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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Attention-deficit/hyperactivity disorder (ADHD) diagnosis relies on behavioral symptoms, lacking objective biological markers.
• Electroencephalography (EEG) offers high temporal resolution and low cost, showing potential for ADHD diagnosis assistance.
• Existing EEG analysis methods often overlook joint three-dimensional time-frequency and spatial features.

Purpose of the Study:

• To propose a novel deep learning framework, TFS-FENet, for effectively modeling EEG time-frequency-spatial characteristics.
• To enhance the objective diagnosis of ADHD by integrating advanced feature extraction techniques.
• To investigate the utility of TFS-FENet in classifying ADHD subtypes and distinguishing ADHD from typical development.

Main Methods:

• Developed TFS-FENet, a deep learning framework integrating convolutional neural networks and an attention mechanism.
• Utilized a resting-state EEG dataset comprising ADHD-Inattentive, ADHD-Combined, and typically developing children.
• Performed three-class and binary classification tasks to evaluate model performance against established methods.

Main Results:

• TFS-FENet achieved 96.89% accuracy in the three-class ADHD classification task.
• Accuracy reached 99.36% in the binary classification task (ADHD vs. typically developing children).
• Interpretability analysis indicated reliance on prefrontal, temporal, and occipital regions, aligning with ADHD neurobiology.

Conclusions:

• TFS-FENet demonstrates superior performance in classifying ADHD using EEG data, outperforming existing models.
• The proposed framework effectively captures complex time-frequency-spatial EEG features for improved diagnostic accuracy.
• Findings support the potential of TFS-FENet as an objective tool for ADHD diagnosis, consistent with neurobiological evidence.