Cough sound analysis and classification

The main objective of this study is to develop a classification system that analyzes cough sounds and distinguishes non-productive (dry) and productive (wet) cough types and cough sounds from non-cough sounds such as breath, voice, and snoring. The study extracted the features of cough sounds in time and frequency domains and considered their effects on classification performance. Autoregressive (AR) coefficients were used in the time domain, and Mel-Frequency Cepstral Coefficients (MFCC) and percentage frequency values representing energy distribution (25th, 50th, 75th, 90th) were used in the frequency domain. The Linear Discriminant Analysis (LDA) method was applied to reduce the feature space size and improve the classification performance. In the optimized low-dimensional feature space obtained with LDA, Support Vector Machines (SVM), Random Forest, k-Nearest Neighbors (k-NN), and Decision Tree algorithms from Machine Learning techniques were tested. The study used cough, breath, and speech sounds obtained from the COVID-19 audio database created by Cambridge University. In addition, snoring sounds were obtained from Google's Audio Set database. The performance of the classifiers was measured using the Leave-One-Out Cross-Validation (LOOCV) approach and evaluated with accuracy, sensitivity, specificity, and F1 scores. The results reveal the importance of MFCC, AR coefficients, and energy percentage frequencies in analyzing cough sounds. The results of this thesis study aim to contribute to non-invasive and mobile systems developed for diagnosing respiratory diseases by analyzing cough sounds.