ADHD/CD-NET: automated EEG-based characterization of ADHD and CD using explainable deep neural network technique.
Hui Wen Loh1, Chui Ping Ooi1, Shu Lih Oh2
1School of Science and Technology, Singapore University of Social Sciences, Singapore, Singapore.
Cognitive Neurodynamics
|August 6, 2024
Summary
A new deep learning system, ADHD/CD-NET, uses EEG data to accurately distinguish between attention deficit hyperactivity disorder (ADHD) and conduct disorder (CD), improving diagnostic accuracy for these complex conditions.
Area of Science:
- Neuroscience
- Computational Psychiatry
- Medical Imaging
Background:
- Attention deficit hyperactivity disorder (ADHD) and conduct disorder (CD) are common childhood neurodevelopmental and behavioral disorders, respectively.
- Distinguishing between ADHD and CD is challenging due to overlapping symptoms, increasing the risk of misdiagnosis and inappropriate treatment.
- Accurate differentiation is critical as treatment plans differ significantly for ADHD and CD.
Purpose of the Study:
- To develop and validate an electroencephalogram (EEG)-based deep learning system, ADHD/CD-NET, for objective differentiation of ADHD, ADHD+CD, and CD.
- To assess the diagnostic performance of the proposed system using both internal and external datasets.
- To enhance the interpretability of the deep learning model's predictions using Gradient-weighted Class Activation Mapping (Grad-CAM).
Main Methods:
- EEG signals were processed into channel-wise continuous wavelet transform (CWT) correlation matrices.
- A convolutional neural network (CNN) model (ADHD/CD-NET) was trained using these matrices and evaluated with 10-fold cross-validation.
- Grad-CAM was employed to identify significant EEG channels contributing to the diagnostic outcomes.
Main Results:
- ADHD/CD-NET achieved high classification accuracy: 93.70% (internal) and 98.19% (external).
- The system demonstrated excellent sensitivity (90.83% internal, 98.36% external) and specificity (95.35% internal, 98.03% external).
- Grad-CAM successfully highlighted key EEG channels, providing insights into the model's decision-making process.
Conclusions:
- ADHD/CD-NET offers a robust and objective method for distinguishing between ADHD, CD, and their comorbidity using EEG data.
- The system's ability to perform temporal localization and identify significant channels aids clinical diagnosis.
- This deep learning approach provides valuable objective analysis for mental health professionals in diagnosing ADHD and CD.
