Search papers, labs, and topics across Lattice.
This paper surveys transformer-based models for EEG motor imagery (MI) decoding, revealing that reported advances often suffer from protocol heterogeneity, inconsistent preprocessing, and non-standard data splits, leading to inflated generalization claims. Session- and subject-aware evaluations on BCIC IV 2a/2b datasets show performance clustering in the high-80% range for binary MI and mid-70% range for multi-class tasks, with only modest gains from complex hybrid architectures. The authors identify three priorities for translating transformer-driven BCIs: protocol discipline, task-relevant benchmarks for denoising, and adaptivity at scale through self-supervised pretraining and resource-aware co-optimization.
Transformer-based brain-computer interfaces (BCIs) are overhyped: current research suffers from methodological flaws leading to inflated performance, and real-world applicability remains limited.
Transformer-based models have accelerated EEG motor imagery (MI) decoding by using self-attention to capture long-range temporal structures while complementing spatial inductive biases. This systematic survey of Scopus-indexed works from 2020 to 2025 indicates that reported advances are concentrated in offline, protocol-heterogeneous settings; inconsistent preprocessing, non-standard data splits, and sparse efficiency frequently reporting cloud claims of generalization and real-time suitability. Under session- and subject-aware evaluation on the BCIC IV 2a/2b dataset, typical performance clusters are in the high-80% range for binary MI and the mid-70% range for multi-class tasks with gains of roughly 5–10 percentage points achieved by strong hybrids (CNN/TCN–Transformer; hierarchical attention) rather than by extreme figures often driven by leakage-prone protocols. In parallel, transformer-driven denoising—particularly diffusion–transformer hybrids—yields strong signal-level metrics but remains weakly linked to task benefit; denoise → decode validation is rarely standardized despite being the most relevant proxy when artifact-free ground truth is unavailable. Three priorities emerge for translation: protocol discipline (fixed train/test partitions, transparent preprocessing, mandatory reporting of parameters, FLOPs, per-trial latency, and acquisition-to-feedback delay); task relevance (shared denoise → decode benchmarks for MI and related paradigms); and adaptivity at scale (self-supervised pretraining on heterogeneous EEG corpora and resource-aware co-optimization of preprocessing and hybrid transformer topologies). Evidence from subject-adjusting evolutionary pipelines that jointly tune preprocessing, attention depth, and CNN–Transformer fusion demonstrates reproducible inter-subject gains over established baselines under controlled protocols. Implementing these practices positions transformer-driven BCIs to move beyond inflated offline estimates toward reliable, real-time neurointerfaces with concrete clinical and assistive relevance.
