Review of Performance Improvement of a Noninvasive Brain-computer Interface in Communication and Motor Control for Clinical Applications

Brain-computer interface (BCI) connects the brain and the external world by identifying brain activity and translating it into messages or commands without relying on peripheral nerves and muscles. Noninvasive electroencephalogram (EEG)-based BCI has been developed for clinical purposes, with potential applications in stroke rehabilitation, tactile communication and control for patients with impaired eye movements or vision, and prognosis for patients with cognitive motor dissociation. However, despite progress, challenges and limitations for clinical use remain.

The review traces EEG-based BCI developments in communication and motor control, highlighting preprocessing and classifier evolution. Early embedded systems recovered visual evoked potentials using FFT and SLIC with around 70% accuracy (offline). A smart multimedia controller using a wireless multichannel EEG module on a tablet achieved 71.4% online accuracy, supporting real-time applicability and portability. To reduce cost and increase portability, a custom low-cost BCI showed comparable decoding performance to conventional systems (about 70% offline using LDA). Machine learning approaches, notably SVM with features like SCSSP, MI, and LDA, improved two-class motor imagery accuracies to about 82% (offline) with minimal processing time on embedded devices. Deep learning methods further advanced performance: EEGNet (a compact CNN) yielded around 70% accuracy; FPGA-accelerated CNNs reached 80.5% and were approximately eight times faster and more power-efficient; LSTM-based models achieved up to 97.6% (offline) with substantial power savings and throughput gains. Overall, BCI systems progressed from statistical models to machine learning and deep learning, benefiting from increasing computational power and hardware accelerators. The literature also underscores a gap between offline and online performance and motivates data augmentation and generative modeling to address limited training data, especially in clinical populations.

This is a mini-review of BCI papers from January 2011 to January 2021. Selection criteria: articles indexed in the Web of Science with more than 100 citations; title, abstract, or keywords containing the phrase "brain-computer interface"; document type restricted to "article" (excluding proceedings and reviews). Performance evaluation emphasizes classification accuracy (percentage of correctly classified trials), commonly considering <70% as unacceptable and >75% as successful. The review distinguishes offline validation (using pre-recorded datasets to identify signal processing techniques) from online validation (real-time extraction and classification from live data), noting that system development often begins offline with subsequent online testing.

- Accuracy is the predominant performance metric, with >75% considered successful. Many systems meet or exceed this offline but often underperform online. - Representative application results: smart multimedia controller (online accuracy 71.4%); portable low-cost BCI vs conventional (offline decoding ~70% vs 68% with high correlation ρ=0.79); SVM-based motor imagery classification ~82.1% (offline) with 0.11 s processing; SCSSP+MI+LDA+SVM ~81.9% (offline); EEGNet ~70%; FPGA-accelerated CNN 80.5% (offline), ~8× faster with improved power efficiency; LSTM-based model 97.6% (offline), reducing power consumption by 62.7% and improving throughput power by 168%. - Offline-to-online performance drop is substantial (approximately 20%–50%). Examples: EEG-controlled FES for foot dorsiflexion, 98.8% (offline) vs 50% (online); tactile oddball paradigms SSP 64.5% (offline) vs 50.0% (online) and LSP 75.5% (offline) vs 53.0% (online); SSVEP study 97.0% (offline) vs 83.0% (online). - Data scarcity, particularly for patient populations, limits generalizable training and hinders deep learning, which typically requires large datasets. - Data augmentation and generative modeling improve performance: cDCGAN augmentation increased motor imagery CNN accuracy from 82.8% to 85.8%; selective WGAN (sWGAN) augmentation improved EEG-based emotion recognition from 83.3% to 92.2%, outperforming cWGAN (to 90.7%), sVAE (80.6%), Gaussian noise (85.8%), and rotational augmentation (75.7%). GAN-generated data preserved key motor imagery features (e.g., power variations) when compared in time–frequency analyses.

The review shows a clear trajectory from classical signal processing and statistical classifiers to machine learning and deep learning, aided by embedded and accelerator hardware that enhance speed and energy efficiency. Despite promising offline accuracies, online performance often drops markedly, emphasizing challenges such as nonstationarity, noise, user state variability, and limited training data—especially acute in clinical cohorts. The findings suggest that robust real-time performance requires not only improved architectures (CNNs, LSTMs) and hardware accelerators (FPGAs) but also strategies to mitigate data scarcity and distribution shift. Generative augmentation (e.g., GANs) demonstrably boosts accuracy by enriching training distributions with realistic synthetic EEG, narrowing the offline-to-online gap. Portability and user acceptability (e.g., wireless modules, mobile integration) further support clinical translation. However, subject-specific calibration and task-specific tuning limit versatility; progress toward subject-independent models and more generalizable latent representations is pivotal for broader clinical deployment.

BCI systems for communication and motor control have evolved substantially over the last decade, progressing from statistical methods to machine learning and deep learning, with notable gains in accuracy, efficiency, and portability. Nevertheless, achieving >90% accuracy online remains challenging, and many clinical-grade applications still suffer from substantial offline-to-online performance declines. The review highlights data scarcity as a key bottleneck and identifies data augmentation—particularly GAN-based methods—as an effective pathway to improve robustness and reduce data collection burden. Future work should focus on improving online reliability, developing subject-independent and task-generalizable models, leveraging hardware accelerators for real-time deployment, and exploring frameworks inspired by Global Workspace Theory to unify latent representations across tasks and modalities—potentially enabling more versatile, amodal BCI systems suitable for diverse clinical applications.

- Scope restricted to January 2011–January 2021 and to highly cited Web of Science articles (>100 citations), potentially introducing citation and selection bias. - Exclusion of proceedings and review papers may omit cutting-edge but less-cited or recent advances. - Many reported performances are from offline evaluations; generalizability to online, real-world clinical settings is limited and often markedly lower. - Data constraints, small sample sizes, and emphasis on healthy subjects limit applicability to patient populations. - The review synthesizes reported results without performing a meta-analysis; heterogeneity in paradigms, datasets, and metrics complicates direct comparisons.