TFS-FENet: A time–frequency spatial deep learning framework for EEG-based ADHD subtype classification

Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder, with clinical diagnosis primarily relying on behavioral manifestations and lacking objective biological markers. Electroencephalography (EEG), due to its high temporal resolution and low cost, has been considered to hold great potential for assisting ADHD diagnosis. However, existing methods mostly focus on modeling two-dimensional features such as time-space, frequency-space, or time-frequency representations of EEG, while very few approaches simultaneously capture the joint three-dimensional time-frequency and spatial features. In this paper, we propose a novel deep learning framework-TFS-FENet (Time-Frequency Spatial Feature Extraction Net)-which integrates convolutional neural networks with an attention mechanism to effectively model the time-frequency-spatial characteristics of EEG. Using a resting-state EEG dataset collected at Southeast University, consisting of three categories (ADHD-Inattentive, ADHD-Combined, and typically developing children), experimental results show that TFS-FENet achieves an accuracy of 96.89% in the ADHD three-class classification task, significantly outperforming comparison models such as ShallowConvNet, EEGNet, ResNet18, and DeepConvNet. In the binary classification task (ADHD vs. typically developing children), accuracy further improved to 99.36%. Moreover, interpretability analysis based on channel weights revealed that the model relied more heavily on the prefrontal, temporal, and occipital regions, which is highly consistent with existing neurobiological findings of ADHD.