A Study of EEG Signal Classification for Alzheimer's Disease Based on Feature-Fused Two-Branch Parallel Networks

Alzheimer's disease (AD) is a neurodegenerative disease that is mainly characterised by an insidious onset and subtle clinical symptoms, making it difficult to diagnose by conventional means. Electroencephalography (EEG) has been widely used to detect Alzheimer's disease as an important tool for assessing and aiding the diagnosis of brain disorders. Accurate diagnosis is essential to prevent progression from early cognitive impairment to AD and to provide early treatment for AD patients. Most classification methods focus only on the time or frequency domain, but ignore the relationship between the two as well as the spatial information of the channels. To address this problem, this paper proposes a feature fusion two-branch parallel network (FFDBPNet) for AD diagnosis, which consists of a convolutional neural network (CNN) and a branch of a long short-term memory (LSTM) recurrent neural network. First, the EEG data are pre-processed and transformed into time-frequency spectrograms, which are dimensionally reduced and fed into the FFDBPNet network. The CNN extracts the time-frequency spectral spatial features of multi-band EEG signals, and the LSTM extracts the time-series features of EEG signals. In addition, a flattening layer is added after the pooling layer to further extract intermediate features to improve the generalisation ability of the model. Finally, feature fusion is performed to obtain the classification results. The classification accuracy of the proposed model for AD EEG signals reaches 96.09%, the sensitivity is 98.04%, and the specificity is 94.14%, which exceeds the traditional classification methods. The experimental results show that the feature fusion model proposed in this paper can extract effective time-frequency features and spatial features from EEG for classification.