Towards high-accuracy classifying attention-deficit/hyperactivity disorders using CNN-LSTM model

Area of Science:

• Neuroscience
• Computational Psychiatry
• Medical Imaging

Background:

• Attention-deficit/hyperactivity disorder (ADHD) is a common childhood neuropsychiatric disorder.
• Current ADHD diagnosis relies on subjective evaluations, lacking objective indicators.
• Defects in neurocognitive attention functions across brain regions are implicated in ADHD.

Purpose of the Study:

• To propose an effective method for identifying children with ADHD using objective indicators.
• To develop a model capable of classifying ADHD, attention deficit disorder (ADD), and healthy children.
• To leverage electroencephalogram (EEG) signals for ADHD detection.

Main Methods:

• A CNN-LSTM model was developed to classify ADHD, ADD, and healthy children.
• The model was trained on a public EEG dataset of 144 children, including event-related potential (ERP) signals.
• Convolution visualization and saliency map methods were employed for feature interpretability.

Main Results:

• The CNN-LSTM model achieved 98.23% accuracy in a five-fold cross-validation, outperforming state-of-the-art CNN models.
• Extracted features were primarily located in frontal and central brain areas.
• Significant differences in time-period mappings, including P300 and contingent negative variation (CNV) ERPs, were observed among the groups.

Conclusions:

• The CNN-LSTM model effectively identifies children with ADHD and its subtypes.
• Visualized features provide interpretable insights into ERP differences between groups.
• The model shows potential as reliable neural biomarkers for more accurate clinical ADHD diagnosis.