Cardiovascular Anomaly Detection: Integration of Phonocardiogram Signals with Ensemble Learning and Deep Learning Techniques

Auscultation, i.e., listening to heart sounds using a stethoscope, is one of the oldest and most practiced diagnosis methods for doctors. However, the doctor's expertise and experience are critical to the success of this method. Phonocardiogram (PCG) signal is used to automate this sector and reduce the human factor. A PCG signal is a visual representation of the sounds generated by the heart. In this paper, machine learning and deep learning techniques are leveraged to train and test some open-access data and develop a model. Binary classification is done by algorithms like Random Forest, Logistic Regression, AdaBoost, Support Vector Machine, and Decision Tree. The best accuracy is found while stacking these machine learning models. The accuracy from ensemble learning is 94.60%. The same accuracy is found by using Artificial Neural Network (ANN). To back our binary classification model and to prove its universality, heart sounds are collected from the patients of a clinic in Chattogram, Bangladesh, and this dataset is used only for testing the previously formed ANN model. With testing the data using the binary classification technique, an accuracy of 94.74% is found. Multiclass classification of the publicly available dataset is also done and the best accuracy was found to be near 78% for the Random Forest Classifier.