The Development of Medical Devices for COPD Exacerbation Development Monitoring

Keywords

For patients with COPD: To improve self-efficacy for patients with COPD, they need to objectively judge their symptoms before acute exacerbation occurs.
For doctors to help COPD detect acute-exacerbations, the service gap is shown below

We Develop abnormal lung sound detection system and then build up a model for early-detection of acute-exacerbations

For Abnormal Lung Sound Detection:
- Audio Source: We collaborated with medical center in Taiwan to collect lung sounds from patients with COPD
- Audio Annotation: Normal, Wheeze, Rhonchi, Crackle
- Audio Preprocessing: (for noisy background) Pre-emphasis with band-pass filter
- Stage 1 Modeling:
  - Audios were transformed into spectrograms.
  - Use Convolutional Neural Network (CNN) for abnormal dectection for each spectrogram
- Stage 2 Modeling:
  - Audios were transformed into spectrograms.
  - Use Recurrent Neural Network for abnormal detection for the spectrograms from each audio
For Early Detection of Acute Exacerbation:
- Use the results from abnormal lung sound detection model and add other data, including respiratory pattern and SpO2
- Build up a decision tree model to determine the occurrence of acute exacerbation

Abnormal Lung Sound Detection
- Training Data: ~6.5k, Testing Data: ~1.3k
- Current Model Performance: Overall, 4-class classification accuracy is 87.13%. Precisions for four classes are all above 80%. Recalls for Normal and Wheeze are above 85%. Recalls for Rhonchi and Crackle are near 80%.
- The visuals for our model
Eary Detection of Acute Exacerbation:
- With respiratory pattern features, SpO2 and class probabilities output from abnormal lung sound detection model, we trained a decision tree classifier and got 80% accuracy on test data.
All the experiments above are still ongoing. More data will be included for analysis.

After more abnormal lung sounds (Rhonchi & Crackle) were added to train the model, the precision / recall for the Rhonchi & Crackle were increased. Howevere, adding more data didn’t work after the precision / recall of Crackle reached 80%. It might be related to the nature of Crackle - discontinous, intermittent and brief sounds. Currently, I’m experimenting on how bidirectional RNN could help detect Crackle better.