Abstract
Atrial fibrillation (AF) is a prevalent cardiac condition characterized by irregular heart rhythms, which can lead to severe consequences. Early detection of AF is essential for effective treatment. This project proposes a transformer-based approach for AF detection using multivariate time series representation of short electrocardiogram (ECG) recordings. Our method converts ECG signals into a multivariate representation that captures the underlying patterns and dynamics of the signals to classify them as normal or AF. The performance of our approach was evaluated on two publicly available ECG datasets and compared with state-of-the art methods. We aim to develop an efficient and accurate AF detection system suitable for clinical use. Our proposed approach utilizes the Transformer Model, which is trained using both supervised and unsupervised pre-training. To evaluate the Transformer model's performance, we compared it to some baseline models such as the Long Short-Term Memory (LSTM) model, a Convolutional Neural Network (CNN) model, RESNET, and the XGBOOST model, which showed that the Transformer outperformed all those baselines in terms of precision, recall, and F1 score.