Feature Extraction Model Based on Inception V3 to Distinguish Normal Heart Sound from Systolic Murmur

In this study, we propose a model for classifying normal and abnormal sounds by extracting characteristics from abnormal heart sounds in which normal and symbolic murmurs appear. Heart sound data obtained through an electronic stethoscope are converted into mel-spectrogram images. The pre-trained Inception V3 model that carries out fine-tuning uses the mel-spectrogram image as input. Convolutional layers of fine-tuning completed Inception V3 models were used as feature extractors. A point-binary correlation analysis technique was used to select effective features for classification from the features extracted through the feature extractor. A crystal coefficient value, which is the square of the correlation coefficient value, is used for an accurate comparison between the features. We used an artificial neural network as a classifier in this experiment. Fine-tuned Inception V3 has an average accuracy of 87.7%. When 5-fold class validation is advanced by selecting the top 30 characteristics with high crystal coefficient values, the accuracy is 97.5%. These results can greatly assist physicians trying to detect a systolic murmur.