A new study has shown no significant differences to demonstrate that using AI to aid clinical decision-making in assessing heart ultrasounds is as effective as current practice at identifying suspected heart disease.  

However, the findings presented at the ESC Congress 2024 in London, showed that AI improved decision-making for less-experienced clinicians and had promising results in important sub-groups that are clinically complex.

Stress echocardiography (SE) is a common ultrasound of the heart during rest and stress that determines risk of heart attack and death in patients with known or suspected CAD. However, its accuracy varies widely depending on the expertise of the clinician assessing the scans and image quality.

Dr Ross Upton, lead author, and Prof Paul Leeson, University of Oxford. developed software designed to provide automated interpretation of SE images by combining novel image features with AI.

“Integration of AI into healthcare holds great promise as a tool to help medical professionals diagnose patients faster and more accurately, allowing them to start treatment sooner,” said Dr Upton.

The PROTEUS trial enrolled patients aged 18 and older who were referred to SE clinics for investigation of suspected CAD at 20 hospitals across the UK between November 2021 and June 2023. 

Some 2,341 patients were randomly allocated to standard clinical-decision making or AI-augmented decision-making, for which clinicians received an AI image analysis report to use during image interpretation, indicating likelihood of severe CAD.

The primary analysis evaluated the appropriateness of standard clinical decision-making vs AI-augmented decision-making when selecting patients for coronary angiograms and related acute coronary events within six months. By December 2023, 2,213 patients (94.5%) achieved six-month follow-up. 

The analyses found that AI-assisted decision-making did not demonstrate non-inferiority vs clinical decision-making for correctly selecting patients for coronary angiography. Of those sent for angiography, 27 out of 36 referrals were correct in the control arm and 34 out of 49 referrals were correct in the AI arm.   

Of the patients that should have been sent for an angiogram who subsequently experienced an event, there were 22 in the control group and 19 in the AI group; the difference was not statistically significant. 

However, further analyses found that AI may benefit clinical decision-making in less experienced clinicians and in important sub-groups whose images are difficult to interpret.