Fonokardiyografi Sinyalleri İle Derin Öğrenme Tabanlı Karar Destek Sistemi Tasarımı
Özet
This study aims to classify cardiovascular diseases using phonocardiogram (PCG) signals through deep learning methods, thereby laying the foundation for developing a decision support system for doctors. In this way, it is planned to support physicians in the preliminary diagnosis process.
In this study, the classification of heart sounds was performed using the HU-PCG dataset, which was previously collected within our department. Additionally, comparative analyses were conducted using commonly used datasets in heart sound classification, including PhysioNet/CinC 2016, PASCAL, and HVD datasets. Both traditional machine learning methods, which involve multiple stages, and end-to-end deep learning-based methods were evaluated for heart sound classification. For the machine learning-based approach, a five-stage classification process was examined. In this approach, after the preprocessing step of the signals, a heart sound identification algorithm was employed using feature vectors created with MFCC and DWT techniques, finally segmentation is done. Features were extracted from the segmented signals, and classification was performed using Linear Classifier, k-NN, SVM, and MLP-BP methods. The classification was conducted in two levels: first, the identification of normal and abnormal sounds, followed by the classification of diseases. For classifying PCG signals using deep learning methods, a 1D convolutional neural network (1DCNN) model was developed, where the PCG signal, which underwent simple preprocessing steps, served as the input. Additionally, a 2D convolutional neural network (2DCNN) model was developed, using mel spectrograms created from preprocessed PCG signals as the input. Another objective of this study is to perform a two level classification of PCG signals and compare it with single level classification. In this context, models such as 2H-1DCNN and 2S-1DCNN were developed using 1DCNN, and 2S-2DCNN models were developed using 2DCNN. With the 1DCNN and 2DCNN models, binary classification of normal versus abnormal was conducted, and three-class classification based on normal and disease states was performed using the 1DCNN, 2H-1DCNN, 2S-1DCNN, 2DCNN, and 2S-2DCNN models.
The study explored methods to improve the model's performance, such as using data augmentation by dividing the PCG signal into segments of equal length, random crop of the PCG signal of equal length in each epoch, and segmenting the PCG signal to create input data for each segment. Furthermore, transfer learning was employed to enhance classification performance by retraining a model, initially trained on larger datasets, on smaller datasets. The use of these approaches significantly improved the performance of deep learning-based models, providing an effective solution for classifying heart sounds.
In this study, for normal-abnormal classification on the HU-PCG dataset, the highest accuracy metric of 1.0 was achieved using 1DCNN and 2DCNN models with random crop from within the signal and transfer learning methods. For three-class models trained to distinguish between normal, aortic stenosis, and mitral regurgitation, the highest performance value of 0.93 accuracy was achieved using the 2S-1DCNN model with transfer learning and random crop methods, while the second-highest accuracy (that is 0.87) was achieved by the 2S-2DCNN model using the same methods.
