<![CDATA[Electroencephalography (EEG) allows noninvasive access to temporal brain activity; however, many EEG classifiers are tailored to a single dataset or disorder and are not easily compared across heterogeneous recording settings. A gap that this study fills is the evaluation of a fixed hybrid temporal–spatial EEG representation pipeline with an identical processing flow on public datasets for epilepsy, attention-deficit/hyperactivity disorder (ADHD), and schizophrenia. Each window passes through a one-dimensional temporal branch and a two-dimensional spatial branch. Their outputs are projected and concatenated into a 256-dimensional fused vector, then standardized and reduced to 64 components by principal component analysis (PCA) fitted only on the training data. The six post-fusion classifiers were tested under dataset-specific validation protocols, with experiment-specific blending applied where applicable, without using held-out test data for scaling, PCA fitting, model selection, threshold adjustment, or blending. Subject-wise ADHD classification achieved around 98.62% accuracy and approximately 98.60% macro-F1. Under 28-fold leave-one-subject-out evaluation, schizophrenia classification yielded an aggregated segment-level accuracy of 96.28% and a fold-averaged macro-F1 score of 73.92%. The selected intra-subject CHB-MIT chb02 experiment attained a seizure-class F1 of 92.31%. The internal four-class experiment was considered exploratory feature-space analysis rather than evidence of dataset-independent clinical diagnosis, as the classes were drawn from disparate datasets and acquisition conditions. The results provide a consistent engineering baseline across the investigated protocols and may inform future EEG decision-support studies, although broader clinical interpretation requires harmonized cohorts, broader multi-subject evaluation, and external validation.]]>
Copyrights © 2026
