EEGConvNeXt: A novel convolutional neural network model for automated detection of Alzheimer's Disease and Frontotemporal Dementia using EEG signals

EEGConvNeXt: A novel convolutional neural network model for automated detection of Alzheimer's Disease and Frontotemporal Dementia using EEG signals

Background and objective
Deep learning models have gained widespread adoption in healthcare for accurate diagnosis through the analysis of brain signals. Neurodegenerative disorders like Alzheimer's Disease (AD) and Frontotemporal Dementia (FD) are increasingly prevalent due to age-related brain volume reduction. Despite advances, existing models often lack comprehensive multi-class classification capabilities and are computationally expensive. This study addresses these gaps by proposing EEGConvNeXt, a novel convolutional neural network (CNN) model for detecting AD and FD using electroencephalogram (EEG) signals with high accuracy.
Materials and method
In this research, we employ an open-access EEG signal public dataset containing three distinct classes: AD, FD, and control subjects. We then constructed a newly proposed EEGConvNeXt model comprised of a 2-dimensional CNN algorithm that firstly converts the EEG signals into power spectrogram-based images. Secondly, these images were used as input for the proposed EEGConvNeXt model for automated classification of AD, FD, and a control outcome. The proposed EEGConvNeXt model is therefore a lightweight model that contributes to a new image classification CNN structure based on the transformer model with four primary stages: a stem, a main model, downsampling, and an output stem.
Results
The EEGConvNeXt model achieved a classification accuracy of ∼95.70% for three-class detection (AD, FD, and control), validated using a hold-out strategy. Binary classification cases, such as AD versus FD and FD versus control, achieved accuracies exceeding 98%, demonstrating the model's robustness across scenarios.
Conclusions
The proposed EEGConvNeXt model demonstrates high classification performance with a lightweight architecture suitable for deployment in resource-constrained settings. While the study establishes a novel framework for AD and FD detection, limitations include reliance on a relatively small dataset and the need for further validation on diverse populations. Future research should focus on expanding datasets, optimizing architecture, and exploring additional neurological disorders to enhance the model's utility in clinical applications.

Files associated with this item cannot be displayed due to copyright restrictions.