Performance of artificial intelligence system for prescriptive acquisition guidance of transthoracic echocardiography by novice users combined with automated quantification of ejection fraction

Abstract

Abstract Funding Acknowledgements Type of funding sources: Private company. Main funding source(s): Caption Health Background Artificial intelligence (AI) has the potential to expedite the acquisition of transthoracic echocardiograms (TTE) and provide automated quantitative data including left ventricular ejection fraction (LVEF). Specifically, AI-based imaging systems may permit less experienced individuals to obtain quantitative measurements of LVEF, with important implications for clinical workflow. Purpose We sought to evaluate the accuracy of an AI-based imaging system for the evaluation of LVEF in a spectrum of novice TTE imagers in a real-world clinical setting, and hypothesized that after appropriate training, it can perform as well as experienced sonographers. Methods Consecutive exams (N = 102; BMI 29 ± 7; most common indications for TTE: heart failure, arrhythmia, valve assessment) were performed by a cohort of physician trainees (MD) and nurses (RN) with no prior TTE experience, using an AI-guided imaging system equipped with real-time prescriptive guidance software that automatically measures LVEF. Guided imaging included 3 views: parasternal long-axis (PLAX), apical 4-and apical 2-chamber (AP4, AP2), with the software recognizing when a satisfactory imaging window is obtained and then auto-capturing and automatically generating an AI based LVEF. AI-EF measurements were compared against the reference standard EF (Ref EF) measured by 2 expert sonographers according to ASE chamber quantification guidelines. Bland-Altman analysis was performed to determine inter-technique agreement. Results Feasibility was 80%. AI-EF and Ref EF demonstrated strong correlation when all 3 views were obtained, with a small bias (Table). In cases when <3 views were available, the combination of PLAX and AP4 views achieved comparable results, with a slight increase in bias and limits of agreement. When results were stratified by cohort (MD vs RN), MD AI-EF attempts showed greater feasibility (100%, n = 51) with stronger correlation (r = 0.93) and smaller bias (-1.9%) compared to RN (feasibility of 61%; n = 31 and r= 0.85, bias -2.1%). Conclusions Use of an AI-assisted imaging system for limited TTE imaging by novices is feasible in a real-world setting, with the AI based EF in good agreement with reference standard. Acquisition of all 3 views provided optimal results, but the combination of AP4/PLAX views performed reasonably well, without the AP2 view that is more difficult to acquire for less skilled users. Untrained MD were more successful when compared to RN, suggesting additional training may be needed for specific user groups. Abstract Figure.