Automatic Classification of Heart Sounds Using Long Short-Term Memory

Abstract

Heart diseases are serious and must be detected early using an auscultation examination. To explore and diagnose heart problems, several signal processing and machine learning approaches are used. From a Phonocardiogram (PCG) signal, the heart sound (HS) can be categorized into normal and abnormal. This paper presents an improvedcomputer-aidedtechniquefor classification of HS using long short-term memory (LSTM)deployed withdifferent time and frequency domain features, i.e., discrete wavelet transform (DWT) and Mel-frequency cepstral coefficients (MFCCs). The overall score, accuracy, sensitivity, and specificity of the LSTM classifier are calculated for the performance evaluation. With the proposed set of experimentsthe classification algorithm achieved a final score of 90.04% (Accuracy 90%, Sensitivity 92.30%, and Specificity 87.69%).