Explainable artificial intelligence-driven visual task-specific electroencephalogram analysis for attention deficit...

Attention deficit hyperactivity disorder (ADHD) is a neurological disorder that primarily develops in early childhood and affects motor development, vision, information processing, memory, and sensory functioning in children. Although many works have been focused on developing machine learning (ML) models to diagnose ADHD from electroencephalography (EEG) data, they lack in interpreting the predictions. This article addresses this gap by proposing an explainable artificial intelligence (AI)-driven approach that computes entropy, mutual information, and transfer entropy from EEG data using the 10-20 system, aggregates electrode-level features into lobe-wise representations, applies ML models, and selects the optimal model for interpretable predictions. Experimental and statistical results show that support vector machine provides better results, 92% accuracy of classification of ADHD and proves to be suitable for clinical screening tasks than other state of the art algorithms. Furthermore, explanations are obtained using Local Interpretable Model-agnostic Explanations (LIME), Shapely Additive Explanations (SHAP), and Partial Dependency Plot (PDP). These reveal that higher values of entropy and transfer entropy in frontal region indicates uncertainty; whereas less participation of occipital region denotes dysfunctionality toward visual perception in children with ADHD. The PDP proves that the maximum values of the three measures in frontal, occipital, and central lobes contribute to the feature importance for the prediction class.