Attention-deficit hyperactivity disorder (ADHD) is a prevalent neurodevelopmental condition in children characterized by impairments in attention, hyperactivity, and impulse control. Despite extensive research, the underlying cause of ADHD remains unclear. Electroencephalography (EEG), a noninvasive method for recording brain activity, is valuable for studying ADHD-related neural patterns. This study explored the potential of EEG data to differentiate children with ADHD and healthy controls (HC) to enhance diagnostic accuracy. We analyzed EEG recordings from 61 children with ADHD and 60 healthy controls. The EEG data comprised signals from 19 scalp channels. Our primary objective was to develop a machine learning model capable of classifying ADHD subjects with ADHD from HC using EEG data as discriminatory features. To select the most relevant features, we utilized mutual information (MI), a measure of the statistical dependence between two variables. The top features were selected based on their minimum MI values, ensuring that they captured meaningful information from both ADHD and HC groups. Principal component analysis was employed to reduce dimensionality while preserving the essential features, aiming to mitigate computational complexity. The selected features were then used to train ten different classifiers: random forest, multilayer perceptron (MLP), k-nearest neighbors, extra tree classifier, XGBoost, support vector machines, logistic regression, AdaBoost, classification and regression trees, and gradient boosting machines. A stacked classifier was constructed by combining the outputs of all 10 individual classifiers, with the MLP acting as a meta-classifier. The stacked classifier outperformed individual models, achieving an impressive accuracy of 92%. Its precision (91%) and sensitivity (93%) were also higher than those of the individual models, indicating its ability to correctly identify ADHD-positive cases. Furthermore, the specificity of the stacked classifier (93%) was superior, highlighting its improved proficiency in correctly classifying HC. This comprehensive evaluation established the stacked classifier as an effective approach for ADHD classification, surpassing the performance of several standalone models. Our proposed method offers a noninvasive, objective, and cost-effective method for identifying children with ADHD, leading to earlier diagnosis, intervention, and improved treatment outcomes.

Keywords: EEG, ADHD, Machine learning, Mutual information, PCA, Stacked classifier

No potential conflict of interest relevant to this article was reported.