Subject-independent classification of ADHD from pediatric EEG: a leakage-aware deep learning framework with interpretable neurophysiological mapping

Abstract

The use of electroencephalography (EEG) in automated screening of Attention-Deficit/Hyperactivity Disorder (ADHD) has created a large body of literature that suggests promising outcomes, with reported classification accuracies often above 90%. A major methodological issue suggests that a substantial part of these results are actually due to data leakage, whereby EEG epochs pertaining to the same subject end up in both training and testing sets, allowing classifiers to capitalize on features specific to the individual rather than on genuine disorder-related biomarkers. We assessed the performance of three classical machine learning algorithms, namely, Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM), as well as three deep learning models: Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and CNN-BiLSTM, on a standard pediatric EEG dataset containing 121 children and 4,173 epochs. The performance of each model was compared in terms of both segment-wise cross-validation (Protocol A) and subject-wise cross-validation (without any data leakage, Protocol B).To evaluate neurotopographic plausibility without making unsupported causal assertions, we applied Deep SHapley Additive exPlanations (SHAP) to the best-generalizing architecture. The performance discrepancy was substantial. Under segment-wise evaluation, the SVM operating on power spectral density (PSD) features reached 90.0% accuracy (95% CI: [89.1%, 90.9%], AUC = 0.96). However, the same model dropped to 65.6% (95% CI: [62.1%, 69.1%], AUC = 0.72) under subject-wise evaluation—a statistically significant 24.4 percentage point collapse ( p < 0.001, Wilcoxon signed-rank). In contrast, deep architectures learning directly from raw EEG were resilient: the stand-alone CNN lost only 2.5 percentage points, achieving 72.7% accuracy (95% CI: [66.0%, 79.4%], AUC = 0.72, recall = 88.0%) on completely unseen participants. Post-hoc SHAP attribution identified the highest relative contributions in frontal (F3, Fz, F4) and central (Cz) electrodes, aligning with established ADHD neurophysiology. Hand-engineered spectral features are highly vulnerable to identity leakage, whereas end-to-end convolutional models generalize across subjects with realistic, uninflated performance. Strict subject-wise validation is strongly recommended as standard practice for translational EEG research.