Cross-Paradigm Consistency of EEG Decoding Pipelines: A Six-Dataset LOSO Benchmark of Handcrafted, Raw-EEG, and Topomap-Sequence Approaches

Electroencephalography (EEG) decoding pipelines are commonly tailored to individual brain–computer interface (BCI) paradigms and datasets, making dataset-specific peak performance difficult to distinguish from consistency under a shared design. We examined whether one representation–architecture pipeline, reused without paradigm-specific architectural redesign but trained separately within each dataset, could remain competitive across motor imagery (MI), P300 event-related potential (ERP/P300), and steady-state visual evoked potential (SSVEP) tasks. Common handcrafted feature families, paradigm-specific representations and specialized classical methods, two established raw-EEG neural references (EEGNet and EEG Conformer), and a prespecified topomap-sequence implementation (TopoSeqNet) were evaluated on six public datasets using within-dataset leave-one-subject-out (LOSO) cross-validation. TopoSeqNet serves as the focal implementation of the shared topomap-sequence strategy, enabling the same representation-construction rule and architecture template to be evaluated across all three paradigms. Among the five principal comparison series, TopoSeqNet ranked first on two datasets and second on four. Its equal-dataset mean chance-corrected balanced accuracy was 0.682, compared with 0.654 for the dataset-selected specific/classical reference, 0.650 for EEG Conformer, 0.648 for EEGNet, and 0.438 for the dataset-selected common reference. Because three dataset entries share the OpenBMI acquisition framework, these summaries are descriptive; an acquisition-source-weighted sensitivity analysis yielded corresponding values of 0.645, 0.630, 0.617, 0.617, and 0.365. Together, these results demonstrate competitive cross-paradigm consistency of the complete TopoSeqNet pipeline under within-dataset subject generalization and establish a focused basis for future component-level, direct-transfer, interpretability, and deployment-efficiency studies.