Explainable electroencephalography-based attention-deficit/hyperactivity disorder detection model with a combination of ternary pattern and twin wavelet transform

Abstract

BACKGROUND Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental condition characterized by inattention, impulsivity, and hyperactivity. Traditional diagnosis relies on clinical evaluation, which is time-consuming and subjective. Electroencephalography (EEG) signals provide an objective alternative, and machine learning methods can improve their diagnostic utility. AIM To develop an explainable EEG-based model for ADHD detection by integrating a novel combination ternary pattern (CTP) feature extractor with twin wavelet transform (TWT) for multilevel signal analysis, and to evaluate its effectiveness in providing accurate, channel-wise, and fusion-based classification results for objective and rapid ADHD diagnosis. METHODS A new EEG dataset containing more than 7000 segments from 137 ADHD patients and 150 controls was studied. A novel feature engineering framework was developed, combining a new CTP extractor with statistical features. A multilevel feature extraction structure was designed using a newly proposed TWT for signal decomposition. Extracted features were reduced to the most informative 263 using neighborhood component analysis. Channel-wise classification was performed with k-nearest neighbors, followed by iterative majority voting across 20 EEG channels. RESULTS Single-channel analysis achieved up to 99.12% accuracy. By applying majority voting, overall classification accuracy increased to 99.97%, with similarly high sensitivity and specificity. CONCLUSION Our study introduces a large ADHD EEG dataset and a novel model integrating TWT and CTP. The model provides highly accurate, channel-wise, and fusion-based results, offering a promising objective tool for rapid ADHD diagnosis.