Cardiovascular Anomaly Detection Using Deep Learning Techniques

Cardiovascular diseases (CVD) refer to a group of health conditions that affect the heart and blood vessels. This can also include arterial damage in organs such as the kidneys, the heart, the eyes, and the brain. Electrocardiograms (ECGs) are a quick, safe, and non-intrusive cardiac exploration, to check for heart rate, heart rhythm, and signs of potential heart disease. The interpretation of ECGs can be vital in determining the condition of the human body, and it is important to obtain accurate results. Deep learning techniques are being used in this work to automatically analyze ECG recordings. We regenerated certain datasets from the Physionet data bank, such as the MITBIH dataset, and others from a cardiology challenge in 2020 by performing some transformations and adaptations. We have implemented different models to detect ECG anomalies using both single and multiple-lead ECG datasets. CNN family model has the highest detection performance when trained on a single lead ECG dataset. Its performance decreased significantly when anomaly classes and lead numbers increased. In fact, accuracy passed from 0.96 to 0.60 whereas the F1 score passed from 0.98 to 0.55 when trained on a 12-lead dataset with 27 classes of anomalies. Given the regular time series pattern of the ECG signal, we propose a combination of a CNN-LSTM model for classification. This model achieved 0.668 accuracy. Combining all models into one ensemble learning model increased significantly the detection accuracy on the 12-lead ECG dataset to reach 0.82. Our combined architecture has proven to achieve state-of-the-art accuracy in ECG anomaly detection and could help health professionals better manage CVD.