Abstract:
Alzheimer’s disease is a chronic disease which is characterized by the gradual loss of
cognitive functions and the impairment of the ability to perform activities of daily
living. This paper presents a new approach for the diagnosis of AD during its early
stage through the use of EEG signals and deep learning. EEG is a cheap, noninva sive and portable neuroimaging technique which makes it a potential method for the
diagnosis of early stage neurological diseases. The suggested framework incorporates
state-of-the-art preprocessing techniques, channel filtering, and Continuous Wavelet
Transform (CWT)-based spectrograms for converting the raw EEG signals into time frequency domain. These spectrograms are used as input for the conventional con volutional neural networks (CNNs) including VGG16, ResNet50, InceptionV3 and
InceptionResNetV2 for the feature extraction and classification. To determine the
models, optimizers, and hyperparameters, several experiments were carried out to
find the best fit. It also shows that the ResNet50 model with Adam optimizer and
the learning rate of 0.0001 has the best classification performance of 96.69%, which
is better than other models. The results also show that the choice of channels, the
adjustment of the hyperparameters and the use of deep learning classifiers are crucial
to increase the accuracy of the AD detection model based on EEG. This framework
offers a convenient, non-invasive and low-cost way for the early diagnosis of AD and
thus is applicable for the clinical practice and the field of precision medicine. In the
future, the work will be extended in terms of increasing model interpretability and
using more data and developing lighter models to be implemented on the edge to in crease the chances of real-life application.
