Development of a non-invasive Covid-19 detection framework using explainable AI and data augmentation1

This paper investigates the potential of COVID-19 detection using cough, breathing, and voice patterns. Speech-based features, such as MFCC, zero crossing rate, spectral centroid, spectral bandwidth, and chroma STFT are extracted from audio recordings and evaluated for their effectiveness in identifying COVID-19 cases from Coswara dataset. The explainable AI SHAP tool is employed which identified MFCC, zero crossing rate, and spectral bandwidth as the most influential features. Data augmentation techniques like random sampling, SMOTE, Tomek, and Edited Nearest Neighbours (ENN), are applied to improve the performance of various machine learning models used viz. Naive Bayes, K-nearest neighbours, support vector machines, XGBoost, and Random Forest. Selecting the top 20 features achieves an accuracy of 73%, a precision of 74%, a recall of 94%, and an F1-score of 83% using the Random Forest model with the Tomek sampling technique. These findings demonstrate that a carefully selected subset of features can achieve comparable performance to the entire feature set while maintaining a high recall rate. The success of the Tomek undersampling technique highlights the ability of model to handle sparse clinical data and predict COVID-19 and associated diseases using speech-based features.