Artificial intelligence and the promise of uplifting echocardiography

Abstract

Echocardiography is the most frequently used imaging modality to assess patients with cardiac diseases and the one with the most favourable cost/effectiveness ratio. Since its introduction in the clinical arena, echocardiography has changed our understanding of the pathophysiology of many cardiac diseases and the way we assess them, replacing invasive cardiac catheterisation in most of the cases. However, both the acquisition and the interpretation of the echocardiography studies, as well as the reproducibility and the repeatability of the measurements, rely heavily on the expertise and the experience of the operators. This is why the practice of echocardiography is considered to be a mixture of craft and science. Nowadays, the art of echocardiography is in danger. The progressive ageing of the general population and the related increased prevalence of cardiovascular disease on one end, and the ageing of the operators associated to the time constrain to perform more and more echocardiography studies on the other end, have created an unprecedented time crunch to perform and interpret an increasing number of studies that may lead to burnout and reporting errors. The recent development of artificial intelligence (AI) techniques to make automated segmentation and quantitative analysis of the echocardiography images offer a solution to reduce echocardiographer workload and to improve the reproducibility and the repeatability of the measurements (figure 1). What remains to be documented is the accuracy of the fully automated measurements performed by the AI algorithm and their clinical significance in comparison to the performance of skilled echocardiographers. To this end, the study by Kitano and coworkers in this issue of Heart adds to the growing enthusiasm for the use of AI algorithms that automates several facets of echocardiography measurement and interpretation. Kitano and coworkers used a novel, commercially available software package able to automatically determine the endocardial border of the cardiac chambers using a knowledgebased AI algorithm in each apical view and to perform speckletracking analysis through one complete cardiac cycle. Their aim was to investigate whether the fully automated twodimensional longitudinal strain of the left ventricle (LV), the left atrium (LA) and the right ventricle (RV) could predict future outcomes in 340 asymptomatic patients with aortic stenosis. They demonstrated that the fully automated analysis of the 2D longitudinal strain was highly feasible (95% for the LV and the LA) provided that adequate 2D views were acquired. 4 Interestingly, the feasibility of RV strain was 94% (84/89) when a RVfocused view was available for analysis, but only 76% (191/251) when an apical fourchamber view was used. Moreover, the procedure was also timeeffective. Total analysis time was 36±2 s in the 264 patients in whom they analysed all four strains. These data confirm the costeffectiveness of the applications of AI to quantitate echocardiography studies. However, Kitano et al notably looked also at the prognostic value of the automated longitudinal strain measurements. During a median 29month followup, 51/340 patients reached the composite primary endpoint of cardiac death, heart failure hospitalisation, myocardial infarction or ventricular arrhythmia. Multivariable analysis confirmed that LV longitudinal strain was significantly associated with cardiac events, even after adjusting for the haemodynamic parameters of aortic stenosis severity and LV ejection fraction, as previously reported. However, the really interesting novel finding by Kitano et al was that the accuracy of fully automated LV longitudinal strain (evaluated using the receiveroperating characteristic analysis) in predicting that patient outcome was similar to that of manually edited measurements performed by an expert at strain analysis (area under the curve 0.66 vs 0.67, respectively). In summary, the study by Kitano et al shows that fully automated LV longitudinal strain is highly feasible, fast, highly reproducible and has similar prognostic value to the measurements obtained by an expert echocardiographer. In addition, the same software package can automatically provide the LA and RV longitudinal strain values that have been reported to hold prognostic value in patients with severe aortic stenosis 8 (figure 2). One important point that should be underlined in order to put the study by Kitano et al in the proper clinical perspective is that AI can identify the correct views and perform accurate measurements only if the echocardiographic views are properly acquired. Acquisition of LV foreshortened views, conventional fourchamber instead of the RVfocused apical view or nondedicated apical views for the LA has not been recognised by the current version of the software package and will provide misleading results. The improvement of the AI algorithm is expected to