Діагностична програмна система для автоматизованої класифікації звуків легень на основі параметрів спектрів вищих порядків

This paper presents a method for the automated classification of lung sounds based on Higher-Order Spectra (HOS) parameters. The system is designed to differentiate between healthy subjects and patients with bronchitis or Chronic Obstructive Pulmonary Disease (COPD). The study was conducted using a verified database of 806 recordings from 275 adult patients, obtained via the ''KoRA-03M1'' digital system, the Littmann 3200 stethoscope, and the open international ICBHI 2017 Respiratory Sound Database (Kaggle). Particular attention is paid to the signal preprocessing stage, which includes an algorithm for removing impulse noise (''spikes'') present in the recorded signals using a bidirectional filtering method. Unlike traditional linear analysis methods, the proposed HOS-based approach accounts for the nonlinear properties of signals and identifies hidden phase coupling between harmonics, which is critical for analyzing noisy biomedical data. Seven optimal features providing maximum class separability were used: the amplitudes of dominant bispectrum peaks, the coordinates of their corresponding bifrequencies, maximum bicoherence, and third- and fourth-order statistical moments (skewness and kurtosis). Based on the compiled database, machine learning models were trained. Multilayer Perceptron (MLP) neural networks demonstrated the highest performance, achieving an accuracy of 97.8%. The results confirm the high diagnostic value of using bispectrum parameters as reliable markers of respiratory pathologies, supporting the integration of the developed software system into modern telemedicine systems for remote monitoring.