Respiratory Sound Classification for Remote Diagnosis

Abstract

With rapid development of electronic control technology in internet technology in modern medical industry and with the development of electronic stethoscope we are one step closer to the remote diagnosis as the hardware part of it is readily available but there are limited software system that can classify stethoscope sound. Diagnosis or classification requires recognizing patterns. If the quantity of input is huge, it becomes difficult to identify these patterns. The collected data is often a non linear data, the conventional models fail to identify patterns. These challenges can be tacle with the use of latest technologies like Machine learning. Till date, many approaches were successful and now, with the use of Neural networks the loss factor is close to nil In this paper machine learning algorithm is implemented on Mel- spectrogram images in the convolutional neural network (CNN). Since considering MFCC features to classify sounds is generally accepted classification method for audio, its scale is used to find patterns in Spectrogram of samples using CNN. Sound is classified in four classes: (1) healthy (2) Containing Wheezes (3) Containing Crackles and (4) Containing both. Accuracy results of the experiments were around 63%-65%.