SCANet: An Innovative Multiscale Selective Channel Attention Network for EEG-Based ADHD Recognition

Attention Deficit Hyperactivity Disorder (ADHD) is one of the most common neurodevelopmental disorders of childhood that significantly impacts the patient’s cognitive and behavioral functions. Traditional diagnostic methods are time-consuming, highly subjective, and prone to misdiagnosis. Electroencephalogram (EEG) data, due to its high temporal resolution and non-invasiveness, can help mitigate these issues. Current approaches using EEG for ADHD identification face challenges such as limited accuracy and generalizability. In this paper, we propose a novel Selective Channel Attention Network (SCANet) that integrates attention mechanisms to improve the classification of EEG signals for ADHD, Attention Deficit Disorder (ADD), and Healthy Controls (HC). SCANet employs depthwise separable convolutions, a multi-scale and dual-branch architecture, to effectively extract features from EEG signals. We introduce the Selective Channel Attention Mechanism (SCAM) combined with self-attention to emphasize inter-channel interactions and global temporal features. Our model demonstrated exceptional performance across both public and private datasets. The model achieved remarkable performance with 99.78% accuracy, 99.78% precision, and 99.79% F1-score on the public three-class dataset, and 87.12% accuracy, 88.64% precision, and 89.14% F1-score on the private binary dataset. In comparison with EEGNet, EEG-Transformer, CNN-LSTM, ablation studies, SCANet shows superior performance and stability for diagnosing ADHD. Additionally, we apply Gradient-weighted Class Activation Mapping (Grad-CAM) to analyze the contribution of EEG channels and time points to the diagnosis, thereby enhancing the model’s interpretability. In summary, the SCANet model shows significant potential for clinical application in diagnosing ADHD, and could provide a robust, efficient alternative for current EEG data classification applications.