Machine-learning based feature selection for a non-invasive breathing change detection.

Area of Science:

• Respiratory Medicine
• Biomedical Engineering
• Data Science

Background:

• Chronic Obstructive Pulmonary Disease (COPD) is a leading global cause of death, necessitating improved patient monitoring for early detection of acute exacerbations.
• Current monitoring methods often focus on breathing rate but lack sufficient sensitivity and specificity.
• This study investigates novel breathing features to better characterize respiratory patterns during load-capacity-drive balance adjustments.

Purpose of the Study:

• To identify breathing features that enhance the early detection of acute exacerbation events in COPD patients.
• To evaluate the effectiveness of different breathing signal features in classifying respiratory states during exercise.
• To improve patient monitoring technologies for COPD management.

Main Methods:

• Analysis of breathing data collected during controlled exercise conditions.
• Evaluation of breathing rate alone versus combined features, including Fourier coefficients, signal amplitude, and ARIMA coefficients.
• Assessment of classification performance for distinguishing rest and effort periods.

Main Results:

• Breathing rate alone demonstrated poor capability in classifying rest and effort periods.
• The highest classification performances were achieved using Fourier coefficients or a combination of breathing rate with signal amplitude and/or ARIMA coefficients.
• Combining features significantly improved the classification power compared to using breathing rate alone.

Conclusions:

• Breathing rate is an inadequate predictor of respiratory changes in COPD patients.
• Incorporating additional features, such as Fourier or ARIMA coefficients, substantially enhances the classification power of breathing signals.
• Feature combination strategies hold promise for improving COPD exacerbation prediction methods.