Automated attention-deficit hyperactivity disorder detection using an explainable EEGNet model based on electroencephalogram signals

Attention-deficit/hyperactivity disorder (ADHD) is a heterogeneous neurodevelopmental condition where traditional diagnostics often suffer from subjective biases and a lack of objective biomarkers. In this study, we developed a novel, automated, and interpretable diagnostic framework using the specialized EEGNet model to extract spatio-temporal patterns directly from raw EEG signals, ensuring robust ADHD detection and model generalizability across a clinical lifespan ranging from 5 to 89 years. The primary innovation of this research lies in a specialized, lightweight EEGNet model that directly processes raw 5-second spatio-temporal EEG segments, eliminating the need for manual feature engineering or signal-to-image transformation. Our findings demonstrate that EEGNet achieves superior performance, reaching an average accuracy of 89.48% (±4.21%) and an AUC of 0.96 (±0.03), significantly outperforming the transfer learning models. A distinguishing feature of this work is the transition from black-box modeling to explainable artificial intelligence (XAI) by integrating Saliency Maps and SHAP values, which decoded the model's decision-making logic to identify key contributing channels (Fp1, Fp2, C3, T5, and A2) whose spatial attributions align with established neuroscientific literature on ADHD. The high ranking of the Fp1 (prefrontal) channel aligns with established neuroscientific evidence regarding executive dysfunction in ADHD. Furthermore, the implementation of a rigorous inter-subject evaluation demonstrates the preliminary clinical feasibility of this interpretable framework, providing a foundation for future validation on larger multi-center cohorts and offering a robust assistive tool for real-world neurological assessments.