Ankara Yıldırım Beyazıt Üniversitesi Fen Bilimleri Enstitüsü, Ankara, 2018
Previous studies on classifying Electroencephalography (EEG) sleep data generally
use the signal itself. Many of these studies need series of pre-processing operations,
manual feature extraction, complex and hard application processes. There is no need
for lots of pre-processing stages in Convolutional Neural Networks (CNN) and
features can be learned automatically instead of using manually extracted features.
Also CNNs have better performance over most of other methods in visual
The study presented in this thesis is based on applying transfer learning with CNNs
via spectrogram images, that were obtained by using Short-Time Fourier Transform
(STFT), on ISRUC-Sleep dataset (ISRUC) for classifying EEG signals. AlexNet
trained with 100 patient subjects from ISRUC was used as pre-trained CNN. For
classifying sleep stages, single-channel EEG data, that was taken from 10 healthy
people, was used as target domain. To reduce overfitting, an image translation
operation was employed and images were augmented horizontally. The main purpose
of using transfer learning method in this study is achieving better training duration and
accuracy. Applying transfer learning increased the accuracy of the classification by
3.11 percent, when compared to the result of using non-pretrained AlexNet, which was
trained from scratch.
When the source domain of transfer learning was ImageNet, the accuracy was
decreased by 2.73 percent compared to the result of training from scratch. Results
showed that transfer learning increases the accuracy when target and source domains
are similar, but it may decrease the accuracy when used on different domains (i.e.,
ImageNet is an image dataset and ISRUC consists of signals). In addition to the study,
scalogram images from Continuous Wavelet Transform (CWT) were also tested.