Multi-Modal Point-of-Care Diagnostics for COVID-19 Based on Acoustics and Symptoms.

Area of Science:

• Biomedical Engineering
• Data Science
• Respiratory Medicine

Background:

• The COVID-19 pandemic necessitates improved diagnostic methods focusing on speed, cost, and safety.
• Acoustic bio-markers for respiratory diseases are gaining attention as a potential solution.

Purpose of the Study:

• To design and evaluate a COVID-19 diagnostic tool utilizing acoustic signals (cough, breathing, speech) and health symptoms.
• To explore the efficacy of multi-modal data integration for enhanced diagnostic performance.

Main Methods:

• Collected a multi-modal dataset including acoustic signals and health symptoms over twenty months via a web application.
• Investigated time-frequency features for acoustic data and binary features for symptoms.
• Employed machine learning models including logistic regression, support vector machines, and LSTM for acoustic analysis, and decision trees for symptom analysis.

Main Results:

• Achieved a 96.3% area-under-curve (AUC) through multi-modal integration, significantly outperforming individual modalities.
• Mel-spectrogram features demonstrated superior performance across different acoustic signal types.
• The diagnostic approach showed generalization capabilities with data from newer SARS-CoV-2 variants.

Conclusions:

• The proposed multi-modal approach offers a promising direction for accurate COVID-19 detection.
• The method demonstrates robustness and adaptability to emerging COVID-19 variants.