Abstract
Electroencephalographic (EEG) signals are acquired non-invasively from electrodes placed on the scalp. Experts in the field can use EEG signals to distinguish between patients with Alzheimer’s disease (AD) and normal control (NC) subjects using classification models. However, the training of deep learning or machine learning models requires a large number of trials. Datasets related to Alzheimer’s disease are typically small in size due to the lack of AD patient samples. The lack of data samples required for the training process limits the use of deep learning techniques for further development in clinical settings. We propose to increase the number of trials in the training set by means of a decomposition–recombination system consisting of three steps. Firstly, the original signals from the training set are decomposed into multiple intrinsic mode functions via multivariate empirical mode decomposition. Next, these intrinsic mode functions are randomly recombined across trials. Finally, the recombined intrinsic mode functions are added together as artificial trials, which are used for training the models. We evaluated the decomposition–recombination system on a small dataset using each subject’s functional connectivity matrices as inputs. Three different neural networks, including ResNet, BrainNet CNN, and EEGNet, were used. Overall, the system helped improve ResNet training in both the mild AD dataset, with an increase of 5.24%, and in the mild cognitive impairment dataset, with an increase of 4.50%. The evaluation of the proposed data augmentation system shows that the performance of neural networks can be improved by enhancing the training set with data augmentation. This work shows the need for data augmentation on the training of neural networks in the case of small-size AD datasets.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have