Authors: Sandeep B. Sangle, Chandrakant J. Gaikwad

Journal: Circuits, Systems, and Signal Processing (Springer Nature)

Overview

The COVID-19 pandemic has underscored the need for early and efficient detection methods to curb the spread of the virus. Traditional methods rely on molecular tests, which, while effective, are resource-intensive and not always conducive to rapid screening. With the virus exhibiting continuous mutation, alternative approaches have been explored to support early-stage COVID-19 detection. This study introduces a novel framework for identifying COVID-19 through respiratory sounds, specifically vocalizations, sneezing, and breathing sounds, by leveraging higher-order statistics (HOS) and Linear Frequency Cepstral Coefficients (LFCC) combined with deep learning techniques.

Objectives

Develop a Non-Invasive COVID-19 Detection Approach: Create an accurate, rapid screening method based on sound signal analysis to identify COVID-19 in patients.

Utilize Advanced Audio Signal Processing Techniques: Apply HOS and LFCC to capture unique audio characteristics associated with COVID-19, providing an alternative to traditional MFCC-based models.

Implement and Benchmark CNN and ResNet-50 Models: Assess the performance of convolutional neural network (CNN) and ResNet-50 architectures on COVID-19 sound signal classification.

Methodology

The research employed higher-order statistical (HOS) characteristics, specifically accumulated bispectrum, to analyze COVID-19 sound signals in contrast to the widely used Mel Frequency Cepstral Coefficients (MFCC). Key methodological components include:

HOS-Based LFCC Extraction: This approach extracts LFCC features from respiratory sounds, capitalizing on HOS to highlight non-linear and intricate characteristics in the audio signal, which are often linked to COVID-19’s unique respiratory signatures.

CNN and ResNet-50 Models for Classification: The study uses CNN and ResNet-50, two deep learning architectures well-suited for feature extraction and classification tasks. CNN models excel in capturing spatial dependencies in sound data, while ResNet-50’s residual connections offer depth and robustness, which are advantageous for distinguishing subtle COVID-19 audio cues.

Novel Detection Approaches: Three distinct detection methodologies are proposed to capture unique COVID-19 signal traits effectively. Each method emphasizes different aspects of the audio signal, enhancing the model's ability to detect COVID-19 sounds accurately.

Key Findings

The study presented several findings regarding the effectiveness of HOS-based LFCC in COVID-19 detection:

Enhanced Detection Accuracy: The models incorporating HOS-based LFCC features showed a significant improvement in accuracy compared to traditional MFCC-based methods. This enhancement suggests that LFCC, supported by HOS, captures COVID-19-specific acoustic features that MFCC may overlook.

Model Performance with CNN and ResNet-50: Both CNN and ResNet-50 models achieved high performance, but ResNet-50, due to its deeper architecture, demonstrated a slight edge in accuracy and stability. The use of accumulated bispectrum enabled the models to process complex respiratory signal patterns, yielding improved classification accuracy.

Reduced False Positives: The proposed methods resulted in fewer false positives, a crucial factor in diagnostic applications, ensuring that individuals are not misdiagnosed based on respiratory sound signals alone.

Practical Implications and Future Research

This work demonstrates the potential of using audio-based diagnostic tools for COVID-19 detection. Practical implications and future directions include:

Potential for Non-Invasive Screening: Sound-based COVID-19 detection offers a non-invasive, cost-effective alternative for early-stage screening, particularly in high-density or resource-limited settings where traditional testing may not be feasible.

Scalability in Mobile Health Applications: Integrating these sound classification models into mobile applications could provide rapid COVID-19 detection, making it accessible to the broader population. Such applications could alert users to seek further testing if COVID-19-like audio signals are detected.

Future Research on Other Respiratory Conditions: The successful application of HOS-based LFCC in COVID-19 detection highlights its potential for other respiratory diseases, such as pneumonia or tuberculosis, which also alter respiratory sounds in unique ways. Expanding this research to a broader spectrum of respiratory conditions could enhance public health diagnostics.

Conclusion

This study demonstrates the efficacy of using HOS-based LFCC features and deep learning models to detect COVID-19 from respiratory sound signals. By focusing on the unique audio characteristics of COVID-19, the proposed methodologies achieved substantial improvements in detection accuracy and reliability. As a non-invasive and scalable approach, sound-based COVID-19 detection can serve as a supplementary tool to existing testing methods, providing accessible early-stage diagnostics that can be widely implemented through mobile health platforms. This research paves the way for broader applications in audio-based disease detection, advancing AI's role in public health diagnostics and preventive care.