A novel method for diagnosing Alzheimer's disease using deep pyramid CNN based on EEG signals

Abstract

Background: The diagnosis of Alzheimer's disease (AD) using electroencephalography (EEG) has garnered more attention recently.

New methods: In this paper, we present a novel approach for the diagnosis of AD, in terms of classifying the resting-state EEG of AD, mild cognitive impairment (MCI), and healthy control (HC). To overcome the hurdles of limited data available and the over-fitting problem of the deep learning models, we studied overlapping sliding windows to augment the one-dimensional EEG data of 100 subjects (including 49 AD subjects, 37 MCI subjects and 14 HC subjects). After constructing the appropriate dataset, the modified DPCNN was used to classify the augmented EEG. Furthermore, the model performance was evaluated by 5 times of 5-fold cross-validation and the confusion matrix has been obtained.

Results: The average accuracy rate of the model for classifying AD, MCI, and HC is 97.10%, and the F1 score of the three-class classification model is 97.11%, which further proves the model's excellent performance.

Conclusions: Therefore, the DPCNN proposed in this paper can accurately classify the one-dimensional EEG of AD and is worthy of reference for the diagnosis of the disease.

Keywords: Alzheimer's disease; Deep learning; Deep pyramid convolutional neural network; Electroencephalography; Mild cognitive impairment.

Fig. 1

Overall flow of the study.

Fig. 2

One subject's frequency domain features of all channels.

Fig. 3

Structure diagram of the deep learning model suitable for one-dimensional frequency domain feature classification of AD.

Fig. 4

Performance of the model (with different step lengths but a fixed window size of 76) using the data augmentation scheme.

Fig. 5

Performance of the model (with different step lengths but a fixed window size of 152) using the data augmentation scheme.

Fig. 6

Performance of the model (with different step lengths but a fixed window size of 228) using the data augmentation scheme.

Fig. 7

Confusion matrix results under the optimal data augmentation scheme of the deep learning model.

Fig. 8

Box plot of 5 times 5-fold cross validation of each evaluation indicator.