Enhancing machine learning algorithms in healthcare with electronic stethoscope

In this study, our aim is to classify respiratory sounds and diseases via audio and text data recorded by an electronic stethoscope using convolutional neural networks (CNNs), support vector machines (SVMs), k-nearest neigbor (k-NN) and Gaussian Bayes (GB) algorithms on a dataset that contains 17,930 lung sounds that were recorded from 1630 subjects. For classifying respiratory sounds, we employed; SVM, k-NN and GB with mel frequency cepstral coefficient (MFCC) features and CNN with 28x28 and 600x600 spectrogram images. We prepared 4 datasets to classify respiratory audio into: (1) healthy versus pathological; (2) rale, rhonchus, and normal sound; (3) singular respiratory sound type; and (4) audio type with all sound types classification. Accuracy results in percent were; (1) CNN 86 and 95, SVM 86, k-NN 85, GB 58, (2) CNN 80 and 93, SVM 80, k-NN 79, GB 42, (3) CNN 76 and 85, SVM 75, k-NN 76, GB 22 and (4) CNN 62 and 77, SVM 62, k-NN 61, GB 15 respectively. For classifying respiratory diseases, SVM, k-NN and GB algorithms were run on 6 datasets to classify patients into; (1) ill or healthy with text data, (2) ill or healthy with audio MFCC features, (3) ill or healthy with the text data and audio MFCC features, (4) 12 diseases with text data, (5) for 12 disease with audio MFCC features, (6) for 12 disease with the text data and audio MFCC features. Accuracy results in percent for SVM were 75, 88, 64, 73, 63, 70; for k-NN 95, 92, 92, 67, 64, 66; for GB 98, 91, 97, 58, 48, 58 respectively. To compare the electronic and traditional stethoscope, 3 chest physicians assessed 100 audio clips. We observed; good level consistency between physicians 2 and 3, average level consistency between physicians 1, 3 and 1, 2 via kappa statistic method.