Transformer-based CNNs: Mining Temporal Context Information for Multi-sound COVID-19 Diagnosis

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Bioacoustics

Background:

• Early COVID-19 screening is crucial for transmission control.
• Computer audition presents a rapid, cost-effective, and eco-friendly diagnostic approach.
• Respiratory and speech sounds contain valuable clinical information for COVID-19 detection.

Purpose of the Study:

• To develop and evaluate deep neural networks for COVID-19 detection using respiratory and speech sounds.
• To improve diagnostic performance by assembling multiple models trained on different sound types.
• To leverage advanced deep learning architectures, including CNNs and transformers, for enhanced feature extraction and temporal context mining.

Main Methods:

• Training three deep neural networks on breathing, counting, and vowel sounds.
• Employing Convolutional Neural Networks (CNNs) for spatial feature extraction from log Mel spectrograms.
• Utilizing a multi-head attention mechanism within a transformer to capture temporal dependencies.

Main Results:

• Transformer-based CNNs demonstrated effective COVID-19 detection on the DiCOVA Track-2 database.
• Achieved an Area Under the Curve (AUC) of 70.0%.
• Outperformed traditional CNNs and hybrid CNN-Recurrent Neural Networks (RNNs).

Conclusions:

• Deep learning models, particularly transformer-based CNNs, show significant potential for COVID-19 detection via computer audition.
• Assembling models trained on diverse sound types enhances diagnostic accuracy.
• This approach offers a promising avenue for accessible and efficient COVID-19 screening.