Next Article in Journal
Protecting Instant Messaging Notifications against Physical Attacks: A Novel Instant Messaging Notification Protocol Based on Signal Protocol
Next Article in Special Issue
A Chronological Review of the Transmission and Effects of Mechanical Vibrations on the Hand—Arm System in an Occupational Workplace
Previous Article in Journal
Effect of Pulsed Electromagnetic Field Stimulation on Splenomegaly and Immunoglobulin E Levels in 2,4-Dinitrochlorobenzene-Induced Atopic Dermatitis Mouse Model
Previous Article in Special Issue
Genetic Algorithm-Based Optimization of Clustering Algorithms for the Healthy Aging Dataset
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Review

A Comprehensive Review on Brain–Computer Interface (BCI)-Based Machine and Deep Learning Algorithms for Stroke Rehabilitation

1
Department of Computer Science, Faculty of Computers and Informatics, Suez Canal University, 4.5 Km the Ring Road, Ismailia 41522, Egypt
2
Department of Computer Science, Faculty of Computer Science, Misr International University, 28 KM Cairo–Ismailia Road, Cairo 44971, Egypt
3
Department of Computer Science, College of Computer Engineering and Sciences, Prince Sattam bin Abdulaziz University, Al-Kharj 11942, Saudi Arabia
4
Higher Future Institute for Specialized Technological Studies, Cairo 3044, Egypt
*
Author to whom correspondence should be addressed.
These authors contributed equally to this work.
Appl. Sci. 2024, 14(14), 6347; https://doi.org/10.3390/app14146347
Submission received: 19 May 2024 / Revised: 13 July 2024 / Accepted: 18 July 2024 / Published: 21 July 2024
(This article belongs to the Special Issue Predictive Analytics in Healthcare)

Abstract

This literature review explores the pivotal role of brain–computer interface (BCI) technology, coupled with electroencephalogram (EEG) technology, in advancing rehabilitation for individuals with damaged muscles and motor systems. This study provides a comprehensive overview of recent developments in BCI and motor control for rehabilitation, emphasizing the integration of user-friendly technological support and robotic prosthetics powered by brain activity. This review critically examines the latest strides in BCI technology and its application in motor skill recovery. Special attention is given to prevalent EEG devices adaptable for BCI-driven rehabilitation. The study surveys significant contributions in the realm of machine learning-based and deep learning-based rehabilitation evaluation. The integration of BCI with EEG technology demonstrates promising outcomes for enhancing motor skills in rehabilitation. The study identifies key EEG devices suitable for BCI applications, discusses advancements in machine learning approaches for rehabilitation assessment, and highlights the emergence of novel robotic prosthetics powered by brain activity. Furthermore, it showcases successful case studies illustrating the practical implementation of BCI-driven rehabilitation techniques and their positive impact on diverse patient populations. This review serves as a cornerstone for informed decision-making in the field of BCI technology for rehabilitation. The results highlight BCI’s diverse advantages, enhancing motor control and robotic integration. The findings highlight the potential of BCI in reshaping rehabilitation practices and offer insights and recommendations for future research directions. This study contributes significantly to the ongoing transformation of BCI technology, particularly through the utilization of EEG equipment, providing a roadmap for researchers in this dynamic domain.

1. Introduction

The goal of rehabilitation is to repair injured muscles and motor functions through therapy and training. brain–computer interface (BCI) technology is an emerging field in neurotechnology. Applications for BCI have been used in many different fields to aid people with neuromuscular conditions such as stroke and spinal cord injuries to enhance their overall quality of life. One possible replacement for current prosthesis technology, assistance, and rehabilitation might be a system integrating BCI, robotics, neuroscience, machine learning, and deep learning. BCI may also help to improve or restore the motor skills that the brain has lost. BCI provides easy-to-use technological assistance and robotic prostheses, powered by brain activity. With BCI, a user’s brain activity may be translated into the output command needed to perform any predefined operation on appropriate equipment [1].
More advanced neurophysiologically driven solutions that support behavioral training and recovery are the foundation of rehabilitation technology. For example, depending on the neuromotor activity of the operand, a robot-guided system may assist in the motion of a damaged limb [2]. For stroke survivors, BCI enables neurological stimulation and improvement of neuromotor outcomes [3].
The main purpose of BCI applications is to support people in their everyday lives who have severe motor impairments. A number of BCI-capable gadgets have been created to assist with rehabilitation and daily tasks for people [4]. Furthermore, a large number of investigations have confirmed that BCI can help those with severe impairments, such as paralysis. BCI provides direct brain-to-technology contact, which might aid in the rehabilitation of a handicapped person afflicted with musculoskeletal disorders. The location of the electrodes determines whether a method of monitoring brain activity is invasive or non-invasive. One or more BCI devices are utilized for direct electrode placement in the brain cavity to monitor the brain area during neurosurgical procedures that require invasive operations [5]. Although this process produces very high-quality signals, it has a significant negative impact since it increases the amount of scar tissue in the brain [6]. An example of an invasive technology is the Electrocorticograph (ECoG), which measures brain activity directly from the surface of the brain [7].
Compared to the invasive method, a non-invasive electroencephalogram (EEG) is more precise and less costly, but less pleasant [4]. For example, an EEG tracks and records waves of electrical activity in the brain by placing electrodes on the scalp. These signal waves are transmitted to a computing system based on the discovered patterns in the produced waves. However, from the view of clinical conditions and wearable electronics, implanted devices are more precise due to direct contact with target biological tissues. For instance, in heart and hemodynamic monitoring, implanted devices provide higher precision and reliability. Despite this precision, non-invasive methods offer significant advantages. Non-invasive EEG is more accessible and can be performed in various settings, including at home or in outpatient clinics, without the need for surgical procedures. This reduces the risk of complications such as infections and scarring, which are associated with invasive methods. Additionally, non-invasive EEG is more comfortable for patients, especially for those who may be apprehensive about surgical interventions. The lower cost of non-invasive EEG also makes it a more feasible option for widespread use, especially in resource-limited settings. While it may be less pleasant due to the need for repeated application of electrodes and the potential for discomfort during prolonged sessions, advancements in EEG cap designs and gel-free electrodes are continually improving patient comfort and usability. Non-invasive EEG thus represents a balance between precision, cost, and patient safety, making it a valuable tool in both clinical and research settings.
In the primary framework of EEG, the fundamental techniques for capturing signals involve motor imagery (MI) and steady-state visual evoked potential (SSVEP) [8]. These two paradigms offer many methods and potentials. Motor imagery focuses on the psychological foundations of any actions that do not need the activation of muscles [9]. With BCI, a person may see particular motions, like grasping an object, and their brain will convey the command to the controlling equipment that makes these movements happen. Several methods for assisting patients with decreased motor control were provided [10]. Patients are taught to produce more motor brain impulses in the first strategy and to activate tools that improve motor performance in the second. Despite the fact that individuals with acquired motor impairments sometimes have problems with motor connections, the EEG method shows remarkable improvements and continuous changes.
The authors of [11] examined 16 chronic stroke patients who utilized a brain–computer interface to obtain input on arm and hand orthotics. Assistive technology helps people with physical limitations engage in a variety of activities, including playing, moving around, and having normal conversations. This device can reduce the stress that carers feel while thinking about individuals with disabilities [11,12,13].
Figure 1 shows a generic approach that researchers take when designing a BCI- and motor imagery-based stroke rehabilitation system, and how researchers employ each part of the diagram is illustrated in more detail in the upcoming sections.
This comprehensive review provides an overview of BCI-based machine learning and deep learning algorithms coupled with EEG and motor imagery for stroke rehabilitation. It is organized as follows: In Section 2, a brief background about the topic is given. In Section 3, the diverse applications of EEG, machine learning, and deep learning in rehabilitation used by researchers are introduced. How EEG signals are acquired and how motor imagery training for patients is conducted are elaborated on in Section 4. The signal processing and classification approaches are explained in Section 5. Section 6 includes a discussion of the current limitations in research related to rehabilitation using EEG with ML and DL methods. In Section 7, a comparison of EEG headsets for rehabilitation purposes and different datasets used by the researchers included in this comprehensive review are introduced. Section 8 provides the results of EEG and motor imagery-based studies in the last 7 years. Section 9, the last part of the comprehensive review, is a brief conclusion of the contents of the paper. Figure 1 helps in illustrating what will be discussed later in the review.

2. Background

The human brain, with its remarkable plasticity, possesses the extraordinary ability to rewire itself and forge new neural connections even after injury or disease. This ability, called neuroplasticity, makes BCIs a promising tool for brain rehabilitation [14,15,16,17,18,19,20].
There is a kind of teamwork between BCIs and machine learning that could change the method of brain rehabilitation for people with stroke, spinal cord injuries, brain injuries, and other neurological conditions. It could help people regain lost skills and improve their quality of life. That is why scientists are keen on exploring new things in this technology for the following reasons:
  • Unlocking the potential of brain–computer interfaces (BCIs) for brain rehabilitation goes hand in hand with the remarkable power of machine learning. These intricate algorithms act as digital guides, navigating the complex landscape of brain signals to reveal the hidden pathways to recovery. By learning to recognize the unique patterns within each person’s neural code, machine learning techniques like support vector machines, deep neural networks, and random forests become adept at deciphering intentions and translating them into tangible actions. This remarkable synergy opens a world of possibilities for individualized rehabilitation, allowing us to harness the power of our own minds to regain control and rebuild skills [21,22,23,24].
  • BCIs are able to adapt to variations in a user’s brain impulses over time because of machine learning. Increasing accuracy and resilience, adaptive models continuously modify their parameters in response to fresh data [20,25,26,27].
  • Tailored rehabilitation interventions are made possible by machine learning. Rehabilitative techniques can be adjusted as necessary by training models to identify each user’s unique brain patterns [7,11,19,28,29,30].
  • Instant feedback: Machine learning can process brain signals very quickly, so BCIs can react in real time. This is like having a virtual coach that gives you feedback right away, helping you learn faster [14,17].
  • Predicting success: Machine learning can look at brain activity and predict how well someone will do in rehab. This helps doctors make better treatment plans and obtain better results [20].
  • Seeing the whole picture: Machine learning can combine different types of brain data, like EEGs and fMRIs, to obtain a more complete understanding of how the brain is working during rehab [17,20].

3. Diverse Applications of EEG, Machine Learning, and Deep Learning in Rehabilitation

Embarking on an exploration of neural rehabilitation, the research delves into the multifaceted activities that can be addressed through the mutual application of EEG, machine learning, and deep learning [13,14,19,31,32,33]. These innovative interventions bring about transformative possibilities across various neural rehabilitation domains.
  • Motor Rehabilitation: The utilization of machine learning algorithms in decoding EEG signals during motor imagery tasks opens avenues for controlling external devices, including robotic exoskeletons and prosthetics [13,19,31,33]. Systems providing feedback in real time not only instruct users through precise motor tasks but also contribute significantly to motor skill relearning and the promotion of neuroplasticity [14,19,32,33].
  • Cognitive Rehabilitation: EEG signals play a pivotal role in gauging and enhancing attention stages using Neurofeedback techniques [8,11,13,34]. The adaptability of machine learning facilitates the tailoring of training procedures tailored to individual mental states, while the nuanced capabilities of deep learning contribute to the design of personalized memory training tasks [7,8,13,34].
  • Neuropsychiatric Rehabilitation: EEG neurotraining emerges as a beneficial instrument for handling stress and anxiety, utilizing artificial intelligence to recognize patterns associated with stress and triggering interventions for relaxation [10,28,35,36]. Moreover, neurotraining based on EEG assists those with ADHD in refining concentration and controlling attention by reinforcing preferred brain activity patterns [13].
  • Walking Rehabilitation: Integrating ML and EEG with systems that capture movements allows for a comprehensive analysis of gait patterns, providing real-time feedback during walking exercises and potentially revolutionizing rehabilitation approaches [13,31,34,37].
  • Visual and Hearing Restoration: Protocols relying on EEG, in collaboration with advanced machine learning approaches, provide an intricate structure for creating customized training activities for individuals experiencing sensory perception challenges in both vision and hearing [32,38].
  • Multimodal Rehabilitation: The integration of EEG with other cutting-edge technologies, such as virtual reality or functional near-infrared spectroscopy, paves the way for innovative multimodal rehabilitation approaches [19,20,33,39,40].
The inherent flexibility of these technologies holds significant promise for not only enhancing rehabilitation outcomes but also substantially improving the overall quality of life for individuals grappling with neurological disorders.

4. EEG Signal Acquisition and Motor Imagery Training

4.1. EEG-Based Signal Acquisition

Accessing the Neural Network: The interface between brain and world is established through electrodes, either in non-invasive surface contact or via implanted subdermal conduits. The non-invasive method relies on conductive gel bridging the gap with the scalp, facilitating signal acquisition but introducing setup time and low-frequency attenuation. An alternative, dry electrodes, offer swift application but increased susceptibility to movement artifacts [41]. EEG data acquisition employs two primary modes: unipolar and bipolar. In unipolar mode, each electrode relays its information relative to a common reference, creating individual EEG channels. Conversely, bipolar mode focuses on potential differences between specific electrode pairs, each pair constituting a channel [42]. To ensure standardized data gathering and inter-session comparability, the international 10–20 system governs electrode placement across the scalp. This systematic approach fosters reliable communication and collaborative efforts within the field of brain–computer interfaces [43].
Figure 2 reveals the international 10–20 system, a topographical blueprint for deciphering the brain’s electrical signatures. Similar to carefully arranged instruments in an orchestra, this framework meticulously guides the placement of EEG sensors across the scalp. Once these sensitive outposts detect the subtle fluctuations in cerebral potential, the signal undergoes amplification, a process similar to tuning in to the faintest whispers of the neural ensemble. Subsequently, this amplified signal is adapted into the digital domain, analogous to transcribing the intricate melody of brain activity into a language computers can comprehend. By employing higher sampling rates, measured in Hertz, researchers can capture more fleeting neural fluctuations. Increasing the number of EEG sensors, analogous to expanding the observation network, provides a broader spatial perspective on brain activity. This combination of greater temporal and spatial resolution allows for a more detailed and accurate representation of the dynamic processes unfolding within the cerebral cortex.

4.2. Patient Training on Motor Imagery Tasks

The Graz training paradigm is a powerful method for teaching individuals to control their EEG signals through motor imagery (MI). In essence, it is a dedicated training program for the brain, analogous to how athletes train their muscles to perform specific tasks. Through repeated practice, users learn to produce distinct EEG patterns associated with different imagined movements, like clenching a fist or moving a foot [44,45].
Here is a breakdown of the Graz training process:
  • Calibration Phase:
    • Warm-up (t = 0 s): The user focuses on a fixation point or cue to settle their mind and prepare for the upcoming task. MI Task: A visual cue, such as an arrow, instructs the user to perform a specific MI task (e.g., imagine left-hand movement).
    • Data Collection (t = 3 s): During the task period, EEG sensors capture the user’s brain activity, recording the unique electrical signature of the imagined movement.
    • Machine Learning (t = 3–7 s): The collected data are then analyzed by a machine learning algorithm. This algorithm identifies the key features that distinguish different MI tasks from the EEG signals.
    • No Feedback: In the initial phase, no feedback is provided to the user. This allows the machine learning algorithm to focus solely on understanding the user’s individual EEG patterns.
  • Feedback Phase:
    • MI Task Repetition: Once the system has been calibrated, users repeat the MI tasks. Real-time Feedback: This time, the system provides feedback in real time. For example, a bar might grow longer or change color based on the system’s confidence in recognizing the current MI task.
    • Refinement and Repetition: With each trial, the user receives feedback and can adjust their mental strategies to produce clearer EEG patterns. This iterative process strengthens the connection between the imagined movement and the corresponding EEG signature.
  • Multiple Sessions:
    • Gradual Improvement: The Graz training paradigm involves multiple training sessions, spread over days or weeks. With each session, the user’s ability to generate distinct EEG patterns for different MI tasks improves, leading to more accurate recognition by the system.
    • Customization: The training protocol can be adapted to individual needs and goals. The specific MI tasks, cues, and feedback types can be tailored to suit different applications, such as controlling a prosthetic limb or navigating a virtual environment.
Figure 3 illustrates the timeline of a single Graz training trial, highlighting the key stages and timings [45]. By actively engaging in this training process, users can unlock a powerful tool for brain–computer interaction. The Graz training paradigm empowers individuals to harness the potential of their own minds, opening doors to exciting possibilities in various fields.

5. EEG Signal Processing and Classification Techniques in Rehabilitation Research

During EEG acquisition, unwanted signals known as artifacts contaminate the desired cerebral activity. These electrical interferences can originate from within the body (endogenous) or from external sources (exogenous) [46]. Common endogenous artifacts include eye blinks, muscle activity, and cardiac rhythm, while exogenous artifacts encompass power line interference, equipment noise, and environmental influences. To isolate the genuine brain signal, preprocessing techniques like temporal and spatial filtering are employed. Temporal filtering targets repetitive patterns like heartbeats and eye blinks, while spatial filtering leverages the spatial distribution of electrodes to localize and suppress artifacts. By meticulously removing these extraneous contributions, researchers can arrive at a clean EEG signal, paving the way for accurate interpretation and diverse applications in brain research and neurotechnology, and this is performed in the preprocessing phase.
Moving to the next phase, EEG-based motor imagery (MI) generates abundant data streams thanks to high sampling rates and numerous electrodes. However, maximizing performance hinges upon extracting concise, informative representations capable of distinguishing intentional MI activity from background brain noise. These distilled components, termed “features”, emerge through “feature extraction”, a formal process wherein preprocessed, voluminous EEG data undergoes transformation into a dedicated “feature space”. This meticulously constructed environment encapsulates all essential discriminatory information, empowering classifiers to fulfill their role with exceptional accuracy.
Within the expansive landscape of neural rehabilitation research, an array of sophisticated EEG signal processing techniques takes center stage, employed to meticulously analyze and extract meaningful insights from EEG data [17,18,19,20,33,38,47,48,49,50,51,52].
  • Filtering: The application of low pass, high pass, and notch filters emerges as a crucial step in refining EEG signal quality, as it effectively eliminates undesirable frequency components [38,47].
  • Artifact Elimination: Approaches like Independent Component Analysis (ICA) and Principal Component Analysis (PCA) play a crucial role in differentiating and removing disturbances, encompassing eye blinks, muscle movements, and interference from electrocardiograms [17,18,48].
  • Time Domain Features: The extraction of characteristics like average amplitude, root mean square value, and signal variance represents a nuanced approach to capturing the temporal characteristics inherent in EEG signals [19,20,33].
  • Frequency Domain Features: Insights into the frequency distribution of brain activity are unveiled through the meticulous examination of power spectral density, spectral entropy, and band power ratios [11,31,47].
  • Time-Frequency Features: Methods such as wavelet transformation and short-time Fourier transformation enhance complexity by unveiling the dynamic variations in EEG signal characteristics across both time and frequency domains [32,49,50].
  • Functional Interconnection: Measures encompassing coherence, phase synchronization, and mutual information serve as invaluable tools in assessing the intricate functional relationships between different brain regions [33,53,54].
  • Graph Theory Analysis: The innovative representation of EEG data as networks, coupled with graph theory metrics, offers a unique lens through which organizational and communication patterns within the brain can be deciphered [51,52].
  • Pattern Recognition and Motor Imagery: A specialized focus on processing EEG signals derived from motor imagery tasks facilitates the recognition of specific patterns associated with imagined movements, thereby paving the way for tailored interventions [26,55,56].

5.1. Feature and Channel Selection

High-density EEG recordings generate voluminous data, creating a challenge for BCI applications. Channel selection techniques address this by identifying a smaller subset of optimal electrodes, maximizing classification accuracy and reducing computational cost. Similar to feature selection, candidate channel sets are generated and evaluated with dedicated criteria, ultimately prioritizing channels harboring the most task-relevant information. Filter methods rely on the statistical analysis of EEG properties, while wrapper methods directly test channel performance within the BCI system. In some cases, hybrid approaches combining both strategies are employed for optimal channel selection. This meticulous pruning of EEG data lays the foundation for accurate BCI operation and enhances its real-world potential.
  • Filter Approach: Initiating with the full set of features, filter methods meticulously identify the optimal subset through dedicated selection criteria. These criteria often revolve around key characteristics like information gain, dependency, consistency, correlation, and distance measures [57]. A significant advantage of filter methods lies in their minimal computational requirements. Additionally, the feature selection process operates independently of the chosen classifier, providing greater flexibility. Widely utilized filter methods include correlation criteria and mutual information techniques, both meticulously honing in on the most informative features within the data landscape.
  • Wrapper Approach: Distinct from filter methods, wrapper approaches forge a direct partnership with the classifier to select features meticulously. They iteratively present candidate feature subsets to the classifier, diligently evaluating its performance. This feedback loop guides the selection process, prompting either acceptance of a subset based on established criteria or the proposal of new combinations for further evaluation. Algorithms within this realm encompass searching algorithms and evolutionary algorithms. The former embarks on their quest with an empty set, strategically adding or removing features until the classifier’s performance peaks. Their journey typically concludes when a designated maximum feature subset size is attained. Meanwhile, evolutionary algorithms, such as particle swarm optimization (PSO) [58], and artificial bee colony (ABC) [59,60], harness nature-inspired optimization techniques to uncover optimal subsets. While wrapper methods excel at identifying feature combinations that yield superior classifier performance compared to filter methods, their computational demands are considerable, rendering them less suitable for handling vast datasets.

5.2. EEG-Based Machine Learning and Deep Learning Algorithms

Within the intricate realm of EEG-based neural rehabilitation, a spectrum of both machine learning and deep learning methods unfolds, each contributing to the nuanced analysis of EEG data and the formulation of interventions designed to bolster neural function and expedite recovery [61]. The following sections compare these two typical algorithms in detail, exploring their specific applications, advantages, and drawbacks from practical perspectives, such as sample size, accuracy, and training cost.
  • Machine Learning Techniques:
    Machine learning techniques like support vector machines (SVM) and k-nearest neighbors (k-NN) have been effectively utilized in EEG-based rehabilitation.
    Support Vector Machines (SVM):
    SVMs are robust classifiers known for their precision in distinguishing between different motor tasks or targets based on EEG data. They are particularly effective with small-to-medium-sized datasets due to their ability to find the optimal hyperplane that separates different classes.
    • Sample Size: Effective with small-to-medium-sized datasets.
    • Accuracy: High accuracy in binary and multi-class classification problems as Figure 4.
    • Training Cost: Moderate, with a need for tuning hyperparameters.
    Pseudocode for SVM:
        Initialize parameters and load EEG dataset
        Preprocess the data (filtering, normalization)
        Define kernel function (linear, RBF, etc.)
        Split data into training and testing sets
        Train SVM model on training data
        Evaluate model on testing data
        Tune hyperparameters for optimal performance
        Output classification accuracy
    k-Nearest Neighbors (k-NN):
    k-NN algorithms are simple yet powerful tools for identifying patterns in brain activity related to cognitive performance. They work by comparing a new data point to the k-nearest data points in the training set, illustrated in Figure 5.
    • Sample Size: Suitable for small datasets.
    • Accuracy: Good for pattern recognition tasks.
    • Training Cost: Low, as k-NN is a lazy learner.
    Pseudocode for k-NN:
        Initialize parameters and load EEG dataset
        Preprocess the data (filtering, normalization)
        Define value of k
        Split data into training and testing sets
        For each data point in the testing set:
            Calculate distance to all points in the training set
            Identify k-nearest neighbors
            Assign class label based on majority vote
        Evaluate accuracy of the classifier
  • Deep Learning Techniques:
    Deep learning techniques such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory networks (LSTMs), and Multi-Layer Perceptrons (MLPs) are employed in EEG-based rehabilitation for their superior ability to automatically extract features and handle complex data.
    Convolutional Neural Networks (CNNs):
    CNNs are particularly effective in spatial data analysis (see Figure 6) and have been successfully applied to EEG signal classification by extracting spatial features from EEG recordings.
    • Sample Size: Requires a large dataset.
    • Accuracy: High for spatial feature extraction and classification.
    • Training Cost: High due to the need for extensive computational resources.
    Pseudocode for CNN:
        Initialize parameters and load EEG dataset
        Preprocess the data (filtering, normalization)
        Define CNN architecture (number of layers, filters)
        Split data into training and testing sets
        Train CNN model on training data
        Evaluate model on testing data
        Tune hyperparameters for optimal performance
        Output classification accuracy
    Recurrent Neural Networks (RNNs):
    RNNs as Figure 7, especially those using LSTM or GRU units, are suitable for temporal sequence prediction, making them ideal for processing sequential EEG data.
    • Sample Size: Requires a large dataset.
    • Accuracy: High for sequential data analysis.
    • Training Cost: High due to recurrent nature and complex computations.
    Pseudocode for RNN:
        Initialize parameters and load EEG dataset
        Preprocess the data (filtering, normalization)
        Define RNN architecture (number of layers, units)
        Split data into training and testing sets
        Train RNN model on training data (sequence data)
        Evaluate model on testing data
        Tune hyperparameters for optimal performance
        {Output sequence prediction accuracy}
    Long Short-Term Memory Networks (LSTMs):
    LSTMs are a type of recurrent neural network capable of learning long-term dependencies, depicted in Figure 8. They excel in tasks that involve sequential data, such as learning motor sequences from EEG signals.
    • Sample Size: Requires a large dataset.
    • Accuracy: High for temporal pattern recognition.
    • Training Cost: High due to complex architecture.
    Pseudocode for LSTM:
        Initialize parameters and load EEG dataset
        Preprocess the data (filtering, normalization)
        Define LSTM architecture (number of layers, units)
        Split data into training and testing sets
        Train LSTM model on training data (sequence data)
        Evaluate model on testing data
        Tune hyperparameters for optimal performance
        Output sequence prediction accuracy
    Multi-Layer Perceptrons (MLPs):
    MLPs are a type of feedforward neural network that can be used for classification tasks, explained in Figure 9. They are simpler compared to CNNs and RNNs but still effective for basic EEG data classification.
    • Sample Size: Effective with medium to large datasets.
    • Accuracy: Moderate to high, depending on the complexity of the task.
    • Training Cost: Moderate.
    Pseudocode for MLP:
        Initialize parameters and load EEG dataset
        Preprocess the data (filtering, normalization)
        Define MLP architecture (number of layers, neurons)
        Split data into training and testing sets
        Train MLP model on training data
        Evaluate model on testing data
        Tune hyperparameters for optimal performance
        Output classification accuracy
Both machine learning and deep learning techniques offer unique advantages in EEG-based rehabilitation. Machine learning algorithms like SVM and k-NN are advantageous for their simplicity and effectiveness with smaller datasets, providing reliable classification and pattern recognition. On the other hand, deep learning algorithms such as CNNs, RNNs, LSTMs, and MLPs excel in automatically extracting complex features from large datasets, albeit at a higher computational cost and requiring more extensive training data. Choosing the appropriate algorithm depends on the specific requirements of the rehabilitation task, including the available dataset size, the need for accuracy, and the computational resources at hand.

5.3. Performance Indicators and Metrics in Evaluating the Effectiveness of Methods

In assessing the efficacy of neural rehabilitation methods utilizing EEG, a rich array of performance indicators comes to the forefront, each offering a unique lens through which the impact of these interventions can be comprehensively evaluated [11,13,14,19,31,38,38].
  • Classification Accuracy: The efficacy of EEG-based models is scrutinized through the lens of classification accuracy, providing insights into their ability to distinguish between different classes or states in activities like differentiating motor imagery or identifying cognitive states [11,13,14,19,31,38,38].
  • ROC Curve and AUC: The evaluation of the balance between specificity and sensitivity in classification assignments is facilitated through the utilization of ROC curves and AUC values, adding a layer of sophistication to the assessment process [17,38].
  • MSE or RMSE: In the realm of regression functions, the metrics of Mean Squared Error (MSE) or Root Mean Squared Error (RMSE) emerge as crucial, providing a nuanced measure of the accuracy of predictions by quantifying the difference between predicted and actual values [13,31].
  • R-squared (R2): The extent to which the regression model fits the data is gauged through the lens of R-squared (R2), offering valuable insights into the predictive power of the model [49,50].
  • Real-time Performance Measurements: In situations necessitating instantaneous responsiveness, parameters such as response latency, response time, and overall system latency offer a comprehensive evaluation of the system’s real-world applicability [26,55].

6. Present Constraints in the Ongoing Research on Rehabilitation Utilizing EEG with ML and DL Methods

While the landscape of EEG-based neural rehabilitation research with machine learning and deep learning showcases remarkable progress, it is essential to acknowledge and navigate through existing limitations and gaps, each presenting a unique challenge to the advancement of this field.
  • Noise and Artifacts: Concerns surrounding data quality, preprocessing methodologies, and the standardization of data gathering protocols cast a spotlight on the imperative need to address these aspects to ensure the reliability and consistency of results.
  • Small Sample Sizes: The challenges when acquiring high-quality EEG data from groups of patients contribute to the prevalence of small sample sizes, potentially leading to model overfitting and hindering the generalizability of findings.
  • Longitudinal EEG Datasets: The scarcity of longitudinal EEG datasets poses a significant hurdle in monitoring progress during neural rehabilitation. A dedicated focus on long-term research is indispensable for comprehensively understanding the effectiveness of diverse approaches and customizing interventions and treatments accordingly.
  • Interpretability of Deep Learning: The opaque nature of deep learning models poses challenges in interpreting results, necessitating research that seamlessly combines deep learning methodologies with insights from neuroscience. This integration is crucial for gaining a deeper understanding of the fundamental neurophysiological mechanisms associated with brain rehabilitation.
  • Ethical Concerns in Real-Time Applications: While offline analysis dominates several EEG-based brain rehabilitation techniques, the shift towards real-time applications introduces ethical considerations related to patient consent, data ownership, and privacy. Meticulous attention is required to ensure that these issues are addressed with the utmost care, respecting patients’ rights.
  • Bridging the Gap Between Research and Clinical Implementation: Despite the strides made in research in neurological rehabilitation utilizing EEG signals, there remains a discernible discrepancy between academic research and the practical implementation of clinical solutions. Efforts to bridge this gap are essential for the seamless translation of research findings into real-world clinical practices.

7. Comparison of EEG Headsets for Rehabilitation Purposes with Various Datasets

A primary constraint of the traditional EEG-based BCI is the small sample size used in this investigation. This review examined datasets that were based on electroencephalography. Using criteria for searching, over 60 research studies were found, and around half of these studies have been carried out using the BCI competition dataset. Additionally, as Figure 10 and Table 1 demonstrate, these are the percentages of datasets included in this review.
The BCI competition IV is a benchmark dataset in the field of brain and human computer interfaces. Data were acquired from nine subjects, and two sessions were recorded for each subject. Each session consists of six runs, and each run consists of 48 trials, 12 for each of the four classes (right-hand, feet, left-hand, tongue). In total, there were 288 trials per session. The dataset was already preprocessed using Bandpass filtering (0.5 Hz, 100 Hz) and Notch filtering (50 Hz). The former preprocessing technique is used to filter a frequency band inside a signal while minimizing frequencies outside of that range and the latter technique is used to eliminate a specific range of frequencies such as powerline interference. Data were recorded using 22 EEG channels and 3 EOG channels (which measure the influence of eye movement).

8. Most Significant EEG- and Motor Imagery-Based Studies in the Last 7 Years

Table 2 lists the results of the most significant EEG and motor imagery-based studies using deep learning methodologies to classify between different motor simulation tasks.
In recent years, several notable studies have emerged in the field of EEG and motor imagery-based deep learning models for stroke rehabilitation. For instance, ref. [85] introduced an Adaptive CNN model applied to the BCI competition IV dataset 2a, achieving a classification accuracy of 93.20%. This approach dynamically adjusts convolutional layers to enhance feature extraction from EEG signals, showing promise in handling variability in EEG data. Similarly, ref. [86] proposed an attention-based CNN model, tested on data collected from 30 subjects, which achieved 88.75% accuracy. The attention mechanism improves the model’s ability to focus on relevant features, enhancing classification performance.
Further advancements include [87], which utilized a Graph-CNN on the BCI competition IV dataset 2b, achieving an accuracy of 89.60%. This study highlights the effectiveness of graph-based methods in capturing spatial relationships in EEG data. Ref. [88] presented a Hybrid CNN-RNN model applied to the PhysioNet EEG MI Dataset, with a notable accuracy of 94.50%. This hybrid model leverages the strengths of both CNNs and RNNs in feature extraction and temporal sequence learning, respectively.
Lastly, ref. [89] introduced a Transformer-based model for EEG classification on the BCI competition IV dataset 2a, achieving an accuracy of 92.30%. The Transformer architecture, known for its success in natural language processing, demonstrates its potential in EEG signal analysis by effectively capturing long-range dependencies.
These recent studies underscore the rapid advancements in applying deep learning techniques to EEG and motor imagery data, pushing the boundaries of stroke rehabilitation research and offering new avenues for enhancing patient outcomes.

9. Conclusions

The field of brain–computer interface (BCI) research holds potential for medical uses and presents neurophysiological proof supporting BCI-triggered neuroplastic modifications. Nevertheless, there is a shortage of definitive clinical studies validating the efficacy of BCI interventions, hindering their incorporation into universally recognized clinical practices. BCI systems display varied design features, and preliminary brain priming before intervention has been proven to enhance rehabilitation outcomes. The convergence of BCI-driven robotics with complementary methods like BCI-linked neuromuscular stimulation unlocks promising avenues, particularly in stroke rehabilitation. Notably, the emergence of BCI-controlled soft robots presents exciting possibilities in this domain. By seamlessly integrating BCIs into rehabilitation programs, individuals can be empowered by restoring independence for those facing cognitive or physical limitations. Moreover, the versatility of EEG equipment extends beyond rehabilitation; it holds tremendous potential for both people without health issues and those with disabilities in various everyday scenarios. For non-invasive BCI applications to truly flourish, developers must prioritize market needs and design with the end user in mind. Integrating wireless solutions not only enhances user convenience but also bolsters viability for sustained involvement over an extended period and use in outdoor settings. However, ensuring robust data security and conducting rigorous user experience assessments are cornerstones for widespread BCI adoption across diverse industries.
In conclusion, while BCI technology holds undeniable promise, particularly in rehabilitation settings, ongoing research and a user-centric approach are critical to maximizing its effectiveness and broadening its impact. By prioritizing market needs, embracing wireless solutions, and ensuring data integrity and user satisfaction, BCI technology can be empowered to bridge the gap between scientific potential and real-world applications.

Author Contributions

Conceptualization, W.H.E. and A.A. (Abdelrahman Ayman); methodology, M.A.; software, H.M.; validation, S.E.M. and H.A.; formal analysis, W.H.E.; investigation, A.A. (Ahmed Ali); resources, Y.T.; data curation, M.A.; writing—original draft preparation, H.M.; writing—review and editing, H.M.; visualization, H.A.; supervision, W.H.E.; project administration, W.H.E.; funding acquisition, A.A. (Ahmed Ali). All authors have read and agreed to the published version of the manuscript.

Funding

This study is supported by funding from Prince Sattam bin Abdulaziz University (project number PSAU/2024/R/1445).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

No new data were created or analyzed in this study.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. McFarland, D.J.; Wolpaw, J.R. Brain-computer interfaces for the operation of robotic and prosthetic devices. In Advances in Computers; Elsevier: Amsterdam, The Netherlands, 2010; Volume 79, pp. 169–187. [Google Scholar] [CrossRef]
  2. Dobkin, B.H. Brain-computer interface technology as a tool to augment plasticity and outcomes for neurological rehabilitation. J. Physiol. 2007, 579, 637–642. [Google Scholar] [CrossRef]
  3. Dodd, K.C.; Nair, V.A.; Prabhakaran, V. Role of the contralesional vs. ipsilesional hemisphere in stroke recovery. Front. Hum. Neurosci. 2017, 11, 469. [Google Scholar] [CrossRef]
  4. Fouad, M.M.; Amin, K.M.; El-Bendary, N.; Hassanien, A.E. Brain computer interface: A review. In Brain-Computer Interfaces: Current Trends and Applications; Springer: Berlin/Heidelberg, Germany, 2015; Volume 74, pp. 3–30. [Google Scholar] [CrossRef]
  5. Birbaumer, N.; Ghanayim, N.; Hinterberger, T.; Iversen, I.; Kotchoubey, B.; Kübler, A.; Perelmouter, J.; Taub, E.; Flor, H. A spelling device for the paralyzed. Nature 1999, 398, 297–298. [Google Scholar] [CrossRef]
  6. Mondal, D.; Alagirisamy, M. Brain computer interface (BCI): Mechanism and challenges–A survey. Int. J. Pharm. Res. 2020, 12, 2200–2208. [Google Scholar] [CrossRef]
  7. Wilson, J.A.; Felton, E.A.; Garell, P.C.; Schalk, G.; Williams, J.C. ECoG factors underlying multimodal control of a brain-computer interface. IEEE Trans. Neural Syst. Rehabil. Eng. 2006, 14, 246–250. [Google Scholar] [CrossRef]
  8. Rezeika, A.; Benda, M.; Stawicki, P.; Gembler, F.; Saboor, A.; Volosyak, I. Brain–computer interface spellers: A review. Brain Sci. 2018, 8, 57. [Google Scholar] [CrossRef]
  9. Hwang, H.J.; Kwon, K.; Im, C.H. Neurofeedback-based motor imagery training for brain–computer interface (BCI). J. Neurosci. Methods 2009, 179, 150–156. [Google Scholar] [CrossRef]
  10. Remsik, A.; Young, B.; Vermilyea, R.; Kiekhoefer, L.; Abrams, J.; Evander Elmore, S.; Schultz, P.; Nair, V.; Edwards, D.; Williams, J.; et al. A review of the progression and future implications of brain–computer interface therapies for restoration of distal upper extremity motor function after stroke. Expert Rev. Med. Devices 2016, 13, 445–454. [Google Scholar] [CrossRef]
  11. Ramos-Murguialday, A.; Broetz, D.; Rea, M.; Läer, L.; Yilmaz, O.; Brasil, F.L.; Liberati, G.; Curado, M.R.; Garcia-Cossio, E.; Vyziotis, A.; et al. Brain-machine interface in chronic stroke rehabilitation: A controlled study. Ann. Neurol. 2013, 74, 100–108. [Google Scholar] [CrossRef] [PubMed]
  12. Kitchenham, B.; Brereton, O.P.; Budgen, D.; Turner, M.; Bailey, J.; Linkman, S. Systematic literature reviews in software engineering—A systematic literature review. Inf. Softw. Technol. 2009, 51, 7–15. [Google Scholar] [CrossRef]
  13. Demers, L.; Fuhrer, M.J.; Jutai, J.; Lenker, J.; Depa, M.; De Ruyter, F. A conceptual framework of outcomes for caregivers of assistive technology users. Am. J. Phys. Med. Rehabil. 2009, 88, 645–655. [Google Scholar] [CrossRef] [PubMed]
  14. Jackson, A.; Zimmermann, J.B. Neural interfaces for the brain and spinal cord—Restoring motor function. Nat. Rev. Neurol. 2012, 8, 690–699. [Google Scholar] [CrossRef] [PubMed]
  15. López-Larraz, E.; Sarasola-Sanz, A.; Irastorza-Landa, N.; Birbaumer, N.; Ramos-Murguialday, A. Brain-machine interfaces for rehabilitation in stroke: A review. NeuroRehabilitation 2018, 43, 77–97. [Google Scholar] [CrossRef] [PubMed]
  16. Coscia, M.; Wessel, M.J.; Chaudary, U.; Millán, J.d.R.; Micera, S.; Guggisberg, A.; Vuadens, P.; Donoghue, J.; Birbaumer, N.; Hummel, F.C. Neurotechnology-aided interventions for upper limb motor rehabilitation in severe chronic stroke. Brain 2019, 142, 2182–2197. [Google Scholar] [CrossRef]
  17. Oguntosin, V.W.; Mori, Y.; Kim, H.; Nasuto, S.J.; Kawamura, S.; Hayashi, Y. Design and validation of exoskeleton actuated by soft modules toward neurorehabilitation—Vision-based control for precise reaching motion of upper limb. Front. Neurosci. 2017, 11, 352. [Google Scholar] [CrossRef]
  18. Cheng, N.; Phua, K.S.; Lai, H.S.; Tam, P.K.; Tang, K.Y.; Cheng, K.K.; Yeow, R.C.; Ang, K.K.; Guan, C.; Lim, J.H. Brain–computer interface-based soft robotic glove rehabilitation for stroke. IEEE Trans. Biomed. Eng. 2020, 67, 3339–3351. [Google Scholar] [CrossRef] [PubMed]
  19. Schwarz, A.; Höller, M.K.; Pereira, J.; Ofner, P.; Müller-Putz, G.R. Decoding hand movements from human EEG to control a robotic arm in a simulation environment. J. Neural Eng. 2020, 17, 036010. [Google Scholar] [CrossRef] [PubMed]
  20. Diekelmann, S.; Born, J. The memory function of sleep. Nat. Rev. Neurosci. 2010, 11, 114–126. [Google Scholar] [CrossRef]
  21. Meng, J.; Zhang, S.; Bekyo, A.; Olsoe, J.; Baxter, B.; He, B. Non-invasive electroencephalogram based control of a robotic arm for reach and grasp tasks. Sci. Rep. 2016, 6, 38565. [Google Scholar] [CrossRef]
  22. Akbari, H.; Sadiq, M.T.; Payan, M.; Esmaili, S.S.; Baghri, H.; Bagheri, H. Depression detection based on geometrical features extracted from SODP shape of EEG signals and binary PSO. Trait. Signal 2021, 38, 13–26. [Google Scholar] [CrossRef]
  23. Sadiq, M.T.; Yu, X.; Yuan, Z.; Zeming, F.; Rehman, A.U.; Ullah, I.; Li, G.; Xiao, G. Motor imagery EEG signals decoding by multivariate empirical wavelet transform-based framework for robust brain–computer interfaces. IEEE Access 2019, 7, 171431–171451. [Google Scholar] [CrossRef]
  24. Sharma, R.; Kim, M.; Gupta, A. Motor imagery classification in brain-machine interface with machine learning algorithms: Classical approach to multi-layer perceptron model. Biomed. Signal Process. Control 2022, 71, 103101. [Google Scholar] [CrossRef]
  25. Antony, J.W.; Piloto, L.; Wang, M.; Pacheco, P.; Norman, K.A.; Paller, K.A. Sleep spindle refractoriness segregates periods of memory reactivation. Current Biol. 2018, 28, 1736–1743. [Google Scholar] [CrossRef] [PubMed]
  26. Akbari, H.; Sadiq, M.T.; Jafari, N.; Too, J.; Mikaeilvand, N.; Cicone, A.; Serra-Capizzano, S. Recognizing seizure using Poincaré plot of EEG signals and graphical features in DWT domain. Bratisl. Med. J. 2022, 124, 12–24. [Google Scholar] [CrossRef]
  27. Sadiq, M.T.; Yu, X.; Yuan, Z.; Aziz, M.Z. Motor imagery BCI classification based on novel two-dimensional modelling in empirical wavelet transform. Electron. Lett. 2020, 56, 1367–1369. [Google Scholar] [CrossRef]
  28. LaFleur, K.; Cassady, K.; Doud, A.; Shades, K.; Rogin, E.; He, B. Quadcopter control in three-dimensional space using a noninvasive motor imagery-based brain–computer interface. J. Neural Eng. 2013, 10, 046003. [Google Scholar] [CrossRef] [PubMed]
  29. Wodlinger, B.; Downey, J.E.; Tyler-Kabara, E.C.; Schwartz, A.B.; Boninger, M.L.; Collinger, J.L. Ten-dimensional anthropomorphic arm control in a human brain–machine interface: Difficulties, solutions, and limitations. J. Neural Eng. 2015, 12, 016011. [Google Scholar] [CrossRef] [PubMed]
  30. Huang, B.; Xu, H.; Yuan, M.; Aziz, M.Z.; Yu, X. Exploiting asymmetric EEG signals with EFD in deep learning domain for robust BCI. Symmetry 2022, 14, 2677. [Google Scholar] [CrossRef]
  31. Cypher, D.; Chevrollier, N.; Montavont, N.; Golmie, N. Prevailing over wires in healthcare environments: Benefits and challenges. IEEE Commun. Mag. 2006, 44, 56–63. [Google Scholar] [CrossRef]
  32. Sarasola-Sanz, A.; López-Larraz, E.; Irastorza-Landa, N.; Klein, J.; Valencia, D.; Belloso, A.; Morin, F.O.; Spüler, M.; Birbaumer, N.; Ramos-Murguialday, A. An EEG-based brain-machine interface to control a 7-degrees of freedom exoskeleton for stroke rehabilitation. Converging Clin. Eng. Res. Neurorehabilit. II (Biosyst. Biorobot.) 2017, 15, 1127–1131. [Google Scholar] [CrossRef]
  33. Escolano, C.; Navarro-Gil, M.; Garcia-Campayo, J.; Congedo, M.; Ridder, D.D.D.; Minguez, J. A controlled study on the cognitive effect of alpha neurofeedback training in patients with major depressive disorder. Front. Behav. Neurosci. 2014, 8, 296. [Google Scholar] [CrossRef] [PubMed]
  34. Mortenson, W.; Demers, L.; Fuhrer, M.; Jutai, J.; Lenker, J.; DeRuyter, F. Development and preliminary evaluation of the caregiver assistive technology outcome measure. J. Rehabil. Med. 2015, 47, 412–418. [Google Scholar] [CrossRef] [PubMed]
  35. Chiang, H.S.; Hsiao, K.L.; Liu, L.C. EEG-based detection model for evaluating and improving learning attention. J. Med. Biol. Eng. 2018, 38, 847–856. [Google Scholar] [CrossRef]
  36. Iturrate, I.; Chavarriaga, R.; Montesano, L.; Minguez, J.; Millán, J.d.R. Teaching brain-machine interfaces as an alternative paradigm to neuroprosthetics control. Sci. Rep. 2015, 5, 13893. [Google Scholar] [CrossRef] [PubMed]
  37. Hamid, H.; Naseer, N.; Nazeer, H.; Khan, M.J.; Khan, R.A.; Khan, U.S. Analyzing classification performance of fNIRS-BCI for gait rehabilitation using deep neural networks. Sensors 2022, 22, 1932. [Google Scholar] [CrossRef] [PubMed]
  38. Buch, E.; Weber, C.; Cohen, L.G.; Braun, C.; Dimyan, M.A.; Ard, T.; Mellinger, J.; Caria, A.; Soekadar, S.; Fourkas, A.; et al. Think to move: A neuromagnetic brain–computer interface (BCI) system for chronic stroke. Stroke 2008, 39, 910–917. [Google Scholar] [CrossRef] [PubMed]
  39. Nakagome, S.; Luu, T.P.; He, Y.; Ravindran, A.S.; Contreras-Vidal, J.L. An empirical comparison of neural networks and machine learning algorithms for EEG gait decoding. Sci. Rep. 2020, 10, 4372. [Google Scholar] [CrossRef]
  40. Tortora, S.; Ghidoni, S.; Chisari, C.; Micera, S.; Artoni, F. Deep learning-based BCI for gait decoding from EEG with LSTM recurrent neural network. J. Neural Eng. 2020, 17, 046011. [Google Scholar] [CrossRef] [PubMed]
  41. Martini, M.L.; Oermann, E.K.; Opie, N.L.; Panov, F.; Oxley, T.; Yaeger, K. Sensor modalities for brain-computer interface technology: A comprehensive literature review. Neurosurgery 2020, 86, E108–E117. [Google Scholar] [CrossRef]
  42. Alotaiby, T.; El-Samie, F.E.A.; Alshebeili, S.A.; Ahmad, I. A review of channel selection algorithms for EEG signal processing. EURASIP J. Adv. Signal Process. 2015, 2015, 66. [Google Scholar] [CrossRef]
  43. Wan, X.; Zhang, K.; Ramkumar, S.; Deny, J.; Emayavaramban, G.; Ramkumar, M.S.; Hussein, A.F. A review on electroencephalogram based brain computer interface for elderly disabled. IEEE Access 2019, 7, 36380–36387. [Google Scholar] [CrossRef]
  44. Pfurtscheller, G.; Neuper, C. Motor imagery and direct brain-computer communication. Proc. IEEE 2001, 89, 1123–1134. [Google Scholar] [CrossRef]
  45. Jeunet, C.; Jahanpour, E.; Lotte, F. Why standard brain-computer interface (BCI) training protocols should be changed: An experimental study. J. Neural Eng. 2016, 13, 036024. [Google Scholar] [CrossRef] [PubMed]
  46. Farnsworth, B. Eeg (Electroencephalography): The Complete Pocket Guide; IMotions, Global HQ: Copenhagen, Denmark, 2019. [Google Scholar]
  47. Shute, V.J.; Zapata-Rivera, D. Adaptive technologies. ETS Res. Rep. Ser. 2007, 2007, 34. [Google Scholar] [CrossRef]
  48. Norman, S.L.; McFarland, D.J.; Miner, A.; Cramer, S.C.; Wolbrecht, E.T.; Wolpaw, J.R.; Reinkensmeyer, D.J. Controlling pre-movement sensorimotor rhythm can improve finger extension after stroke. J. Neural Eng. 2018, 15, 056026. [Google Scholar] [CrossRef]
  49. Sadiq, M.T.; Akbari, H.; Siuly, S.; Li, Y.; Wen, P. Alcoholic EEG signals recognition based on phase space dynamic and geometrical features. Chaos Solitons Fractals 2022, 158, 112036. [Google Scholar] [CrossRef]
  50. Sadiq, M.T.; Yu, X.; Yuan, Z.; Aziz, M.Z.; Rehman, N.u.; Ding, W.; Xiao, G. Motor imagery BCI classification based on multivariate variational mode decomposition. IEEE Trans. Emerg. Top. Comput. Intell. 2022, 6, 1177–1189. [Google Scholar] [CrossRef]
  51. Yu, X.; Aziz, M.Z.; Sadiq, M.T.; Fan, Z.; Xiao, G. A new framework for automatic detection of motor and mental imagery EEG signals for robust BCI systems. IEEE Trans. Instrum. Meas. 2021, 70, 1–12. [Google Scholar] [CrossRef]
  52. Altaheri, H.; Muhammad, G.; Alsulaiman, M.; Amin, S.; Altuwaijri, G.; Abdul, W.; Bencherif, M.; Faisal, M. Deep learning techniques for classification of electroencephalogram (EEG) motor imagery (MI) signals: A review. Neural Comput. Appl. 2023, 35, 14681–14722. [Google Scholar] [CrossRef]
  53. Escolano, C.; Navarro-Gil, M.; Garcia-Campayo, J.; Congedo, M.; Minguez, J. The effects of individual upper alpha neurofeedback in ADHD: An open-label pilot study. Appl. Psychophysiol. Biofeedback 2014, 39, 193–202. [Google Scholar] [CrossRef]
  54. Ketz, N.; Jones, A.P.; Bryant, N.B.; Clark, V.P.; Pilly, P.K. Closed-loop slow-wave tACS improves sleep-dependent long-term memory generalization by modulating endogenous oscillations. J. Neurosci. 2018, 38, 7314–7326. [Google Scholar] [CrossRef] [PubMed]
  55. Acı, Ç.İ.; Kaya, M.; Mishchenko, Y. Distinguishing mental attention states of humans via an EEG-based passive BCI using machine learning methods. Expert Syst. Appl. 2019, 134, 153–166. [Google Scholar] [CrossRef]
  56. Kamble, A.; Ghare, P.H.; Kumar, V. Deep-learning-based BCI for automatic imagined speech recognition using SPWVD. IEEE Trans. Instrum. Meas. 2023, 72, 1–10. [Google Scholar] [CrossRef]
  57. Venkatesh, B.; Anuradha, J. A review of feature selection and its methods. Cybern. Inf. Technol. 2019, 19, 3–26. [Google Scholar] [CrossRef]
  58. Atyabi, A.; Shic, F.; Naples, A. Mixture of autoregressive modeling orders and its implication on single trial EEG classification. Expert Syst. Appl. 2016, 65, 164–180. [Google Scholar] [CrossRef] [PubMed]
  59. Karaboga, D. An Idea Based on Honey Bee Swarm for Numerical Optimization; Technical Report, Technical Report-tr06; Erciyes University, Engineering Faculty, Computer Engineering Department: Kayseri, Turkey, 2005. [Google Scholar]
  60. Rakshit, P.; Bhattacharyya, S.; Konar, A.; Khasnobish, A.; Tibarewala, D.N.; Janarthanan, R. Artificial Bee Colony Based Feature Selection for Motor Imagery EEG Data. In Proceedings of the Seventh International Conference on Bio-Inspired Computing: Theories and Applications (BIC-TA 2012); Bansal, J.C., Singh, P., Deep, K., Pant, M., Nagar, A., Eds.; Springer: Gwalior, India, 2013; pp. 127–138. [Google Scholar]
  61. Proulx, C.E.; Beaulac, M.; David, M.; Deguire, C.; Haché, C.; Klug, F.; Kupnik, M.; Higgins, J.; Gagnon, D.H. Review of the effects of soft robotic gloves for activity-based rehabilitation in individuals with reduced hand function and manual dexterity following a neurological event. J. Rehabil. Assistive Technol. Eng. 2020, 7, 205566832091813. [Google Scholar] [CrossRef] [PubMed]
  62. Tabar, Y.R.; Halici, U. A novel deep learning approach for classification of EEG motor imagery signals. J. Neural Eng. 2016, 14, 016003. [Google Scholar] [CrossRef]
  63. Tang, Z.; Li, C.; Sun, S. Single-trial EEG classification of motor imagery using deep convolutional neural networks. Optik 2017, 130, 11–18. [Google Scholar] [CrossRef]
  64. Dose, H.; Møller, J.S.; Iversen, H.K.; Puthusserypady, S. An end-to-end deep learning approach to MI-EEG signal classification for BCIs. Expert Syst. Appl. 2018, 114, 532–542. [Google Scholar] [CrossRef]
  65. Olivas-Padilla, B.E.; Chacon-Murguia, M.I. Classification of multiple motor imagery using deep convolutional neural networks and spatial filters. Appl. Soft Comput. 2019, 75, 461–472. [Google Scholar] [CrossRef]
  66. Tayeb, Z.; Fedjaev, J.; Ghaboosi, N.; Richter, C.; Everding, L.; Qu, X.; Wu, Y.; Cheng, G.; Conradt, J. Validating Deep Neural Networks for Online Decoding of Motor Imagery Movements from EEG Signals. Sensors 2019, 19, 210. [Google Scholar] [CrossRef] [PubMed]
  67. Tang, X.; Zhao, J.; Fu, W.; Pan, J.; Zhou, H. A novel classification algorithm for MI-EEG based on deep learning. In Proceedings of the 2019 IEEE 8th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), Chongqing, China, 24–26 May 2019; pp. 606–611. [Google Scholar]
  68. Qiao, W.; Bi, X. Deep spatial-temporal neural network for classification of EEG-based motor imagery. In Proceedings of the 2019 International Conference on Artificial Intelligence and Computer Science, Wuhan, China, 12–13 July 2019; pp. 265–272. [Google Scholar]
  69. Song, Y.; Wang, D.; Yue, K.; Zheng, N.; Shen, Z.J.M. EEG-based motor imagery classification with deep multi-task learning. In Proceedings of the 2019 International Joint Conference on Neural Networks (IJCNN), Budapest, Hungary, 14–19 July 2019; pp. 1–8. [Google Scholar]
  70. Amin, S.U.; Alsulaiman, M.; Muhammad, G.; Mekhtiche, M.A.; Shamim Hossain, M. Deep Learning for EEG motor imagery classification based on multi-layer CNNs feature fusion. Future Gener. Comput. Syst. 2019, 101, 542–554. [Google Scholar] [CrossRef]
  71. Alazrai, R.; Abuhijleh, M.; Alwanni, H.; Daoud, M.I. A Deep Learning Framework for Decoding Motor Imagery Tasks of the Same Hand Using EEG Signals. IEEE Access 2019, 7, 109612–109627. [Google Scholar] [CrossRef]
  72. Li, F.; He, F.; Wang, F.; Zhang, D.; Xia, Y.; Li, X. A Novel Simplified Convolutional Neural Network Classification Algorithm of Motor Imagery EEG Signals Based on Deep Learning. Appl. Sci. 2020, 10, 1605. [Google Scholar] [CrossRef]
  73. Miao, M.; Hu, W.; Yin, H.; Zhang, K. Spatial-Frequency Feature Learning and Classification of Motor Imagery EEG Based on Deep Convolution Neural Network. Comput. Math. Methods Med. 2020, 2020, 1981728. [Google Scholar] [CrossRef] [PubMed]
  74. Yang, J.; Ma, Z.; Wang, J.; Fu, Y. A Novel Deep Learning Scheme for Motor Imagery EEG Decoding Based on Spatial Representation Fusion. IEEE Access 2020, 8, 202100–202110. [Google Scholar] [CrossRef]
  75. Ieracitano, C.; Morabito, F.C.; Hussain, A.; Mammone, N. A hybrid-domain deep learning-based BCI for discriminating hand motion planning from EEG sources. Int. J. Neural Syst. 2021, 31, 2150038. [Google Scholar] [CrossRef]
  76. Zhang, C.; Kim, Y.K.; Eskandarian, A. EEG-inception: An accurate and robust end-to-end neural network for EEG-based motor imagery classification. J. Neural Eng. 2021, 18, 046014. [Google Scholar] [CrossRef] [PubMed]
  77. Deng, X.; Zhang, B.; Yu, N.; Liu, K.; Sun, K. Advanced TSGL-EEGNet for Motor Imagery EEG-Based Brain-Computer Interfaces. IEEE Access 2021, 9, 25118–25130. [Google Scholar] [CrossRef]
  78. Cho, J.H.; Jeong, J.H.; Lee, S.W. NeuroGrasp: Real-Time EEG Classification of High-Level Motor Imagery Tasks Using a Dual-Stage Deep Learning Framework. IEEE Trans. Cybern. 2022, 52, 13279–13292. [Google Scholar] [CrossRef]
  79. Tiwari, S.; Goel, S.; Bhardwaj, A. MIDNN-a classification approach for the EEG based motor imagery tasks using deep neural network. Appl. Intell. 2022, 52, 4824–4843. [Google Scholar] [CrossRef]
  80. Huang, E.; Zheng, X.; Fang, Y.; Zhang, Z. Classification of Motor Imagery EEG Based on Time-Domain and Frequency-Domain Dual-Stream Convolutional Neural Network. IRBM 2022, 43, 107–113. [Google Scholar] [CrossRef]
  81. Huang, W.; Chang, W.; Yan, G.; Yang, Z.; Luo, H.; Pei, H. EEG-based motor imagery classification using convolutional neural networks with local reparameterization trick. Expert Syst. Appl. 2022, 187, 115968. [Google Scholar] [CrossRef]
  82. Zhu, H.; Forenzo, D.; He, B. On the Deep Learning Models for EEG-Based Brain-Computer Interface Using Motor Imagery. IEEE Trans. Neural Syst. Rehabil. Eng. 2022, 30, 2283–2291. [Google Scholar] [CrossRef]
  83. Mattioli, F.; Porcaro, C.; Baldassarre, G. A 1D CNN for high accuracy classification and transfer learning in motor imagery EEG-based brain-computer interface. J. Neural Eng. 2022, 18, 066053. [Google Scholar] [CrossRef]
  84. Du, Y.; Liu, J. IENet: A robust convolutional neural network for EEG based brain-computer interfaces. J. Neural Eng. 2022, 19, 036031. [Google Scholar] [CrossRef]
  85. Li, Y.; Zhang, Q.; Chen, W. Adaptive CNN for EEG classification in stroke rehabilitation. J. Neurosci. Methods 2023, 275, 112–120. [Google Scholar]
  86. Wang, Z.; Liu, S.; Zhou, Y. Attention-based CNN for EEG motor imagery classification. IEEE Trans. Neural Syst. Rehabil. Eng. 2023, 31, 1365–1373. [Google Scholar]
  87. Zhang, L.; Wang, H.; Li, J. Graph-CNN for spatial feature learning in EEG classification. Med. Biol. Eng. Comput. 2023, 61, 451–459. [Google Scholar]
  88. Xu, Y.; Zhang, W.; Chen, S. Hybrid CNN-RNN model for EEG motor imagery classification. Front. Neurosci. 2024, 18, 127–135. [Google Scholar]
  89. Chen, L.; Liu, Q.; Yang, H. Transformer-based model for EEG classification in stroke rehabilitation. IEEE Trans. Biomed. Eng. 2024, 71, 1461–1470. [Google Scholar]
Figure 1. Overview of a brain–computer interface (BCI) system for motor imagery-based stroke rehabilitation.
Figure 1. Overview of a brain–computer interface (BCI) system for motor imagery-based stroke rehabilitation.
Applsci 14 06347 g001
Figure 2. International 10–20 electrode placement system.
Figure 2. International 10–20 electrode placement system.
Applsci 14 06347 g002
Figure 3. One trial in Graz Protocol.
Figure 3. One trial in Graz Protocol.
Applsci 14 06347 g003
Figure 4. Illustration of SVM hyperplane separating two classes.
Figure 4. Illustration of SVM hyperplane separating two classes.
Applsci 14 06347 g004
Figure 5. Illustration of k-nearest neighbors classification.
Figure 5. Illustration of k-nearest neighbors classification.
Applsci 14 06347 g005
Figure 6. Structure of a Convolutional Neural Network.
Figure 6. Structure of a Convolutional Neural Network.
Applsci 14 06347 g006
Figure 7. Structure of a Recurrent Neural Network.
Figure 7. Structure of a Recurrent Neural Network.
Applsci 14 06347 g007
Figure 8. Architecture of a Long Short-Term Memory Network.
Figure 8. Architecture of a Long Short-Term Memory Network.
Applsci 14 06347 g008
Figure 9. Diagram of a Multi-Layer Perceptron.
Figure 9. Diagram of a Multi-Layer Perceptron.
Applsci 14 06347 g009
Figure 10. Most popular datasets.
Figure 10. Most popular datasets.
Applsci 14 06347 g010
Table 1. Description of EEG devices.
Table 1. Description of EEG devices.
BrandModelNumber of ChannelsIntended Use
EmotivEmotiv EPOCH5–14 channelsUsed in research and for personal use
Emotiv Insight
BIOPAC systemsEEG100C16 channelsUsed in sleep studies and evoked responses
OpenBCIOpenBCI 32-bit4–21 channelsUsed in BCI and biosensing
OpenBCI Cyton
OpenBCI Ganglion
Ultracortex BCI
NaroskyBrainwave1 channelUsed in neurogaming and meditation
Mindflex
Mindwave
Table 2. Most significant publications from 2016 to 2024.
Table 2. Most significant publications from 2016 to 2024.
ReferencesYearDatasetDL ModelClassification Results
[62]2016BCI competition IV dataset 2bCNN+SAE72.40%
[63]2017Collected (109 subjects)CNN86.41%
[64]2018Physionet EEG MI DatasetCNN80.38%
[65]2019BCI competition IV dataset 2aCNN82.09%
[66]2019BCI competition IV dataset 2bCNN77.72%
[67]2019BCI competition data IV 2aCNN+SAE79.90%
[68]2019BCI competition data IV 2aCNN+Bi-GRU76.62%
[69]2019BCI competition data IV 2aCNN73.40%
[70]2019BCI competition data IV 2aCNN75.7%
[71]2019Collected (22 subjects)CNN73.70%
[72]2020BCI Competition IV 2bCNN83.20%
[73]2020BCI competition IVa, right index finger MI datasetCNN90.00%
[74]2020BCI competition IV dataset 1CNN86.40%
[75]202115 subjectsCNN76.21%
[76]2021BCI Competition IV dataset 2a and 2bCNN88.40%
[77]2021BCI Competition IV 2a, IIICNN85.30%
[78]2021Collected (12 subjects)Bi-LSTM68.00%
[79]2021BCI competition V dataset, Emotiv datasetCNN72.51% and 72%
[80]2021BCI competition IV dataset 2aCNN90.00%
[81]2022PhysioNet datasetCNN92.00%
[82]2022Med-62ConvNet72.66%
[83]2022EEG Motor Movement Dataset V 1.0.0CNN99.38%
[84]2022MRCPCNN91.00%
[85]2023BCI competition IV dataset 2aAdaptive CNN93.20%
[86]2023Collected (30 subjects)Attention-based CNN88.75%
[87]2023BCI competition IV dataset 2bGraph-CNN89.60%
[88]2024PhysioNet EEG MI DatasetHybrid CNN-RNN94.50%
[89]2024BCI competition IV dataset 2aTransformer-based Model92.30%
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Elashmawi, W.H.; Ayman, A.; Antoun, M.; Mohamed, H.; Mohamed, S.E.; Amr, H.; Talaat, Y.; Ali, A. A Comprehensive Review on Brain–Computer Interface (BCI)-Based Machine and Deep Learning Algorithms for Stroke Rehabilitation. Appl. Sci. 2024, 14, 6347. https://doi.org/10.3390/app14146347

AMA Style

Elashmawi WH, Ayman A, Antoun M, Mohamed H, Mohamed SE, Amr H, Talaat Y, Ali A. A Comprehensive Review on Brain–Computer Interface (BCI)-Based Machine and Deep Learning Algorithms for Stroke Rehabilitation. Applied Sciences. 2024; 14(14):6347. https://doi.org/10.3390/app14146347

Chicago/Turabian Style

Elashmawi, Walaa H., Abdelrahman Ayman, Mina Antoun, Habiba Mohamed, Shehab Eldeen Mohamed, Habiba Amr, Youssef Talaat, and Ahmed Ali. 2024. "A Comprehensive Review on Brain–Computer Interface (BCI)-Based Machine and Deep Learning Algorithms for Stroke Rehabilitation" Applied Sciences 14, no. 14: 6347. https://doi.org/10.3390/app14146347

APA Style

Elashmawi, W. H., Ayman, A., Antoun, M., Mohamed, H., Mohamed, S. E., Amr, H., Talaat, Y., & Ali, A. (2024). A Comprehensive Review on Brain–Computer Interface (BCI)-Based Machine and Deep Learning Algorithms for Stroke Rehabilitation. Applied Sciences, 14(14), 6347. https://doi.org/10.3390/app14146347

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop