Neural decoding for EEG-BCI: from conventional machine learning to deep learning models

Objective: Brain–computer interfaces (BCIs) decode neural signals to support alternative communication for individuals with motor and speech impairments, including severe speech disorders. We review EEG-BCI decoding research from 2020 to 2025 and analyze how signal-processing protocol decisions influence fair comparison and cross-paradigm generalization. Methods: We systematically reviewed EEG-BCI studies across six paradigms, summarizing commonly used datasets and experimental designs. Focusing on the conversion from raw multichannel EEG to model inputs, we analyzed preprocessing and channel selection and assessed how their implementation and reporting affect downstream modeling and evaluation. Results: Key challenges include dataset heterogeneity, low signal-to-noise ratio, and cross-subject/cross-session nonstationarity, which can confound performance comparisons. We synthesized prevailing practices in bandpass filtering, artifact handling, segmentation, and normalization, emphasizing estimating preprocessing statistics on training data only to prevent test-set leakage and optimistic bias. We also surveyed end-to-end decoding models—including convolutional and hybrid temporal networks, attention-based architectures, self-supervised pretraining, domain adaptation, multimodal fusion, and continual learning—and compared paradigm-specific metrics and evaluation units. Conclusion: Protocol decisions between raw EEG and model inputs should be treated as core experimental variables and reported transparently. Credible generalization and real-world use require leakage-aware training, paradigm-appropriate metrics, and usability-oriented evaluation criteria.