A novel wheezing detection approach based on constrained non-negative matrix factorization

Monophonic and Polyphonic Wheezing Classification Based on Constrained Low-Rank Non-Negative Matrix Factorization

“…Non-negative Matrix Factorization (NMF) or standard NMF [ 61 , 62 ] is a decomposition technique that has attracted special attention in different fields of biomedical signal processing in the last few years [ 63 , 64 ]. Previous works showed the efficiency of the NMF approach at detecting [ 9 , 50 , 51 ] and improving the audio quality of wheezing [ 65 , 66 ]. In general terms, NMF can be defined as an unsupervised learning tool used for linear representation of non-negative two-Dimensional (2D) data where its main advantage is to reduce the dimensionality of a large amount of data in order to find hidden structures by means of part-based representation with non-negative patterns.…”

“…Some of the most popular cost functions are the Euclidean distance, the generalized Kullback–Leibler divergence, the Itakura–Saito divergence, and the Cauchy distribution [ 67 , 68 ]. In this paper, we propose to minimize the generalized Kullback–Leibler divergence (see Equation ( 4 )) because previous works [ 9 , 50 , 51 , 63 , 65 , 66 ] obtained promising results in biomedical signal processing since provides a scale-invariant factorization, that is low energy sound components of bear the same relative importance as high energy ones in the decomposition process. …”

“…The auscultation examination is considered a widely used method of detecting CRDs because it is a non-invasive, inexpensive, easy, comfortable, and fast method regardless of age [ 4 ]. However, the auscultation process has several limitations that reduce the reliability of the diagnosis: (i) high subjectivity conditioned by the physician’s training to recognize and interpret the sounds captured by the stethoscope [ 5 , 6 ]; (ii) the discrimination between adventitious sounds with similar characteristics, such as monophonic and polyphonic wheezing sounds, is a harder task to perform by means of auscultation [ 7 ]; and (iii) normal respiratory sounds and adventitious sounds (abnormal and indicative of a lung disorder) are simultaneously mixed in the time and frequency domain, complicating the physician’s analysis of the valuable clinical information contained in adventitious sounds [ 5 , 8 , 9 ]. Considering the above, a misdiagnosis is the main cause of the patient returning to the health center with a worsening of the disease that was not detected in the first medical examination performed by auscultation, so in recent years, it has become crucial to develop novel approaches to help physicians provide reliable diagnoses applied to lung disorders, with the implicit fact of reducing health care costs [ 10 , 11 ].…”

“…In general, the sounds generated during breathing can be classified into two main categories: normal respiratory sounds and adventitious sounds. Both sounds are mixed in the time-frequency domain as they are simultaneously generated by the same air flow through the bronchial tree of the lungs and also share part of the spectral bands in which they are active [ 5 , 9 ]. Normal respiratory sounds are represented by a wideband spectrum where most of the energy is concentrated in the frequency band of 60 Hz–1000 Hz [ 12 ].…”

“…For this reason, many research efforts have been applied in biomedical signal processing in order to develop reliable methods for early wheezing detection. In this sense, many wheezing detection algorithms, based on different approaches, can be found in the state-of-the-art literature: the Autoregressive (AR) model [ 33 ], auditory modeling [ 34 ], entropy [ 35 ], Neural Networks (NN) [ 36 , 37 ], wavelet transform [ 38 , 39 ], tonal index [ 40 , 41 ], Mel-Frequency Cepstral Coefficients (MFCCs) [ 42 , 43 ], Gaussian Mixture Models (GMMs) [ 44 , 45 ], spectral peaks identification [ 46 , 47 , 48 ], the Hidden Markov Model (HMM) [ 49 ], and recently, Non-negative Matrix Factorization (NMF) [ 9 , 50 , 51 ].…”