Audio cough analysis by parametric modelling of weighted spectrograms to interpret the output of convolutional neural networks

Abstract

This study explores the feasibility of employing eXplainable Artificial Intelligence XAI methodologies for the analysis of cough patterns in respiratory diseases. A cohort of 20 adult patients, all presenting persistent cough as a symptom of respiratory disease, was monitored for 24 hours using a smartphone. The audio signals underwent frequency domain transformation to yield 1-second spectrograms, subsequently processed by a CNN to detect cough events. Quantitative analysis of spectrogram regions relevant for cough detection highlighted by occlusion maps, revealed significant differences between patient groups. Notably, distinctions were observed between the Chronic Obstructive Pulmonary Disease (COPD) patient group and groups with other respiratory pathologies, both chronic and non-chronic. In conclusion, interpretability analysis methods applied to neural networks offer insights into cough-related distinctions among patients with varying respiratory conditions.