EEG Biomarkers for Rett Syndrome: A Comparative Analysis of Machine Learning and Deep Learning with Explainable AI
Journal:
medRxiv
Published Date:
Oct 23, 2026
(2)
Abstract
Rett syndrome (RTT) is a rare neurodevelopmental disorder characterised by abnormal cortical activity, which can be captured using electroencephalography (EEG). Although machine learning (ML) has shown promise for diagnosing several neurological conditions, comprehensive evaluation of classical ML and deep learning (DL) methods, together with physiologically meaningful explanations, remains limited for RTT. This study analyses resting-state, eyes-closed EEG recordings from RTT patients and healthy controls using both power spectral density (PSD)-based ML and end-to-end DL frameworks. Classical ML models, including Support Vector Machines, Random Forest, and XGBoost, were trained on PSD features, while EEG-specific DL architectures, including ShallowConvNet, DeepConvNet, EEGNet, and EEG Conformer, were trained directly on preprocessed EEG signals. Subject-wise cross-validation was used to avoid data leakage, with balanced accuracy as the primary performance metric because it equally accounts for sensitivity and specificity. Random Forest achieved the best performance among the classical ML models, whereas EEGNet achieved the highest DL performance. An ablation study indicated that 4-second EEG epochs offered the optimal trade-off between model performance and robustness. Model interpretability was investigated using SHAP for Random Forest and occlusion sensitivity analysis for EEGNet. Both methods consistently identified delta and theta oscillations, along with centro-frontal brain regions, as key contributors to RTT classification. These findings show that both PSD-based machine learning and end-to-end deep learning methods can reliably identify RTT and highlight common neurophysiological biomarkers, offering an interpretable framework for EEG-based diagnosis and biomarker discovery for Rett Syndrome.
