Abstract 9495: Deep Learning Algorithm for Predicting Atrial Fibrillation Based on Chest Radiography

Abstract

Introduction: Atrial fibrillation (AF) is a common risk factor for stroke and heart failure, with gradually increasing prevalence. AF is usually diagnosed on the basis of electrocardiography. Chest radiography is commonly performed as a screening test among patients with cardiac diseases but cannot be used to detect AF because of its unclear radiographical findings.Hypothesis: We hypothesize that deep learning methods, particularly convolutional neural networks (CNN), can be used to detect AF on chest radiographs. Methods: Chest radiographs used for training were obtained from Yongin Severance Hospital, South Korea. A total of 11,044 images acquired from patients with normal sinus rhythm or AF were used, whereas images from patients with other rhythms, such as paced rhythm or left bundle branch block, were excluded. The training, validation, and test datasets were split 8:1:1, and Resnet was applied as a model architecture. The accuracy, area under the receiver operating characteristic (ROC) curve, area under the precision-recall curve (PRC), precision, and recall were calculated. Gradient-weighted class activation mapping (Grad-CAM) was used to determine the area focused on by the model to predict AF. Results: AF was detected from chest radiographs with an accuracy, AUC, and PRC of 0.95, 0.81 and 0.39 in the validation set, respectively, and 0.94, 0.76, and 0.35 in the test set, respectively (Figure 1-A, B). Grad-CAM showed that the highest predictive value images from each dataset focused on the heart and its border, while the lowest predictive value images focused on the ribs (Figure 1-C, D, E, F). Conclusions: Deep learning algorithms can be used to detect AF on chest radiographs, which can be used as a screening tool for AF patients.