Acoustic COVID-19 Detection Using Multiple Instance Learning.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Computational Biology

Background:

• The COVID-19 pandemic necessitated rapid and widespread diagnostic testing.
• Traditional testing methods can be time-consuming and costly, limiting accessibility.
• Machine learning (ML) offers a potential solution for low-cost, scalable audio-based diagnostics.

Purpose of the Study:

• To develop and evaluate an ML-based diagnostic tool for COVID-19 detection using audio recordings.
• To establish comparability between ML algorithms through the DiCOVA challenge.
• To improve diagnostic accuracy by incorporating diverse audio modalities and patient metadata.

Main Methods:

• Utilized the Coswara dataset comprising cough, speech, breath, and vowel phonation recordings.
• Employed a base ML model pre-trained on short audio intervals.
• Implemented a Multiple Instance Learning (MIL) model with self-attention for analyzing longer recordings.
• Applied linear regression and other fusion methods to combine predictions from different sound modalities for the DiCOVA challenge's fusion category.

Main Results:

• The MIL approach enhanced model generalizability, achieving an AUC ROC score of 86.6% in the fusion category.
• Incorporating sustained vowel phonation and patient metadata significantly improved performance.
• The final model achieved a diagnostic accuracy score of 92.2%.

Conclusions:

• Audio-based ML analysis is a promising, cost-effective method for COVID-19 screening.
• The MIL approach effectively leverages longer audio recordings for improved diagnostic accuracy.
• Integrating diverse data sources, including novel sound modalities and metadata, further enhances ML diagnostic performance.