TK1114 : Automated Detection of Cardiac Abnormalities Using Phonocardiogram Signals and Deep Learning baxsed on Combined MFCC and PLP Features
Thesis > Central Library of Shahrood University > Electrical Engineering > MSc > 2026
Authors:
[Author], Hosein Marvi[Supervisor], [Advisor]
Abstarct: Abstract Cardiovascular diseases are among the leading causes of mortality worldwide, and their early diagnosis plays a crucial role in improving treatment outcomes and reducing healthcare costs. Phonocardiogram (PCG) signals provide a non-invasive and cost-effective approach for assessing cardiac function and detecting heart abnormalities. However, the presence of noise and the inherent complexity of heart sound signals make the automatic diagnosis of cardiac disorders a challenging task. In this study, an intelligent frxamework for the automatic classification of PCG signals is proposed. The Yaseen dataset, consisting of five classes including Normal, Aortic Stenosis (AS), Mitral Regurgitation (MR), Mitral Stenosis (MS), and Mitral Valve Prolapse (MVP), was first subjected to preprocessing steps including resampling, band-pass filtering, noise reduction, amplitude normalization, and signal length standardization. To enhance the diversity of training samples, a novel data augmentation strategy termed Class-Aware Hybrid PCG Augmentation (CHPA) was developed. Subsequently, acoustic and perceptual features including Mel-Frequency Cepstral Coefficients (MFCCs), first- and second-order MFCC derivatives, Perceptual Linear Prediction (PLP) coefficients, and spectral features were extracted. To evaluate the effectiveness of the extracted features, several machine learning classifiers, including Support Vector Machine (SVM), Random Forest, K-Nearest Neighbors (KNN), Logistic Regression, Extra Trees, and Gradient Boosting, were investigated. Furthermore, the performance of a conventional Convolutional Neural Network (CNN) and the proposed CHPA Multi-Scale SE-CNN architecture was evaluated. The proposed architecture employs multi-scale convolutional blocks and a Squeeze-and-Excitation (SE) channel attention mechanism to capture temporal patterns at different scales and enhance discriminative feature representations.The results obtained using five-fold cross-validation demonstrated that the best classical classifier, SVM, achieved an accuracy of 99.00%. The CNN model achieved an accuracy of 99.20%. The proposed CHPA Multi-Scale SE-CNN outperformed all evaluated methods, achieving an accuracy of 99.30%, precision of 99.33%, sensitivity of 99.30%, specificity of 99.82%, and F1-score of 99.30%. These findings indicate that the combination of acoustic feature extraction, the proposed CHPA augmentation strategy, and a multi-scale channel-attention-baxsed deep learning architecture provides an effective frxamework for the automatic diagnosis of heart diseases from PCG signals.
Keywords:
#_Phonocardiogram (PCG) #Heart Disease Diagnosis #CHPA #Deep Learning #Convolutional Neural Network (CNN) #Multi-Scale SE-CNN #MFCC #PLP_ Keeping place: Central Library of Shahrood University
Visitor: