Evaluation of Respiratory Sounds Using Image-Based Approaches for Health Measurement Applications

Area of Science:

• Medical technology
• Signal processing
• Machine learning

Background:

• Audio sensing at home can detect worsening respiratory conditions.
• Classifying respiratory events is crucial for timely medical intervention.

Purpose of the Study:

• To evaluate image-based transfer learning for classifying respiratory events from audio visualizations.
• To compare the performance of five different audio visualizations for respiratory event classification.

Main Methods:

• Applied transfer learning using a pre-trained AlexNet model.
• Utilized five audio visualizations: linear spectrogram, logarithmic spectrogram, Mel-spectrogram, wavelet scalograms, and aggregate images.
• Evaluated three classification tasks: C1 (wet, dry, whooping cough, restricted breathing), C2 (wet vs. dry cough), and C3 (cough vs. restricted breathing).

Main Results:

• Aggregate image-based classifiers achieved high accuracy across all tasks (C1: 0.88, C2: 0.88, C3: 0.91).
• The Mel-spectrogram method showed the highest testing accuracy (0.94) specifically for classifying wet vs. dry coughs (C2).

Conclusions:

• Image-based transfer learning with aggregate audio visualizations shows promise for classifying respiratory events.
• This approach can provide valuable, otherwise unavailable, data to healthcare professionals for remote patient monitoring.