CovNet

CovNet applies transfer learning to detect COVID-19 from cough sounds for non-invasive digital diagnostics.

Key Features:

• Transfer Learning Framework: Employs a dual-strategy transfer learning framework comprising a parameter transferring strategy and an embedding incorporation strategy.
• Utilization of Large-Scale Datasets: Transfers knowledge from the large FluSense multi-sound dataset to improve performance on smaller COVID-19 cough datasets such as COUGHVID.
• Application Across Multiple Datasets: Validated across COUGHVID, the COVID-19 cough sub-challenge (CCS) database from the INTERSPEECH Computational Paralinguistics Challenge (ComParE), and the DiCOVA Track-1 database.
• Improved Diagnostic Performance: Reports an absolute improvement of 3.57% in ROC AUC on DiCOVA Track-1 validation and a 1.73% increase in unweighted average recall (UAR) on the ComParE CCS test versus baseline.

Scientific Applications:

• Digital diagnostics: Detects COVID-19 from respiratory cough sounds to support early detection and monitoring.
• Non-invasive screening: Provides an audio-based method to distinguish COVID-19 positive and negative individuals in cough sound datasets such as COUGHVID.

Methodology:

Implements four simple convolutional neural networks (CNNs) within a transfer learning architecture, applying parameter transferring and embedding incorporation strategies and leveraging FluSense as source data to adapt models to COVID-19 cough datasets (COUGHVID, ComParE CCS, DiCOVA Track-1).