A Multi-Branch EEGNet–Transformer Hybrid Network for Automated ADHD Screening from Multichannel EEG

Abstract

Attention-Deficit/Hyperactivity Disorder (ADHD) diagnosis remains dependent on subjective instruments and interviews, motivating objective screening tools. Electroencephalography (EEG) is non-invasive and cost-effective, but automated ADHD classification is challenged by noisy recordings, inter-subject variability, and complex spatiotemporal structure. We propose MB-EEGNet-T, a compact multi-branch hybrid model that fuses an EEGNet-like convolutional branch for local temporal–spatial feature extraction with a temporal Transformer branch that models longer-range dependencies using multi-head self-attention. Using an open-access 19-channel pediatric EEG dataset of 61 ADHD and 60 control participants [1], we construct overlapping 128-sample segments (≈1 s at 128 Hz) with 50% overlap. Preprocessing applies per-channel standardization and a 0.5–45 Hz Butterworth band-pass filter; training is regularized via physiologically plausible augmentations (temporal shift, amplitude scaling, additive Gaussian noise, and stochastic channel dropout). On a stratified segment-level split, MB-EEGNet-T achieves 92.2% accuracy, 0.922 F1-score, and 0.976 ROC-AUC on a held-out test set, improving over an EEGNet-only baseline. The model uses 61k parameters (<0.25 MB), supporting efficient inference. We discuss evaluation pitfalls (e.g., subject leakage) and outline steps toward subject-independent validation and clinical decision support.