COVID-19 detection from optimized features of breathing audio signals using explainable ensemble machine learning

The automatic detection of COVID-19 using smartphone-recorded breathing signals in a ubiquitous and non-invasive way holds great promise. However, achieving accurate detection is challenging due to breathing signals' noisy and non-stationary nature, lack of distinguishable features, and imbalanced COVID/non-COVID data scenarios. This paper proposes an explainable ensemble learning-based framework for COVID-19 detection that extracts features from breathing signals through multiresolution analysis. First, we extract 165-dimensional features from the decomposed coefficients of a two-level discrete wavelet transformed (DWT) signal. From these, 27 optimized features are selected using the Recursive Feature Elimination with Cross-Validation (RFECV) technique. The level-2 DWT decomposed approximation coefficients retain frequencies in the 0–150 Hz range, aligning with human breathing frequencies. We utilize an ensemble model comprising decision trees, random forests, gradient boost, and XGBoost classifiers with a majority voting strategy for the detection task. A balanced and augmented dataset is prepared using the publicly available Coswara dataset. The results show that the ensemble approach improves accuracy compared to the individual models. Further, we explore the model's interpretability using Shapley additive explanations values, finding that the model places primary importance on features such as the RMS value, higher pitch of short-time Fourier transform, and higher frequency components of the Mel spectrogram, which align well with known COVID-related breathing characteristics. A comparison with related works demonstrates the effectiveness of our proposed feature extraction and ensemble framework, achieving an accuracy of 97.5 % and specificity of 95.24 %. These findings can potentially support smartphone-based COVID-19 detection applications using breathing signals.