Screening COVID-19 by Swaasa AI platform using cough sounds: a cross-sectional study

The COVID-19 pandemic caused by SARS-CoV-2 has resulted in millions of deaths globally, necessitating robust screening techniques to control its spread. Current methods like RT-PCR are often expensive, time-consuming, and require technical expertise. This research addresses the need for a cost-effective and accessible screening method by developing and validating the Swaasa AI platform. This platform leverages the readily available and easily collectable data source of cough sounds, a common symptom of COVID-19, to identify potential cases. The unique acoustic signature of COVID-19 coughs is analyzed using machine learning algorithms to differentiate between COVID-19 positive and negative individuals. The study aims to demonstrate the effectiveness of Swaasa as a preliminary screening tool, potentially reducing the burden on existing diagnostic methods and facilitating early intervention. The high prevalence of the disease and limitations of currently available techniques highlight the importance of this research in developing a more accessible and efficient screening methodology.

Existing literature highlights the potential of AI in analyzing cough sounds for disease diagnosis. Studies have demonstrated the effectiveness of AI models, particularly CNNs and FFANNs, in analyzing various respiratory conditions using audio data. However, challenges remain in acquiring large, high-quality cough datasets from diverse populations and validating the performance of AI models in real-world clinical settings. The lack of readily available and clinically validated tools for COVID-19 screening using cough analysis underscores the novelty and importance of this research. The researchers also note previous research showing the successful application of cough sound analysis in diagnosing diseases like tuberculosis, and the potential for application to COVID-19 detection. Several studies are mentioned which used various machine learning techniques with varying degrees of success, highlighting the need for robust, clinically validated models.

The study involved three phases: model development and training, clinical validation, and external validation (pilot testing). **Sample Size Estimation:** The sample size was calculated using a formula that considered factors such as the desired level of confidence, anticipated prevalence of COVID-19, and the desired precision. A total of 1052 participants (62% controls) were included in the study. The control group included healthy individuals and those with respiratory symptoms but negative for COVID-19 by RT-PCR. **Data Collection:** Cough data was collected from COVID-19 suspects at Andhra Medical College using smartphones. Participants were instructed to cough for 10 seconds, with the recording device held 4-8 inches from their mouth. A noise reduction algorithm was applied to filter out background noise. After cough sample collection, RT-PCR was performed as a reference standard. Strict infection control measures were followed. **Model Development and Training:** Two parallel models, a CNN and an FFANN, were trained using data from 252 COVID-19 positive and 390 COVID-19 negative subjects. Feature extraction included Mel Frequency Cepstral Coefficients (MFCCs), spectral features, chroma features, contrast features, tonnetz features, zero-crossing rate (ZCR), energy, skewness, and kurtosis, resulting in 209 features that were reduced to 170 via correlation-based feature selection. The final output layers of both models were combined to improve prediction accuracy. **Clinical Validation:** The trained model was tested on 233 subjects, comparing the model's predictions to RT-PCR results. **External Validation (Pilot Phase):** The model was further validated on 183 presumptive COVID-19 cases at a peripheral healthcare center, again comparing the model's predictions with RT-PCR results. **Statistical Analysis:** Model performance was evaluated using metrics such as accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), Area Under the ROC Curve (AUC), and the confusion matrix. The Clopper-Pearson method was used to calculate 95% confidence intervals for these parameters. LIME (Local Interpretable Model-agnostic Explanations) was used to interpret the model's predictions and analyze cough sound characteristics associated with COVID-19.

**Model Derivation Phase:** The model achieved 96% accuracy on the test dataset and an AUC of 0.965. When tested on crowdsourced data, it achieved 86% accuracy and an AUC of 0.855. Analysis of features indicated that high spectral content was a distinguishing feature of COVID-19 coughs. LIME analysis showed unique frequency distributions for various respiratory diseases, including COVID-19, highlighting its unique acoustic signature. **Model Validation Phase:** In the clinical validation phase with 234 subjects, the model achieved 75.54% accuracy, 95.45% sensitivity, and 73.46% specificity. The AUC was 0.75. **Model Pilot Phase:** In the pilot testing phase with 183 subjects, the model identified 82 subjects as likely having SARS-CoV-2, and 58 of those were truly COVID-19 positive, resulting in a PPV of 70.73%.

The Swaasa AI platform demonstrated promising results as a cost-effective and accessible screening tool for COVID-19. The high sensitivity (95.45% in the validation phase) indicates the model's ability to correctly identify most COVID-19 positive cases. The use of a combined CNN and FFANN approach, combined with a large, clinically acquired dataset, contributes to the model's robustness. The study's findings align with previous research demonstrating the potential of AI in cough analysis for disease diagnosis. The platform's potential to improve efficiency and reduce the cost associated with COVID-19 screening is significant, particularly in resource-constrained settings. However, the relatively lower specificity (73.46%) suggests a need for further refinement to minimize false positives. The pilot phase results in a real primary care setting further demonstrate effectiveness as a screening and triaging tool.

The Swaasa AI platform presents a valuable tool for preliminary COVID-19 screening, offering a cost-effective and accessible alternative to existing methods. The platform's high sensitivity makes it effective in identifying high-risk patients. While further validation with larger, more diverse populations is needed to enhance accuracy and generalizability, the current findings suggest significant potential for improving COVID-19 screening efficiency and accessibility. Future research should focus on improving specificity, exploring the use of Swaasa in various settings, and investigating its performance with emerging COVID-19 variants.

The study's limitations include the relatively small sample size in the clinical validation and pilot phases, although still larger than many previously reported studies. The study's geographic location (primarily in India) may limit the generalizability of the findings to other populations with potentially different cough characteristics. The use of smartphone recordings could introduce variability due to recording conditions. Further research is needed to address these limitations and improve the robustness of the Swaasa platform.