Unveiling the potential of phase portrait-based recurrence network: a revelation through lung sound analysis

Abstract

Abstract The work reported in this article introduces the novel concept of phase portrait-based recurrence network (RN) analysis in the digital auscultation of expiratory wheeze (ES) and vesicular (VS), with its potential revealed through machine learning techniques (MLTs). The time-series lung sound signals of ES and VS, subjected to power spectral density analysis, revealed information regarding the morphology of the respiratory tract responsible for the generation of signature frequency components. Having constructed the complex network using Pearson’s linear correlation coefficient ($P_{a,b}$) and employing the graph features for classification by principal component analysis (PCA), it is understood that the graph features thus obtained are incapable of classifying the two signals. Hence, a novel method, recurrence network, of constructing a network from the phase portrait of the time series is employed to deduce the network features. The MLTs, K-nearest neighbour (KNN) and PCA, are found to give better classification when RN topological features are used. When PCA separates the two signals with 84.7% total variance between the principal components, KNN yields 100% prediction accuracy. Thus, the study unveils the potential of RN over $P_{a,b}$-based complex network in classifying the lung sound signals ES and VS and thereby opening the possibility of employing the technique in digital auscultation, a best-suited one for the time of the widespread pandemic coronavirus disease-2019.