2D image-based Atrial Fibrillation Classification

Felipe Dias1, Nelson Samesima1, Adele Ribeiro1, Ramon Moreno1, Carlos Pastore1, Jose Krieger1, Marco Gutierrez2
1Heart Institute University of Sao Paulo Medical School, 2Heart Institute University of Sao Paulo


Atrial fibrillation (AF) is a common arrhythmia (0.5% prevalence worldwide) associated with increased risk for several cardiovascular derangements, including stroke. Routine automated AF detection using Electrocardiogram (ECG) is based on the analysis of unidimensional ECG signals and requires dedicated software for each type of device, limiting its broad use, especially with the fast incorporation of telemedicine into the health system. Here, we have implemented an AF classification method using the region of interest (ROI) corresponding to the long DII lead extracted automatically from the 12-lead ECG DICOM images. We designed a Convolutional Neural Network (CNN) to classify these ROIs between AF and non-AF using a dataset composed of 80,178 exams from 64,766 different patients. The dataset was collected retrospectively from the Picture Archiving and Communication System (PACS) of a public cardiovascular tertiary hospital and randomly divided into training, validation, and testing with the proportion of 60%, 20%, and 20%, respectively. The CNN receives as input the DII lead ROI image along with demographic data (age, sex, and ethnicity). The network architecture contains five blocks, each one with the following layers: conv2d, batch normalization, ReLU, conv2d, batch normalization, ReLU, and max pooling. The fully connected layer receives both the output of the last block and the demographic data. The binary cross-entropy was chosen as the cost function for training loss. Using this approach, we observed 94.3%, 98.9%, 99.1%, and 92.2% for sensitivity, specificity, AUC, and F1-score, respectively. These results suggest that the proposed methodology is competitive with other recent literature methods (F1-score<90%) when using one-dimensional ECG signals as input. Moreover, this approach has two major advantages: it allows easier integration into clinical practice in hospital settings since ECGs are usually stored in PACS as 2D images, and it allows classification of ECGs acquired from different devices, e.g., scanners and photographs.