Transformer-Based Modeling of Directed Transfer Entropy Connectivity for EEG-Based ADHD Classification in Children
TL;DR
The Contextualized Transfer Entropy Network (CTE-Net) is proposed, an end-to-end deep-learning architecture that combines global content-based contextualization with nonlinear and directed EEG connectivity estimation and exhibited the lowest within-subject dispersion among the analyzed representation stages.
Abstract
Electroencephalography (EEG) provides a non-invasive and cost-effective tool for supporting the assessment of attention-deficit/hyperactivity disorder (ADHD). However, the nonstationary nature of EEG produces substantial variability among signal windows recorded from the same participant, which can obscure diagnostic structure and lead to inconsistent predictions. To address this problem, we propose the Contextualized Transfer Entropy Network (CTE-Net), an end-to-end deep-learning architecture that combines global content-based contextualization with nonlinear and directed EEG connectivity estimation. CTE-Net first employs a Transformer encoder to contextualize the multichannel representations within each EEG window. The resulting signals are processed using channel-wise nonlinear temporal filters and Takens delay-coordinate embeddings. A differentiable matrix-based Transfer Entropy module, formulated using Rényi’s α-entropy and a rational quadratic kernel, then estimates directed predictive information dependencies between all ordered electrode pairs. The resulting connectivity coefficients are used for ADHD-versus-control classification. The model was evaluated on a publicly available pediatric EEG dataset comprising 120 participants, equally divided between ADHD and control groups, using five fixed subject-wise folds and ten random training repetitions. At the window level, CTE-Net achieved an accuracy of 80.9±1.7%, precision of 82.7±2.1%, and sensitivity of 84.2±2.3%. At the participant level, it achieved an accuracy of 83.4% (95% CI: 78.2–88.2) and an ROC-AUC of 90.2% (95% CI: 85.1–94.6), demonstrating competitive and comparatively balanced classification performance. Beyond classification performance, the directed Transfer Entropy representation exhibited the lowest within-subject dispersion among the analyzed representation stages, with a median reduction of 38.35% relative to raw EEG. This reduction remained consistent across different PCA dimensionalities and distance definitions. These single-dataset findings support CTE-Net as a compact and interpretable methodological framework for representing directed EEG interactions while attenuating window-specific variability within individual participants.