Editorial: Advances and challenges to bridge computational intelligence and neuroscience for brain-computer interface

Brain-computer interfaces (BCIs) enable interaction with external devices through brain signals, bypassing peripheral nerves and muscles. The field has progressed beyond laboratory settings and is highly interdisciplinary, spanning neuroscience, mathematics, computer science, and engineering. Despite abundant open-access tools, datasets, and repositories (e.g., MOABB), major challenges persist in interpreting domain-specific methods, reusing and comparing datasets, and reproducing experiments across disciplines. Critical experimental details (e.g., number and rate of SSVEP stimuli, electrode impedance, reference type, filters, labels) are often difficult for non-specialists to interpret even when present in publications. This editorial presents research that addresses these gaps by proposing unified frameworks, standardized data formats, and approaches to generalizable machine learning for BCIs.

The editorial situates its scope within prior BCI work, citing comprehensive reviews of EEG-based BCI paradigms (Abiri et al., 2019), trustworthy benchmarking initiatives like MOABB for algorithm comparison (Jayaram and Barachant, 2018), and application-focused systems such as wireless multifunctional SSVEP-based BCIs (Lin et al., 2018). It also references efforts to bridge computational intelligence and neuroscience through common descriptions of systems and data (Singh et al., 2021), highlighting the need for shared standards and reproducibility across heterogeneous datasets and methods.

As an editorial synthesizing a Research Topic, the paper highlights methods and contributions from included studies: (1) Benerradi et al. introduced BenchNIRS, a benchmarking framework for fNIRS data, enabling unbiased evaluation and best-practice methodology across five open datasets using six baseline machine-learning models (LDA, SVM, kNN, ANN, CNN, LSTM). (2) Bianchi et al. merged multiple public P300 speller datasets into a unified standardized format spanning 127 subjects, over 6,800 spelled characters, and 1,168,230 stimuli, with detailed annotations (features, labels, metadata) and an emphasis on maintaining raw data for consistent preprocessing. (3) Fox et al. trained and tested 20 SVM models using various neural features extracted from EEG, demonstrating performance variability due to non-stationarity and temporal variability in EEG signals. (4) Singh and Bianchi explored deep convolutional neural networks that encode temporal information directly from raw EEG, combining frequency and temporal features to automate analysis and reduce the need for manual preprocessing.

• BenchNIRS showed that no single machine-learning model consistently outperforms others across all fNIRS datasets; however, LDA remains a robust, reliable baseline despite its simplicity, supporting community standards in unbiased benchmarking. • The unified P300 speller dataset format by Bianchi et al. (127 subjects, >6,800 characters, 1,168,230 stimuli) with comprehensive metadata and preservation of raw data enhances data sharing, reproducibility, and consistency in preprocessing and analysis. • Fox et al. found substantial variability in SVM performance across EEG-based models due to signal non-stationarity and temporal dynamics, underscoring the need for continuous model adaptation. • Singh and Bianchi demonstrated that deep CNNs encoding temporal information, and combining frequency and time-related features, improve EEG classification accuracy and reduce manual preprocessing, supporting real-time BCI applications.

The highlighted works collectively address the central challenge of bridging computational intelligence and neuroscience in BCI by establishing shared evaluation frameworks, standardized data formats, and methods that generalize across experimental conditions. BenchNIRS advances unbiased, comparable machine-learning evaluation for fNIRS; the unified P300 dataset standardizes data sharing and metadata, improving reproducibility; and studies by Fox et al. and Singh and Bianchi tackle generalization and automation in EEG analysis. Together, these contributions reduce barriers to cross-disciplinary understanding, enable more reliable benchmarking and replication, and promote scalable, robust BCI systems in real-world settings.

Standardized machine-learning frameworks and unified data formats are pivotal for advancing BCIs. BenchNIRS and the standardized P300 dataset set precedents for community standards and reproducibility, while generalizable neural feature modeling and deep learning-based automation promise more accurate, robust, and real-time BCIs. These efforts collectively foster collaboration and innovation across disciplines, paving the way for reliable, efficient, and widely applicable BCI systems. Future directions include expanding community standards, improving metadata completeness, addressing EEG non-stationarity with adaptive models, and further integrating temporal-frequency representations for enhanced performance.

The editorial notes persistent challenges that limit interpretation and generalizability across the field: gaps in sharing domain-specific experimental details hinder reproducibility; multidisciplinary differences complicate method comparison; non-stationarity and temporal variability in EEG cause performance fluctuations, indicating the need for continuous model adaptation. While datasets and tools are increasingly available, standardized formats and comprehensive metadata remain essential to overcome these limitations.