Detection of Coronary Artery Disease Using PCG Signals for Wearable Devices: LSTM-Based Approach

It is acknowledged that coronary artery disease (CAD) is the most common type of heart disease. Continuous monitoring to identify the onset of CAD, even before symptoms are identified, can be helpful. This work attempts to efficiently identify CAD with high accuracy using wearable devices that enable continuous cardiovascular system monitoring. The proposed method utilizes single-channel PCG recordings for binary classification using Long Short-Term Memory. For this purpose, we have utilized the open-source dataset from the Physionet/Computing in Cardiology 2016 Challenge. We then provided a comparison of achieved results with state-of-the-art. The method that is being presented stands out because it does not require feature extraction or denoising, which are typically necessary for other similar works. We implemented five-fold cross-validation and achieved an average accuracy of 98.36±1.6 %. F1-score, precision, specificity, and sensitivity are 97.38±2.32 %, 94±3.97%, 96.3±2.32%, and 95.56±3%, respectively. For computational complexity measurement, we have used floating point operations, i.e., FLOPs (965784) and trainable parameters (1885). Additionally, we have developed an inference mechanism in a C++ program utilizing the weights and biases taken from the Keras model. The inference delay was calculated to be 9.664 seconds in a simulated Arm7 environment with Gem5.