Comparative study of multi-headed and baseline deep learning models for ADHD classification from EEG signals.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Rising prevalence of Attention-Deficit/Hyperactivity Disorder (ADHD) in children necessitates improved diagnostic tools.
• Electroencephalography (EEG) offers a noninvasive method for ADHD detection, but full channel utilization poses computational challenges.
• Model overfitting is a risk with complex EEG analyses.

Purpose of the Study:

• To compare a novel multi-headed deep learning framework against a traditional single-model approach for ADHD classification using EEG.
• To evaluate the efficacy of using a reduced set of strategically selected EEG channels.
• To assess the ability of multi-headed models to capture inter-channel relationships and temporal features.

Main Methods:

• Collected EEG data from 79 participants (37 with ADHD, 42 healthy) across four cognitive states.
• Utilized five strategically selected EEG channels to mitigate computational complexity.
• Implemented a multi-headed deep learning framework with parallel branches (Bidirectional Long Short-Term Memory, Long Short-Term Memory, Gated Recurrent Unit).

Main Results:

• The multi-headed framework, specifically the Long Short-Term Memory and Bidirectional Long Short-Term Memory combination, achieved the highest classification accuracy of 89.87%.
• This performance significantly surpassed all tested baseline single-model configurations.
• The multi-headed approach effectively extracted richer temporal features and inter-channel relationships.

Conclusions:

• Multi-headed deep learning models demonstrate significant potential for enhancing EEG-based ADHD diagnosis.
• Integrating diverse deep learning architectures within a multi-headed framework improves classification accuracy.
• This approach offers a computationally efficient and effective method for ADHD detection.