Atrial Fibrillation Diagnosis from ECG Signal Using Statistical and Time-Frequency Attributes-based LSTM Network

Atrial Fibrillation (AF) is currently the most prevalent heart arrhythmia. The number of people affected by AF is rising daily. Recognizing AF in its early stages can help to reduce the causes of heart failure. Manually analyzing long-term electrocardiogram (ECG) signals presents challenges for medical professionals. An automatic computer-aided recognition system is required to analyze long-length ECG signals. This work proposes an effective machine learning-based system to predict subtle abnormalities in ECG signals. The proposed method is a statistical attributes-based LSTM network that will classify normal and AF-affected ECG signals. Noise removal, QRS detection, feature extraction, and classification have been performed using feasible algorithms. Beat-by-beat counting approach is used to extract the statistical features and give these as LSTM network input. The time-frequency (TF) moment features are also been extracted. Both types of signals are classified also using a TF-based LSTM network. Our proposed method is compared with TF based LSTM network. The results indicate that our proposed method is an effective, robust system and comparable to TF moments-based LSTM.