Heart Murmur Prediction with Machine Learning

Heart murmurs can be an effective indicator of medical problems with a person’s heart. If machine learning algorithms can be designed to detect heart murmurs in audio files containing heartbeat sounds, it can aid medical professionals in detecting medical problems with a person’s heart. This can be especially helpful in poorer countries where trained heart specialists are not plentiful. There exist datasets with labeled audio heartbeat sounds, some of which have heart murmurs. The Dangerous Heartbeat Dataset is an excellent example of labeled heartbeat-related data. This paper models this dataset using three machine learning models and one deep learning model, i.e., Random Forest, XGBoost, CatBoost, and Convolutional Neural Network (CNN). We investigate two types of classifiers for this dataset: one where the background laboratory noisy samples are included and one where they were excluded during the training and testing phases. The features in the model are audio features extracted using the Librosa Python package. For the CNN model, the Mel Spectrogram 2-D matrix was used as the input. Model accuracies with the three machine learning models were as high as 77%, while those with the CNN model were as high as 71%.