Temporal autoencoder architectures with attention for ECG anomaly detection

Anomaly detection is a crucial step in any diagnostic procedure. With the advent of continuous monitoring devices, it is inevitable to use technological assistance for the same. Many methods, including autoencoders, have been proposed for anomaly detection in time series ECG data. The attention mechanism dynamically highlights the relevant portion of the input data and provides the decoder with the information from every encoder hidden state in its temporal vicinity. This work proposes a performance enhancement of autoencoders in identifying an ECG anomaly with the help of attention. A comparison of different autoencoder models, LSTM and hybrid, with and without attention to detect an anomaly, is proposed in this work. The comparison of the different models in terms of precision, recall, F1-score, false-positive rate (FPR), false-negative rate (FNR) and area under the ROC curve (AUC) are specified. The obtained results indicate that attention helps to enhance the autoencoder's performance.