Abstract Body (Do not enter title and authors here):
Introduction
Transthoracic echocardiography (TTE) is the standard-of-care imaging modality for management of heart failure (HF). Transthyretin amyloid cardiomyopathy (ATTR-CM) is an important cause of HF that can be difficult to identify and follow longitudinally by TTE, owing to imprecision of measurements and variability in interpretation. Analysis of standard TTE variables using Artificial Intelligence (AI) could improve precision and reproducibility while reducing interpretation time.

Aims
To assess performance of a previously validated, fully automated AI algorithm for TTE interpretation (Us2.ai) in patients with heart failure at risk for ATTR-CM. We hypothesized that AI measurements would be strongly correlated with human interpretation and identify differences between participants with and without ATTR-CM.

Methods
TTE images from participants in the prospective Screening for Cardiac Amyloidosis with Nuclear Imaging in Minority Populations (SCAN-MP) study (n = 586, 36 ATTR-CM cases) were analyzed independently by a human expert and an AI algorithm (Us2.ai). Intraclass correlation and Bland-Altman agreement between human and AI were assessed for clinically relevant TTE variables including left ventricular ejection fraction (LVEF), global longitudinal strain (GLS), maximal wall thickness (MWT), and mitral E/e’ ratio. Additionally, human and AI parameter distributions were compared between participants with and without ATTR-CM.

Results
Comparison of human and AI distributions demonstrated differences in participants with ATTR-CM for most variables, with similar magnitude and directionality (Table 1). Correlation between AI and human measurements was high for the Doppler measure E/e’ (ICC: 0.89, 95% CI: [0.86-0.90]) and moderate for LVEF [0.72 (0.27-0.86)], MWT [0.68 (0.63-0.73)], and GLS [0.64 (0.54-0.72)]. Similarly, Bland-Altman agreement was highest for E/e’, while LVEF, GLS, and MWT had relatively wider limits of agreement and greater bias. AI underestimated most parameters, including LVEF (bias AI -5.3%), MWT (bias AI -0.15cm), and E/e’ (bias AI -0.49), but overestimated GLS (bias AI +1.1%).

Conclusions
Fully automated TTE measurements were correlated with a human reference and reproduced differences between ATTR-CM cases and controls. AI TTE interpretation is a promising tool to democratize echocardiography for clinical care and research, and may be interchangeable with human readers for some measurements relevant to ATTR-CM.