Combining a recursive approach via non-negative matrix factorization and Gini index sparsity to improve reliable detection of wheezing sounds

Abstract

Auscultation constitutes a fast, non-invasive and low-cost tool widely used to diagnose respiratory diseases in most of the health centres. However, the acoustic training and expertise acquired by the physician is still crucial to provide a reliable diagnosis of the status of the lung. Each wrong diagnosis increases the risk to the health of patients and the costs associated with the treatment of the disease detected. A wheezing detection system can be useful to the physician to minimize the subjectivity of the interpretation of the breathing sounds, misdiagnoses due to stress and elucidating complex acoustic scenes (such as louder background noises). Highlight that the presence of wheeze sounds is one of the main indicators of respiratory disorders from airway obstructions. This work presents an expert and intelligent system to detect wheeze sounds based on a recursive algorithm that combines orthogonal non-negative matrix factorization (ONMF) and the sparsity descriptor Gini index. The recursive algorithm is composed of four stages. The first stage is based on ONMF modelling to factorize the spectral bases as dissimilar as possible. The second stage clusters the ONMF bases into two categories: wheezing and normal breath. The third stage proposes a novel stopping criterion that controls the loss of wheezing spectral content at the expense of removing normal breath content in the recursive algorithm. Finally, the fourth stage determines the patient’s condition to locate the temporal intervals in which wheeze sounds are active for unhealthy patients. Experimental results report that the proposed method: (i) provides the best detection performance compared to the recent state-of-the-art wheezing detection approaches, achieving the highest robustness in noisy environments; and (ii) reliably distinguishes the patient’s condition (healthy/unhealthy). The strengths of the proposed method are the following: (i) its unsupervised nature since it does not depend on any training stage to learn in advanced the sounds of interest (wheezing). This fact could make this method attractive to be used in clinical settings because wheezing sound databases are often unavailable; and (ii) the modelling of the spectral behaviour by means of a common feature, the sparsity, that represents the typically energy distributions shown by most of the wheeze and normal breath sounds.