Analysis of cough sound measurements including COVID-19 positive cases: A machine learning characterization

Abstract

Remote monitoring and measurement are valuable tools for medical applications and they are particularly important in the context of pandemic outbreaks, like the current COVID-19. This paper presents an analysis of sound measurements of cough events from the point of view of their predictive content with respect to identification of different types of cough, including positive COVID-19 cases. The data consisted of a collection of audio samples collected from different sources including dry, wet, whooping and COVID-19 coughs. Unsupervised and supervised machine learning techniques were used to reveal the underlying structure of the data, described by dissimilarity spaces constructed from pair-wise dynamic time warping measures derived from the original sound measurements. Intrinsic dimensionality, nonlinear mappings to low-dimensional spaces and visual cluster assessment techniques allowed a representation of the cough types distribution. Supervised classification techniques were used to obtain models identifying cough classes and high performance classifiers were obtained for most of them, including COVID-19. These results are preliminary and there is potential to improve, as they were obtained directly from a small dataset, without signal preprocessing (trimming, filtering, etc.), hyperparameter tuning, ensemble models, and class imbalance handling approaches.