Abstract <p>Objective assessment of spatial cognitive ability is crucial for screening cognitive impairment and in neurorehabilitation. While deep learning methods for electroencephalogram (EEG) analysis show great promise, they increasingly rely on complex, parameter-heavy architectures. This complexity often leads to poor generalization due to overfitting on small datasets and hinders deployment on mobile healthcare devices. To overcome these limitations, we propose a novel lightweight multi-scale channel attention network with depthwise residual blocks. The model incorporates multi-scale convolutional layers to capture diverse temporal and spatial patterns in EEG signals. It then leverages channel attention mechanisms to dynamically prioritize informative channels, focusing on task-critical features. Furthermore, a novel depthwise separable residual block is introduced to significantly reduce computational complexity and maintain stable model performance. Evaluations on a spatial cognition EEG dataset demonstrate that our network achieves higher accuracy than baselines with only 8.453M parameters, making it an efficient and practical solution for mobile deployment. It also holds strong potential for extension to early screening and intervention in a wider range of cognitive disorders.</p> Graphical abstract <p></p>