Left ventricular ejection fraction – human vs artificial intelligence: quo vadis?

Abstract

Abstract Background/Introduction Two-dimensional echocardiography (Echo) is a feasible method for assessing left ventricular (LV) ejection fraction (EF) in daily practice. However, the interpretation of Echo exams depends on the user's expertise and may vary between different operators. A novel, vendor-neutral artificial intelligence (AI) performs both, automated evaluation of Echo exams and calculations of biplane LV EF in one workflow. Purpose We sought to assess the ability of the AI to automatically identify appropriate LV 4- and 2-chamber views (4CV) (2CV) from routine Echo examinations and compare the resulting biplane EF with conventional hand-tracing biplane Simpson method (Human). Methods We prospectively enrolled 311 patients who underwent clinically indicated Echo exams. Biplane LV EF was manually traced online on 4CV and 2CV by cardiologists (Human). After completion of the exam, the AI-based solution recognized the optimal LV 4CV and 2CV according to quality and depth criteria and automatically performed the calculation of biplane EF by endocardial borderline detection without Human's interaction. Spearman's correlation (R) and Bland-Altman analysis with limits of agreement (LOA) were assessed for bias between the two methods. In a subgroup of 20 patients, Echo exams were automatically reanalyzed by the AI, and conventional biplane Simpson of LV EF was performed by two cardiologists blinded to the previous results to determine intraclass correlation (ICC). Significance was defined as a 2-tailed p value <0.05. Results 311 patients (median age 72 years [19–97]; 40% female) received an Echo for valvular heart disease, ischemic and non-ischemic heart failure or other indications (39, 31, 19 and 11%). 16 cases (5%) did not pass AI's criteria due to poor Echo imaging or impaired acoustic window of patients. In 53 patients (17%) either 4CV or 2CV were recognized, but the AI system successfully identified both 4CV and 2CV in 242 patients (overall feasibility 78%). For these 242 patients, correlation between AI and Human biplane LV EF was r=0.83 (p<0.001) (Figure 1 left). The absolute mean bias between methods was 5.2% (p<0.001) and absolute LOA ranged from −9.0 to +19.4% (Figure 1 right). ICC of LV EF by Human was 0.77 (p<0.001). The AI's ability to correctly re-/classify 4CV and 2CV was 100% with an ICC of 1 for fully automated LV EF measurements. Conclusion The results provided by the AI-based software showed very good capability to identify 4CV and 2CV and good LV EF result compared to Human manual tracings, especially since patients were not pre-selected. However, differences between AI and Human measurements are not negligible and warrant further investigations. Funding Acknowledgement Type of funding sources: None.