Integrating EEG microstate dynamics in a stacked ensemble for neurodiagnostic ASD assessment

Autism Spectrum Disorder (ASD) remains diagnostically challenging due to its neurobiological heterogeneity and the current reliance on subjective behavioral assessments. To address this, we propose a novel stacked ensemble machine learning framework that enhances EEG-based ASD classification by integrating both spatial and temporal neural features. Spatial features including spectral power, functional connectivity, and signal complexity were extracted alongside temporal features derived from microstate transitions and Hidden Markov Model (HMM)-based dynamics, capturing complementary aspects of resting-state brain activity. Using Random Forest models for both base learners and the meta-classifier, our ensemble achieved a classification accuracy of 96.3% under rigorous GroupKFold cross-validation, significantly outperforming unimodal models based on spatial (88.15%) and temporal (73.6%) features alone. Bootstrapped confidence intervals confirmed the statistical robustness and generalizability of the ensemble approach. Our framework not only improves diagnostic accuracy but also lays the groundwork for translational neurotechnology aimed at early detection, subtype differentiation, and personalized intervention strategies in ASD.