI-assisted echos administered at home to those administered in a primary care setting, but these two attributes had low importance (CRAI 3.2% and 3.0%, respectively). For AI-assisted echos, respondents were willing to tolerate the following maximum risks in exchange for home-based rather than primary care-based echos: 12.5% risk of unusable images and 11.0% risk of false-negative results. Respondents would tolerate similar maximum risks in exchange for cardiologist confirmation of AI-based interpretation rather than AI-based interpretation alone (13.0% and 11.1%, respectively).

CONCLUSIONS: While respondents preferred AI-assisted to standard echos, image usability and false negatives were stronger choice drivers than setting or cardiologists’ interpretation.