High performance method for COPD features extraction using complex network.

Area of Science:

• Medical Informatics
• Signal Processing
• Machine Learning

Background:

• Chronic Obstructive Pulmonary Disease (COPD) diagnosis relies on clinical assessment and spirometry.
• Objective analysis of respiratory sounds offers a non-invasive diagnostic adjunct.
• Current methods for respiratory sound analysis can be enhanced with advanced feature extraction and classification.

Purpose of the Study:

• To develop and validate a novel methodology for COPD classification using respiratory sound attributes.
• To explore the efficacy of various signal processing techniques and machine learning algorithms for COPD detection.
• To establish a robust and accurate diagnostic tool for COPD based on lung sound spectrograms.

Main Methods:

• Respiratory sounds were segmented into breaths, and features were extracted using Spectral Transforms and Wavelet Transforms (e.g., Daubechies Db2).
• Complex Network analysis was applied to spectrograms to generate novel graph-based representations.
• Six machine learning algorithms were trained and fine-tuned for COPD classification, with hyperparameter optimization.

Main Results:

• The proposed methodology demonstrated high performance, with Area Under the Curve (AUC) consistently exceeding 96% across different time-frequency transformations.
• The Random Forest algorithm achieved a peak AUC of 99.67%, significantly outperforming other evaluated algorithms.
• The Wavelet Daubechies 2 (Db2) transform, particularly with the Naive Bayes algorithm, approached 98% accuracy.

Conclusions:

• The study successfully demonstrated a promising approach for accurate COPD diagnosis using machine learning applied to respiratory sound spectrograms.
• The integration of Complex Network analysis and optimized classification algorithms yielded superior diagnostic performance.
• This methodology offers a potential non-invasive and effective tool for early and accurate COPD detection.