Electroencephalography (EEG) is a vital tool for detecting Alzheimer’s Disease (AD). Although deep learning has been widely applied to EEG analysis, existing methods predominantly focus on time-domain or frequency-domain representations, limiting their ability to fully capture the complexity and multifaceted nature of EEG signals. Time-domain approaches excel at modeling temporal dependencies but often neglect critical spectral structures, while frequency-domain methods emphasize spectral patterns at the expense of temporal dynamics. Given the inherent interdependence of temporal and spectral components in EEG signals, a robust framework integrating both domains is essential to exploit their complementary strengths. We propose EEG-TFNet, a multi-domain neural network that adaptively fuses spatiotemporal and spectral features to enhance classification performance. The time-domain branch employs multi-scale selective convolution guided by frequency-domain features to capture temporal patterns modulated by spectral properties. The time-frequency branch, built on a ResNet architecture, extracts amplitude features from time-frequency spectrograms. By synergistically fusing features from both domains, EEG-TFNet effectively models cross-domain dependencies, providing a comprehensive EEG signal representation for AD detection. Extensive experiments on four public datasets using leave-one-subject-out cross-validation demonstrate the model’s superior cross-subject generalization. Comparative evaluations with 5-fold cross-validation against established EEG models and subject-independent split experiments against state-of-the-art time-series models further confirm that EEG-TFNet outperforms existing approaches in capturing complex EEG features for AD detection.