Artificial intelligence based left ventricular ejection fraction and global longitudinal strain in cardiac amyloidosis.

BACKGROUND

Assessment of left ventricular ejection fraction (LVEF) and global longitudinal strain (GLS) plays a key role in the diagnosis of cardiac amyloidosis (CA). However, manual measurements are time consuming and prone to variability. We aimed to assess whether fully automated artificial intelligence (AI) calculation of LVEF and GLS provide similar estimates and can identify abnormalities in agreement with conventional manual methods, in patients with pre-clinical and clinical CA.

METHODS

We identified 51 patients (age 80 ± 10 years, 53% male) with confirmed CA according to guidelines, who underwent echocardiography before and/or at the time of CA diagnosis (median (IQR) time between observations 3.87 (1.93, 5.44 years). LVEF and GLS were quantified from the apical 2- and 4-chamber views using both manual and fully automated methods (EchoGo Core 2.0, Ultromics). Inter-technique agreement was assessed using linear regression and Bland-Altman analyses and two-way ANOVA. The diagnostic accuracy and time for detecting abnormalities (defined as LVEF ≤ 50% and GLS ≥ -15.1%, respectively) using AI was assessed by comparisons to manual measurements as a reference.

RESULTS

There were no significant differences in manual and automated LVEF and GLS values in either pre-CA (p = .791 and p = .105, respectively) or at diagnosis (p = .463 and p = .722). The two methods showed strong correlation on both the pre-CA (r = .78 and r = .83) and CA echoes (r = .74 and r = .80) for LVEF and GLS, respectively. The sensitivity and specificity of AI-derived indices for detecting abnormal LVEF were 83% and 86%, respectively, in the pre-CA echo and 70% and 79% at CA diagnosis. The sensitivity and specificity of AI-derived indices for detecting abnormal GLS was 82% and 86% in the pre-CA echo and 100% and 67% at the time of CA diagnosis. There was no significant difference in the relationship between LVEF (p = .99) and GLS (p = .19) and time to abnormality between the two methods.

CONCLUSION

Fully automated AI-calculated LVEF and GLS are comparable to manual measurements in patients pre-CA and at the time of CA diagnosis. The widespread implementation of automated LVEF and GLS may allow for more rapid assessment in different disease states with comparable accuracy and reproducibility to manual methods.