COVID-19 Detection From Respiratory Sounds With Hierarchical Spectrogram Transformers.

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Respiratory Medicine

Background:

• Traditional respiratory assessments for diseases like COVID-19 require in-person visits.
• Remote monitoring using portable devices offers a viable alternative for preliminary disease screening.
• COVID-19 primarily impacts the lower respiratory tract, making respiratory sound analysis relevant.

Purpose of the Study:

• To develop a novel deep learning approach for distinguishing COVID-19 patients from healthy individuals using respiratory sounds.
• To evaluate the effectiveness of a hierarchical spectrogram transformer (HST) for analyzing cough and breathing sounds.

Main Methods:

• A novel hierarchical spectrogram transformer (HST) was developed, utilizing self-attention mechanisms on spectrogram representations of respiratory sounds.
• The HST model progressively increases window size to capture local and global contextual information.
• The approach was validated against conventional and deep-learning baselines using crowd-sourced, multi-national datasets.

Main Results:

• The proposed HST deep learning approach demonstrated superior performance compared to existing methods.
• The model achieved an area under the receiver operating characteristic curve (AUC) exceeding 90% in detecting COVID-19 cases.
• The study successfully distinguished COVID-19 patients from healthy controls based on audio recordings.

Conclusions:

• The novel deep learning approach using HST shows significant potential for accurate, non-invasive COVID-19 detection via respiratory sound analysis.
• This method supports remote monitoring capabilities, reducing the need for hospital visits.
• The findings highlight the efficacy of advanced AI techniques in diagnosing respiratory illnesses from audio data.