Abstract Body (Do not enter title and authors here): Background: Transthoracic echocardiography (TTE) is a frontline tool for the detection of cardiac amyloidosis (CA). Yet, accurate and reliable differentiation of CA from its phenotypic mimics remains challenging with conventional TTE techniques achieving sensitivities and specificities ranging around 70.9% and 79.1% respectively. Accordingly, we aimed to assess the accuracy of a novel artificial intelligence (AI) screening algorithm for TTE-based CA detection.
Methods: Utilizing a large, multicenter dataset, we trained a convolutional neural network to predict CA from TTE apical 4 chamber video clips (Ultromics, Oxford, UK). Training was conducted with a 3-D convolutional neural network using 1,349 individuals with adjudicated CA (45% light chain, 40% transthyretin wild type, 15% hereditary transthyretin) and 1,263 controls with increased wall thickness and/or heart failure with preserved ejection fraction and similar clinical characteristics of age, sex, race and history of arrhythmias (Panel A). Next, external validation was conducted globally across 18 sites including 597 CA and 2,025 non-CA controls (Panel A). Classification accuracy was assessed on the whole dataset, and subgroup analyses were performed separately using all patients referred for PyP (n=347, 23% CA) and individuals matched for age, sex and wall thickness (n=936, 50% CA).
Results: CA patients and controls were of similar age (median range; 72 to 78), sex (51 to 76% male), and racial makeup (44 to 63% White), with high proportions of hypertension, obesity, and diabetes in each subgroup. CA patients had greater left ventricular mass index than controls (median; 124 g/m2 vs. 105 g/m2, respectively), which was similar in the wall thickness-matched subgroup (~123 g/m2). Model discrimination and classification were excellent for the entire external validation dataset as well as each subgroup (Panel B). Performance was consistent between CA subtypes (sensitivity: AL=84.7% [77.7,90.6], ATTRwt=85.2% [80.8,89.6], ATTRv=86.4% [77.3,94.6]).
Conclusions: AI screening algorithms may provide effective detection of CA with TTE--irrespective of age, gender, and wall thickness--potentially facillitating timely initiation of diagnostic testing.