Normal / Abnormal Heart Sound Recordings Classification Using Convolutional Neural Network

Abstract

As part of the PhysioNet / Computing in Cardiology Challenge 2016, this work focuses on automatic classification of normal / abnormal phonocardiogram (PCG) recording, with the aim of quickly identifying subjects that need further expert diagnosis. To improve the robustness of the classifiers by increasing the number of training samples, the recordings were windowed into 5 second segments and our classifiers were trained to classify these segments. Overall recording classification was then generated using a voting scheme from classification results of its segments. Our features include spectrograms and Melfrequency cepstrum coefficients. Our best submission result during the official phase (evaluated on a random 20% of the hidden test set) has a score of 0.813, with 0.735 sensitivity and 0.892 specificity. Two more submissions are still being evaluated.

Short-segment Heart Sound Classification Using an Ensemble of Deep Convolutional Neural Networks

“…Applications of DL to cardiac signals are introduced very recently [6][7][8]. CNNs have been used for normal/abnormal PCG classification using input features such as spectrogram and Mel-frequency cepstrum coefficients (MFCCs) in [9] on 5-second windowed segments, and MFCC heatmaps of 3-second segments in [10]. Tschannen et al [11] combined a wavelet-based deep CNN feature extractor with support vector machine (SVM) for heart-sound classification.…”

Comparative Analysis of Rescaled Range Results of Normal and Abnormal Heart Sound Recordings

“…This was done to have a uniform recording length that is consistent with all recordings studied. Based on previous literature Nilanon et al [7], medical practitioners concur with this assumption. Another assumption is that the behaviour of a particular recording would remain the same for a small change in the time interval.…”

Feature extraction and classification of heart sound using 1D convolutional neural networks

“…The minimum duration of heart sound acquisition was 5 s. In this study, we divided the signals into pieces of 5 s, and the most common smoothing windows were selected. First, we trained the classifier for each segment extracted from the records to classify normal and abnormal heart sound signals, rather than extract from the entire record [20]. The classifier not only extended the sample size of the entire dataset but also reduced the overfitting in network training by expanding the sample size of the dataset and then the merging the lengths of all the samples.…”