Deep learning based cough detection camera using enhanced features.

Area of Science:

• Artificial Intelligence
• Acoustic Signal Processing
• Medical Diagnostics

Background:

• Coughing is a primary symptom of COVID-19, necessitating remote detection methods.
• Existing methods for cough detection lack precise localization capabilities.
• Deep learning offers potential for advanced acoustic analysis.

Purpose of the Study:

• To develop and validate a deep learning model for remote cough sound detection and localization.
• To integrate the model with a sound camera for real-time visualization of cough sources.
• To enhance acoustic feature extraction for improved cough classification.

Main Methods:

• A convolutional neural network (CNN) based binary classifier was developed, processing two-second acoustic features.
• Data augmentation techniques were employed to address class imbalance and simulate real-world noise.
• Spectrograms, mel-scaled spectrograms, and MFCCs were enhanced with velocity (V) and acceleration (A) maps.
• Simplified VGGNet, GoogLeNet, and ResNet architectures (V-net, G-net, R-net) were trained and evaluated.
• The best performing model, G-net with MFCC-V-A features (Spectroflow), achieved high F1 scores.

Main Results:

• The Spectroflow model (G-net with MFCC-V-A) achieved a test F1 score of 91.9% and 97.2% accuracy.
• The integrated sound camera system demonstrated a 90.0% F1 score and 96.0% accuracy in pilot tests.
• Real-time tracking of cough locations within camera images was successfully demonstrated.

Conclusions:

• The developed Spectroflow system provides an effective method for remote cough detection and localization.
• This technology has significant potential for public health surveillance and infection control.
• The combination of advanced acoustic features and deep learning architectures enhances cough detection performance.