Table 2 Summary of state-of-the-art models in BCI highlighting contributions, limitations, and references.

Enhanced Feature Extraction: Advanced ML techniques like Common Spatial Patterns (CSP) improve EEG feature extraction, leading to better classification accuracy

High Computational Cost: Complex ML models require significant computational resources, which may limit their real-time applicability in BCIs

Deep Learning Approaches: DL models like CNNs have shown high accuracy in classifying motor imagery tasks from EEG data

Data Requirement: DL models typically require large datasets for training, which can be challenging with the noisiness of EEG signals

Transfer Learning: Transfer learning methods have been applied to EEG classification to reduce the need for large labeled datasets, allowing models to generalize better across different subjects

Cross-Subject Variability: Models often perform poorly when applied to new subjects, due to variability in EEG signals

Real-Time Processing: Advances in hardware and optimization techniques have enabled the real-time processing of EEG signals, critical for responsive BCIs

Signal Noise: EEG signals are highly susceptible to noise from muscle movements, environmental factors, and electrode placement, which can de- grade model performance

Explainable AI (XAI): Development of interpretable models allows for a better understanding of how EEG features correlate with brain activities, enhancing the transparency of BCI systems

Interpretability vs. Performance Trade-off: Highly accurate models are often less interpretable, making them difficult to trust and understand in critical applications

Hybrid Models: Integration of ML with other signal processing methods (e.g., wavelet transforms) improves classification accuracy by capturing both time and frequency domain features

Overfitting: ML and DL models, especially those with high complexity, are prone to overfitting, particularly with small EEG datasets

Cross-Platform Compatibility: Development of more generalized models that work across different EEG acquisition devices, improving the accessibility and usability of BCIs

Device-Specific Tuning: Many models require device-specific tuning, limiting their generalizability and increasing the cost of deployment in real- world settings

Non-Invasive EEG Processing: Improved algorithms for processing non-invasive EEG signals have reduced the need for invasive methods like ECoG, making BCIs safer for users

Limited Signal Resolution: Non-invasive EEG signals typically have lower spatial resolution com- pared to invasive methods like ECoG, which limits the precision of BCIs

Improved Classification: GANs can be used to learn complex data distributions, potentially enhancing the performance of EEG signal classification by generating high-quality, realistic samples that improve model training

Computational Complexity: The process of training GANs is computationally expensive, requiring significant resources, which might limit their use in real-time or resource-constrained environments

Fractal Dimension as a discriminative feature for high accuracy classification in motor imagery EEG-based brain-computer interface

Non-Stationarity of EEG Signals: EEG signals, especially those related to motor imagery, are inherently non-stationary, meaning their statistical properties change over time. This variability can complicate the extraction of consistent features like FD, potentially affecting classification accuracy

Comparing between Different Sets of Pre-processing, Classifiers, and Channels Selection Techniques to Optimise Motor Imagery Pattern Classification System from EEG Pattern Recognition

Computational Complexity: The data-driven automatic channel selection technique, while effective, may introduce computational challenges. Implementing such methods in real-time BCI applications could be resource-intensive, can affecting system performance

A functional source separation algorithm to enhance error-related potentials monitoring in non-invasive brain-computer interface

Artifact Handling: While the FSS algorithm aims to enhance ErrP detection, the study does not extensively address the presence of various artifacts in EEG recordings, such as ocular, muscular, or environmental noise. Effective artifact removal is essential for accurate signal analysis in BCI systems

Hybrid Deep Learning (HDL)-Based Brain-Computer Interface (BCI) Systems: A Systematic Review

Limited Exploration of Preprocessing Techniques: The review highlights that preprocessing is essential for enhancing the signal-to-noise ratio (SNR) of EEG data. However, it notes that only 21.28% of the reviewed papers did not use any preprocessing or did not declare any preprocessing steps. This suggests a need for more in-depth exploration of the impact of various preprocessing methods on classification accuracy