PCG Heart Sounds Quality Classification Using Neural Networks and SMOTE Tomek Links for the Think Health Project

Huisa C.M.; Elvis Supo C.; Edward Figueroa T.; Rendulich J.; Sulla-Espinoza E. · Lecture Notes in Networks and Systems (2023) · DOI: 10.1007/978-981-19-7615-5_65

Resumen

Cardiac PCG signal recordings are an important part of cardiology teleconsultations. The main problems related to high-quality recordings are due to less experienced healthcare personnel taking lower-quality samples and ambient noise, and these scenarios can lead to errors in diagnosis by the physician and PCG heart sound classification algorithms. Given this problem, machine learning algorithms were proposed for quality classification of PCG recordings that aid in accurate diagnosis and faster care. One difficulty in the application of these algorithms is the problems related to class imbalance, which is very common in medical applications that affect model performance. In this study, an artificial neural network (ANN) classifier with a SMOTE Tomek Links class imbalance method is used. Public databases containing 7893 recordings with ten features of each PCG signal are used. We use two types of labeling which have different levels of imbalance. The use of SMOTE Tomek Links in combination with neural networks showed better performance compared to SVM classifier. For future work, we intend to perform laboratory tests in remote areas applying the proposed algorithm and we also intend to use the concept of mel spectrograms and convolutional networks for the classification of heart sounds.

Instituciones: Universidad Nacional de San Agustín de Arequipa