Cochleogram-based adventitious sounds classification using convolutional neural networks

Abstract

Background:The World Health Organization (WHO) establishes as a top priority the early detection of respiratory diseases. This detection could be performed by means of recognizing the presence of acoustic bio-markers (adventitious sounds) from auscultation because it is still the main technique applied in any health center to assess the status of the respiratory system due to its non-invasive, low-cost, easy to apply, fast to diagnose and safe nature. Method:Despite the novel deep learning approaches applied in this biomedical field, there is a notable lack of research that rigorously focuses on different time–frequency representations to determine the most suitable transformation to feed data into Convolutional Neural Network (CNN) architectures. In this paper, we propose the use of the cochleogram, based on modeling the frequency selectivity of the human cochlea, as an improved time–frequency representation to optimize the learning process of a CNN model in the classification of respiratory adventitious sounds. Our proposal is evaluated using the largest and most challenging public database of respiratory sounds. Results:The cochleogram obtains the best binary classification results among the compared methods with an average accuracy of 85.1% in wheezes and 73.8% in crackles, and a competitive performance evaluating a multiclass classification scenario in comparison with other well-known state-of-the-art deep learning models. Conclusion:The cochleogram provides a suitable time–frequency representation since it is able to model respiratory adventitious content more accurately by means of non-uniform spectral resolution and due to its increased robustness to noise and acoustic changes. This fact implies a significant improvement in the learning process of CNN models applied in the classification of respiratory adventitious sounds.