Anomaly Detection of ECG Signals using Attention-derived Convolutional Long Short-term Memory

An electrocardiogram (ECG) is a diagnostic procedure that measures electrical activity in the heart and helps doctors detect any irregularities. Early detection of anomalies in ECG signals is critical to offer patients the best treatment options. An electrocardiogram, or ECG, is time-series data generated by the electrical activity of the heart. Applying machine learning techniques to anomaly detection in these time series data has been the focus of recent research initiatives. These techniques necessitate feature extraction, which introduces complications such as the discrepancy in the features extracted and variability in the ECG parameters. Avoiding the need for human-crafted features from ECG signals is possible with the help of deep learning techniques, such as approaches built on Convolution Neural Networks (CNN). Given the volume of the ECG datasets and the sophistication of recent CNN technologies, However, CNN's performance in detecting anomalies in ECG was hindered by its inability to capture ECG signals of varying lengths. To tackle this issue, we presented C2LSTM, a deep learning model consisting of CNN, LSTM cells, and an attention mechanism. This model has more flexibility in terms of input length and uses half as many parameters compared to the CNN models. We conducted a considerable number of tests on the MIT-BIH dataset in order to give a thorough evaluation of the C2LSTM model against the AAMI standards; the C2LSTM achieved a sensitivity of 96%, positive predictivity of 99%, and F1-score of 96%.