|
|
 |
| |
|
|
|
Predicting heart disease using modified GoogLeNet convolutional neural network architecture based on the heart sound |
|
|
|
PP: 2837-2858 |
|
|
|
doi:10.18576/isl/121101
|
|
|
|
Author(s) |
|
|
|
Abdullah R. Rashwan,
Laila El Fangary,
Safaa M. Azzam,
|
|
|
|
Abstract |
|
|
| Based on the data of the world health organization (WHO), diagnosing heart disease is a great task, as heart disease (HD) is the most prevalent disease worldwide. We suggested a method based on heart sounds to deal with this difficult issue because the heart sound (HS) is an essential component for detecting heart conditions. A feature extraction technique and a classifier are used in the suggested strategy. We use the GoogLeNet convolutional neural network (CNN) architecture with some modifications to separate the most crucial attributes of HS, and the heart condition is classified as diseased or not diseased based on these attributes. The model is trained using the AdaBelief optimizer to tune the parameters of our modified GoogLeNet architecture. The model was trained and validated utilising various datasets from PhysioNet 2016. Additional training samples were provided by integrating the PASCAL dataset with the PhysioNet 2016 dataset. Additionally, the variety of samples from various sources enabled our system to learn about sounds from everyday life more accurately. Our results indicated that using a modified GoogLeNet architecture with the AdaBelief optimizer, the trained model obtained test accuracy of 100% and 99.9% on unseen HS recordings from PhysioNet and the merged datasets, respectively. By comparing our proposed model with the highest-scoring methods listed on the official PhysioNet website in these datasets, the results show significantly improved.
|
|
|
|
|
|
|
|
