Feature evaluation of accelerometry signals for cough detection

Area of Science:

• Biomedical Engineering
• Signal Processing
• Wearable Technology

Background:

• Cough detection is crucial for monitoring respiratory diseases like asthma and COPD.
• Current methods rely on audio analysis, raising privacy concerns due to always-on microphones.
• Machine learning and deep learning are commonly used for audio-based cough detection.

Purpose of the Study:

• To explore an alternative, privacy-preserving method for cough detection.
• To develop a non-contact system using accelerometers to differentiate coughs from other movements.
• To evaluate the efficacy of motion-based signals for continuous respiratory health monitoring.

Main Methods:

• Utilized a non-contact tri-axial accelerometer to capture motion signals.
• Extracted 43 time-domain features from accelerometry data.
• Employed six feature ranking methods, including Recursive Feature Elimination (RFE), to select top features.
• Trained 68 logistic regression models using subject-record-split and leave-one-subject-out (LOSO) cross-validation.

Main Results:

• The best subject-record-split model achieved 92.20% accuracy, 90.87% sensitivity, 93.52% specificity, and 92.09% F1 score using 20 RFE-selected features.
• The best LOSO model achieved 89.57% accuracy, 85.71% sensitivity, 93.43% specificity, and 88.72% F1 score using 10 RFE-selected features.
• Demonstrated the effectiveness of motion-based features for cough detection.

Conclusions:

• A motion-based cough detection system using accelerometers is feasible and effective.
• This approach offers a privacy-preserving alternative to audio-based monitoring.
• The findings support the future development of wearable, non-contact cough detectors for respiratory health management.