Results - Details
Search command : Author="Στυλιανού"
And Author="Ιωάννης"
Current Record: 5 of 31
|
Title |
Phonocardiogram signal classification using deep learning models |
Alternative Title |
Αλγόριθμοι επεξεργασίας ιατρικών σημάτων: φωνοκαρδιογράφημα |
Author
|
Καλαiτζή Ελένη
|
Thesis advisor
|
Στυλιανού, Ιωάννης
Καφεντζής, Γεώργιος
|
Abstract |
Heart disease is the leading cause of death in all continents except Africa. Heart diseases include coronary heart
disease, heart failure, hypertensive heart disease, congenital heart disease, valvular heart disease, and structural
heart abnormalities.
This thesis aims to improve the task of automatic risk detection of structural heart abnormalities from
digital phonocardiogram (PCG) data with applications to pediatric heart disease screening. In a previous study,
a number of convolutional neural network (CNN)-based systems operating on mel-frequency cepstral coefficients
(MFCCs) of phonocardiogram signals are shown to perform better than systems using conventional machine
learning guided by hand-crafted features. These systems were trained using time-frequency representations of
segmental PCG frames. Various techniques for segmentation and time-frequency representations were designed
to model the input recordings. Training and testing has been conducted on high-quality databases while each
test produces five random experiments. Systems were evaluated using Receiver Operating Curve-Area under
Curve, Sensitivity, Specificity, Accuracy, F1 score, and Matthews Correlation Coefficient.
In the context of this thesis, we emphasize on the improvement of pre-existing Convolutional Neural Network (CNN) models. We undertake this endeavor while adhering to the identical preprocessing methodology
employed in a prior study. A Batch Normalization layer is included in pre-existing architectures, showing
improvement in all performance metrics. Additionally, we introduce an enhanced CNN architecture combining Batch Normalization and an additional convolutional layer that outperforms previous models and presents
significant improvement in all performance metrics.
|
Language |
English |
Issue date |
2023-11-22 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Physics--Graduate theses
|
|
Type of Work--Graduate theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/e/9/f/metadata-dlib-1697782568-167239-31750.tkl
|
Views |
329 |