Detection of severe aortic stenosis by clinicians versus artificial intelligence: A retrospective clinical cohort study

Abstract

Many severe aortic stenosis (AS) cases are undetected and/or not considered for potentially life-saving treatment, with a persistent male-bias reported among those undergoing aortic valve replacement (AVR). We evaluated the clinical value of a validated artificial intelligence automated alert system (AI-AAS) that detects severe AS from routine echocardiographic measurements. In a retrospective, clinical cohort of 21,749 adults investigated with transthoracic echocardiography at two tertiary-referral centres, we identified 4057 women (aged 61.6 ± 18.1 years) and 5132 men (60.8 ± 17.5 years) with native aortic valves. We firstly applied the AI-AAS to the cardiologists' reported echo measurements, to detect all AS cases, including guideline-defined severe AS. Two expert clinicians then independently reviewed the original clinical diagnosis/management based on the initial report. Initially, 218/9189 (2.4 %, 95%CI 2.1–2.7 %) severe AS cases were diagnosed. The AI-AAS subsequently increased this number by 158 (52 % women) to 376 cases (4.1 %, 95%CI 3.7–4.5 %) of severe guideline-defined AS. Overall, more women were under-diagnosed (92/169 [54.4 %] versus 80/207 [38.6 %] men – adjusted odds ratio [aOR] 0.21, 95%CI 0.10–0.45). Even when accounting for potential contraindications to valvular intervention, women were persistently less likely to be considered for valvular intervention (aOR 0.54, 95%CI 0.31–0.95) and/or underwent AVR (aOR 0.29, 95%CI 0.09–0.74). Our study suggests an AI-AAS application that is agnostic to gender, haemodynamic bias, symptoms, or clinical factors, provides an objective alert to severe forms of AS (including guideline-defined severe AS) following a routine echocardiogram, and has the potential to increase the number of people (especially women) directed towards more definitive treatment/specialist care.