PO-01-214 DEEP LEARNING EVALUATION OF ECHOCARDIOGRAMS TO IDENTIFY OCCULT ATRIAL FIBRILLATION

Abstract

Atrial fibrillation (AF) can often be missed by routine screening given its frequently paroxysmal and asymptomatic presentation. Deep learning algorithms can identify patients with paroxysmal AF from sinus electrocardiograms. Transthoracic echocardiograms (TTEs) may provide additional structural information complementary to ECGs that could also be used to help identify occult AF. We sought to determine whether deep learning evaluation of echocardiograms of patients in normal sinus rhythm could predict concurrent atrial fibrillation. We identified patients who had TTE performed between 6/2004 and 6/2021. Cases were defined as patients who had a TTE in normal sinus rhythm (NSR) with AF documented on ECG within 90 days before/after the TTE. Controls had a NSR TTE (by report and confirmed by lack of noted mitral A wave velocity) and no documented AF within 90 days by ECG or ICD diagnosis. We trained a video-based convolutional neural network using TTE parasternal long axis (PLAX) videos to predict concurrent AF. Model performance was compared to AF prediction by logistic regression using clinical variables (age, sex, heart failure, hypertension, cerebrovascular disease, peripheral arterial disease, diabetes, height, weight, smoker), CHADS2VASc score, and left atrial (LA) size on TTE. Model was trained on 57,681 TTEs of patients who were on average 66.2 years old (SD 16.9), 44.9% female, with a mean CHADS2VASc score of 3.1 (SD 2.1) and LA area of 19.7cm2 (SD 5.8). When tested on a held-out sample of 7,283 TTEs, the model predicted the occurrence of atrial fibrillation within 90 days of a sinus TTE with an AUC of 0.71 (0.69-0.73). Using the max accuracy threshold, the PPV was 0.20 and the NPV was 0.95. The model performed better than predicting AF using clinical risk factors (AUC 0.67), LA area (AUC 0.63), and CHADS2VASc (AUC 0.61). A deep learning model predicted the occurrence of atrial fibrillation from TTEs moderately well, better than clinical variables or LA size alone. This may open up additional opportunities for screening patients for occult AF.