COVID-19 Cough Classification using Machine Learning and Global Smartphone Recordings (2012.01926v2)

Abstract: We present a machine learning based COVID-19 cough classifier which can discriminate COVID-19 positive coughs from both COVID-19 negative and healthy coughs recorded on a smartphone. This type of screening is non-contact, easy to apply, and can reduce the workload in testing centres as well as limit transmission by recommending early self-isolation to those who have a cough suggestive of COVID-19. The datasets used in this study include subjects from all six continents and contain both forced and natural coughs, indicating that the approach is widely applicable. The publicly available Coswara dataset contains 92 COVID-19 positive and 1079 healthy subjects, while the second smaller dataset was collected mostly in South Africa and contains 18 COVID-19 positive and 26 COVID-19 negative subjects who have undergone a SARS-CoV laboratory test. Both datasets indicate that COVID-19 positive coughs are 15\%-20\% shorter than non-COVID coughs. Dataset skew was addressed by applying the synthetic minority oversampling technique (SMOTE). A leave-$p$-out cross-validation scheme was used to train and evaluate seven machine learning classifiers: LR, KNN, SVM, MLP, CNN, LSTM and Resnet50. Our results show that although all classifiers were able to identify COVID-19 coughs, the best performance was exhibited by the Resnet50 classifier, which was best able to discriminate between the COVID-19 positive and the healthy coughs with an area under the ROC curve (AUC) of 0.98. An LSTM classifier was best able to discriminate between the COVID-19 positive and COVID-19 negative coughs, with an AUC of 0.94 after selecting the best 13 features from a sequential forward selection (SFS). Since this type of cough audio classification is cost-effective and easy to deploy, it is potentially a useful and viable means of non-contact COVID-19 screening.

• The paper introduces a non-contact, cost-effective method by classifying cough sounds from global smartphone recordings for early COVID-19 detection.
• The study employs advanced techniques like SMOTE and nested cross-validation, with Resnet50 achieving an AUC of 0.98 along with high sensitivity and specificity.
• The findings support the integration of this ML-driven approach into health apps for rapid screening, paving the way for broader applications in respiratory diagnostics.

Analysis of COVID-19 Cough Classification using Machine Learning and Smartphone Recordings

This paper presents a paper on the application of machine learning techniques to classify cough sounds as indicative of COVID-19 infection, using recordings obtained from smartphones. The research aims to provide a non-contact, cost-effective, and easily deployable solution to reduce strain on healthcare systems by facilitating early self-isolation of symptomatic individuals.

Study Approach and Datasets

The researchers utilized two primary datasets in their paper. The Coswara dataset, publicly available, consists of recordings from 1171 subjects across six continents, including 92 COVID-19 positive cases and 1079 healthy individuals. The second, smaller Sarcos dataset, comprises mainly South African subjects, with 18 COVID-19 positive and 26 COVID-19 negative individuals. Notably, both datasets showed that COVID-19 specific coughs are approximately 15-20% shorter in duration than non-COVID coughs.

Methodology

To address the class imbalance inherent in these datasets, the Synthetic Minority Oversampling Technique (SMOTE) was employed. The machine learning algorithms evaluated included logistic regression (LR), k-nearest neighbor (KNN), support vector machine (SVM), multilayer perceptron (MLP), convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and a residual-based neural network (Resnet50).

The feature extraction process involved the computation of mel-frequency cepstral coefficients (MFCCs), alongside other audio features, preserving temporal patterns to enhance classification efficacy. A nested leave-$p$-out cross-validation scheme was used for robust evaluation and hyperparameter optimization, ensuring unbiased model assessment.

Key Results

Among the classifiers evaluated, Resnet50 achieved the highest area under the ROC curve (AUC) of 0.98, indicating superior performance in distinguishing COVID-19 positive coughs from healthy ones. The LSTM and CNN models also performed well, achieving AUCs of 0.94 and 0.95, respectively. On the Coswara dataset, the Resnet50 configuration demonstrated 93% sensitivity and 98% specificity, showing its effectiveness in real-world cough audio classification tasks.

When tested on the Sarcos dataset, models exhibited reduced performance, with the LSTM attaining the best AUC of 0.78. Further improvement was achieved using Sequential Forward Selection (SFS) for feature optimization, enhancing the AUC to 0.94 with selected features.

Implications and Future Directions

The findings suggest that automated cough classification using smartphone-recorded audio is a viable approach for COVID-19 screening, with potential applicability worldwide. From a practical standpoint, the method could be integrated into pre-existing health apps to provide users with preliminary assessments of their health status, especially during pandemic contexts, helping manage testing backlogs and reducing transmission rates.

Future work should consider enlarging the dataset for better generalization, integrating transfer learning to leverage larger pre-existing acoustic datasets, and refining the solution for a seamless user experience on widely available smartphone platforms. Such advancements would enhance the robustness and credibility of mobile-based diagnostic tools, potentially extending their application to a broader range of respiratory diseases.