Exploring classical machine learning for identification of pathological lung auscultations.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Machine learning (ML) applications in biomedical research are rapidly expanding.
• Despite advances in digital stethoscopes and AI, clinical applications for pulmonary auscultation analysis are limited.
• Current diagnostic methods for pulmonary conditions often lack accuracy, impacting patient care.

Purpose of the Study:

• To apply ML techniques to analyze pulmonary auscultation signals.
• To differentiate between normal and abnormal lung sounds using audio features.
• To evaluate the effectiveness of various ML algorithms in this diagnostic task.

Main Methods:

• Utilized digital 6-channel auscultations from 45 patients.
• Extracted audio features (e.g., F0-4, loudness, HNR, DFA, log energy, RMS, MFCC) using the Python library Surfboard.
• Applied unsupervised (fair-cut forest, outlier forest) and supervised (random forest, regularized logistic regression) ML models with data preprocessing techniques.

Main Results:

• Supervised ML models outperformed unsupervised methods.
• Random forest achieved a mean AUC ROC of 0.691 (accuracy 71.11%) for side-based detection and 0.721 (accuracy 68.89%) for patient-based detection.
• Decision fusion by averaging outputs improved performance.

Conclusions:

• ML, particularly supervised learning with random forest, shows significant potential for accurate pulmonary sound analysis.
• This approach can enhance preliminary diagnosis accuracy, improving patient care.
• Further development is needed to integrate these ML solutions into clinical practice.