Keywords

1 Introduction

Electroencephalogram (EEG) is an effective method for deciphering brain signals that originate from the central nervous system and respond quickly to different brain states [1]. The brain receives information from its surroundings, and environmental recognition and perception are one of humans’ special characteristics. Establishing communication between the human brain and the computer helps with this environmental perception. People with mental disabilities and who cannot communicate with the environment are rehabilitated with this system [2].

The brain-computer interface (BCI) is a system that determines the functional goal, the desire to change, move, control, or interact with something in the environment directly from brain activity. In other words, BCIs make it possible to control a program or device using only the mind. Using a BCI ignores the need for voluntary muscle control to interact with the devices around us. Rather than execute a physical movement, the computer interprets the desired action and controls the device directly [3].

The BCI system comprises a communication pathway that directly links the brain to an external device. This is followed by signal processing, which is achieved without motor activities. The aforementioned system has been documented in references [4,5,6]. These systems often utilize EEG signals obtained either from the cranium or from electrodes implanted in the cortical neurons of an individual [7]. Electronic computerized devices require direct physical interaction between the user and the device. Illustrative instances include input devices such as a keyboard, a mouse, or a trackball. A BCI or Brain-Machine Interface (BMI) is a communication pathway that offers an alternative method for humans to interact with the external environment. BCI captures cerebral activity during cognitive processes and converts it into a directive for a designated output [8,9,10,11,12]. Multiple modes of brain activity detection can be employed in BCI applications to determine the user’s intention. Motor imagery (MI) is widely recognized as one of the most prominent modes. Scholarly literature defines MI as a cognitive process in which an individual envisions a specific bodily movement using neural activity in the sensorimotor cortex. This is rather than physically executing the movement. This process is characterized by oscillatory activity that corresponds to an imagined movement [13]. Prior to the commercial application of BCI from MI, several challenges need to be addressed. For a motor imagery-based BCI to be considered feasible, it should be utilized by individuals across various skill levels. This necessitates the interface to be designed plug-and-play, allowing ease of use. Additionally, the interface should be self-paced, highly responsive, and maintain performance consistency. This objective could be attained by surmounting the ensuing challenges through the implementation of the following approaches. Enhancement of Motor Imagery-Based Brain-Computer Interfaces (MI-BCIs) Minimize or eliminate calibration duration. BCI research suggests that an unsynchronized motor imagery-based brain-computer interface (MI-BCI) may indicate literacy lack. Enhance the number of commands. The online MI-BCI Instructional Procedure is also known as Flexible BCI [14, 15].

The non-stationary characteristics of EEG signals pose a challenge for machine learning methods in accurately extracting complex non-linear features for MI-EEG classification, despite some progress being made in this area. Hence, deep learning (DL) techniques were employed to augment nonlinear characteristics extraction from EEG signals. Tabar et al. [16] employed the short-time Fourier transform (STFT) technique to transform motor imagery electroencephalography (MI-EEG) time series into two-dimensional (2D) images. Subsequently, either a convolutional neural network (CNN) or a stacked autoencoder (SAE) was used to classify MI-EEG signals. Schirrmeister et al. [17] introduced deep neural network architectures, specifically shallow and deep convolutional neural networks. These architectures can be used to classify motor imagery and electroencephalography data [18,19,20,21]. These models perform better than traditional techniques. Big data from healthcare applications and BCI usages are becoming more common because of the cloud and edge computing [22], machine learning [23,24,25], and other models [26, 27].

Recent years have seen researchers focus on models based on machine learning and deep learning in healthcare data analysis, which can be used in the processing of a variety of data sets. Sakhavi et al. [28] constructed a deep learning algorithm that acquires envelope representations for classifying MI-EEG. Furthermore, they conducted a fine-tuning process on the model, resulting in a 7% increase in classification accuracy on the BCI competition dataset IV 2a. Furthermore, Liu et al. [29] enhanced the detection of four-class motor imagery electroencephalogram (MI-EEG) signals through the utilization of a parallel spatial–temporal self-attention-based convolutional neural network (CNN) approach.

However, the outcomes produced by deep learning models were constrained by insufficient annotated data and diversity of MI-EEG signals acquired from participants. To tackle this problem, a number of studies in the field of Brain-Computer Interfaces (BCI) have utilized transfer learning. This approach involves utilizing the knowledge gained from the source domain to aid in the acquisition of information in the target domain [30]. Raza et al. [31] utilized an innovative approach to detecting and adapting to covariate shift. This approach can reduce differences between two feature sets. Zanini et al. [32] employed a Riemannian alignment technique to reduce the distance between separate domains in Riemannian space. This approach yielded favorable results in MI-EEG data classification across subjects. Azab et al. [33] proposed a novel similarity metric that utilizes Kullback–Leibler divergence (KL) to enhance logistic regression classifier performance. The aim was to evaluate the similarity between two feature spaces and enable weighted transfer learning.

For the development of BCI models using EEG signals, machine learning techniques present a significant opportunity. This paper provides an overview of recent developments in disease analysis, rehabilitation, and machine learning through BCI investigations. Furthermore, the potential and challenges of implementing machine learning in BCI systems are discussed. As shown in Fig. 1, BCI’s effects on diseases and their analysis have been studied in the last few years. Clearly, over the past few years, more associations have been discovered between diseases and disorders that lead to disability. As a non-invasive method of monitoring and controlling diseases, BCI has become increasingly popular.

Fig. 1
An arrow illustration with plots from 1 to 7 represents the introduction, statistical studies, epilepsy review, Alzheimer review, Parkinson check, review of multiple sclerosis, and investigation of sleep disorders.

During the last few years, the investigation of BCI and brain diseases has led to more associations with disorders and some disabilities

Full size image

Personalized and effective patient care can be made possible by this technology.

Here are the remaining sections of our study: The second section provides a brief overview of statistical analysis and progress using BCI systems. A discussion of the challenges and limitations of BCI systems is presented in Sect. 3. The study’s opportunities, future work, and conclusions are discussed in Sect. 4.

2 Progress in BCI Systems

Based on recent data compiled, it is estimated that 5–9% of students are affected by learning disabilities. Males exhibit higher prevalence than females [34]. Based on supplementary information, it can be inferred that roughly 6.2 million individuals in the United States are afflicted with disabilities stemming from traumatic brain injuries. The aforementioned may lead to extensive deficits in the cognitive, communicative, locomotive, perceptual, attitudinal, and psychological domains [35, 36]. Consequently, there has been an increased focus on brain-computer interface systems compared to previous times.

A BCI system quantifies neural activity, consolidates it, and generates a corresponding output. BCI is an entity that innovates in medical research. BCIs enable covert communication between the human brain and a machine, rendering them a viable instrument for neurological disorders management. The integration of techniques for gathering data, preparing it for analysis, extracting relevant features, and translating it into actionable insights is an integral component of a comprehensive BCI framework [37]. Brain signals such as EEG are captured by invasive or noninvasive sensors. Following that, the BCI system performs data preprocessing. Subsequently, the BCI system engages in machine learning, where it extracts pertinent data features and inputs them into translation algorithms. Subsequently, the computer-generated output device issues instructions to a mechanical apparatus, such as a robotic extremity, to exert an impact on either the user or the adjacent milieu.

The EEG-based BCI is a pivotal tool in rehabilitation for individuals diagnosed with attention deficit hyperactivity disorder (ADHD). Attention Deficit Hyperactivity Disorder (ADHD) is a notable neurological condition that affects many adolescents. ADHD is characterized by inattention, hyperactivity, and impulsivity. Medication is the conventional approach to managing ADHD. Nonetheless, it yields a plethora of detrimental outcomes for underage individuals. According to study [38], the treatment may have adverse effects on cardiovascular systems, height, and somatic complaints in individuals under the age of majority. The EEG-based BCI system employs the P300 potential in a range of virtual reality (VR) feedback games [39] to enhance participant attention. The advancement of neural cursors and brain-computer interface (BCI) spelling is beneficial for individuals diagnosed with Amyotrophic Lateral Sclerosis (ALS). ALS is a condition that disrupts neural networks, resulting in interference with communication pathways between the brain and other organs. The condition predominantly impacts muscle motor control. ALS shares similarities with Locked-in Syndrome (LIS) in that there is a gradual loss of voluntary muscle control, including the tongue muscle. The absence of conversational and sign language skills poses growing difficulty communicating with individuals diagnosed with ALS [40]. Neural Cursor and BCI Spellers are two communication modalities that serve as viable options for individuals diagnosed with ALS to restore their social communication abilities. BCI technology for seizure anticipation has shown promise in drug-resistant epilepsy patient management. Epilepsy is a persistent neurological condition distinguished by unprovoked seizures. Approximately 1% of the global populace is impacted by epilepsy, and approximately 30% of individuals diagnosed with epilepsy do not experience seizures. BCI can predict seizures in epilepsy patients [41]. This is because the epileptic area of the patient can only be identified and managed through the use of strategically positioned electrodes, as opposed than surgical elimination.

The use of MI-BCI has been observed to be advantageous in upper-limb stroke rehabilitation. Stroke is a highly deleterious neurological disorder in humans [42], often resulting in fatality or impairment. A considerable segment of stroke victims remains chronically impaired due to cerebral injury. Physical practice (PP) forms the basis of most rehabilitation approaches employed today. As per prior studies, individuals who have experienced a stroke affecting their upper limbs can reactivate authentic movements through the application of MI-BCI. Additionally, a subset of these individuals exhibits comparable levels of proficiency in controlling non-invasive, EEG-based MI-BCI as their healthy counterparts [43].

3 BCI Analysis and Challenges

The performance of Brain Computer Interface systems has been disrupted by problems found in research. Friedrich et al. [44] and Alvarsson et al. [45] both found that environmental auditory disorders could affect an individual’s behavior. The performance of the BCI system is directly affected by them. When the eyes are open or closed, the brain’s movement patterns change, and these changes will be problematic for BCIs [46, 47].

There are several types of constraint satisfaction problems (CSPs) [48]. In recent years, researchers have commonly used this method. In the past decade, a variety of CSPs have been optimized for EEG signal decoding, including Novietal et al. [49].

A subband method was proposed to extract the function of each frequency band in EEG signals. Noise is one of the challenges in decoding brain signals. To perform correct decoding, complex algorithms must be designed that detect and remove noise. Currently, many traditional machine learning algorithms have been used for EEG classification, including Random Forest (RF) [49], Support Vector Machine (SVM) [50], and Linear Discriminant Analysis (LDA) [51]. However, because they lacked automatic feature extraction and noise detection, their output was not precise enough.

Machine learning and artificial intelligence have greatly benefited from research into deep learning [52]. A number of algorithms are used in this method to automatically extract features from different layers. It involves learning the hierarchy of features, i.e., higher-level features are defined by lower-level ones. Deep learning methods have gained special significance in Brain Computer Interface [53] due to their better performance characteristics compared to traditional methods and machine learning, as well as their ability to comprehend complex correlations between brain responses and reduce the need to extract handcraft features.

A graph attention network [54] displays nodes in the entire community according to their attention [55]. A model called Simplifying Graph Convolutional Networks was proposed by Wu et al. [56]. Machine learning problems involving structured graph data have been solved using graph neural networks [57].

In 2015, ResNet was discovered and was the first deep learning network to train thousands of layers without gradient fading; ResNet relies on residual learning to ensure the inputs from the previous layer are correct [58].

Using 15 healthy right-handed men between the ages of 23 and 54 years, Shelishiyah et al. [59] proposed a method in which the brain communicates with the hardware without the involvement of the nervous or muscular systems. EEG-fNIRS hybrid systems have improved classification accuracy and classified classes [60]. As a result of this hybrid system, noise is reduced and accuracy is increased.

According to Choi et al. [61], in their methodological study on SMR [62, 63], they proposed to examine the three-dimensional movement of the arm non-invasively. In order to estimate a user’s functional status and improve diagnosis accuracy [64], convolutional neural networks are used. However, this method requires long training times for motion analysis [65]. Timeline of decoding brain signals and making changes in the meaning of decoding from 2008 to 2022 is shown in Fig. 2.

Fig. 2
A timeline illustration from 2008 to 2022. Some of the events are brain messages being decoded using brain signals in 2008, advances in the field of deep learning and artificial intelligence in 2016, and articles written regarding rehabilitation and decoding of brain signals in 2022.

Timeline of decoding brain signals and making changes in the meaning of decoding from 2008 to 2022

Full size image

4 Neurological Diseases and BCI

Neurological diseases have been studied for centuries, with the most notable advances occurring in the past two centuries. In the last two decades, BCI technology has been developed to allow for direct communication between a computer and the brain. This technology is expected to revolutionize neurological studies and treatments. The diseases related to the world of BCI will be introduced in the following, and the important parts related to Ange can be seen in Fig. 3.

Fig. 3
A timeline illustration from 1999 to 2023. Some of the events include decoding brain signals as a new way to diagnose diseases in 1999, solving diagnosis problems and creating optimal solutions in 2021, and expanding studies based on A I in 2023.

Basically, the first two stages of brain signal decoding are the first stages, the middle two stages are growth and troubleshooting stages, and the last two stages are development and improvement stages

Full size image

4.1 Epilepsy

A person with epilepsy suffers from a brain disorder characterized by frequent and unpredictable interruptions in the brain’s normal functioning.

Rather than a specific disease, epilepsy refers to a set of abnormalities in brain function caused by various factors [66]. In the studies of Namazi et al. [67], this issue has been investigated, as aging can affect the neuroplasticity of epilepsy patients. According to their study, fractal EEGs have been calculated for people in different age groups, and aging is associated with epilepsy exacerbation.

Using deep learning, Gao et al. [68] classified epileptic EEG signals (EESC). Their classification accuracy was 90% using power spectrum density energy diagrams (PSDEDs).

With CNN, Zhou et al. [69] were able to detect epilepsy with 93% accuracy. Their work has several disadvantages, including the time domain signals, which are accurate for some patients but inaccurate for others.

One method used is combining several deep learning methods. According to Tian et al. [70], by combining fast Fourier transform (FFT), wavelet packet decomposition (WPD), and convolutional neural network (CNN), this method increases feature extraction accuracy by at least 1% and classification accuracy by at least 4%. It is possible to analyze epileptic spikes in brain signals (EEG) using tensors. A tensor and simultaneous multilinear low-rank tensor (SMLRAT) can be used to obtain local optimal solutions. This method was used for the first time to diagnose epileptic hyacinths by Dao et al. [71].

4.2 Alzheimer Diseases

Topographic maps can be used to diagnose and examine Alzheimer’s disease (AD). The method can be used to diagnose mild cognitive impairment (MCI), mild Alzheimer’s disease (MAD), and advanced Alzheimer’s disease (AAD). Rodrigues et al. [72] obtained a 95.55% accuracy using topographic maps.

Safi et al. [73] used Hjorth parameters and deep learning classification networks to accurately diagnose Alzheimer’s disease. Combining Hjorth parameters with common features can improve classification accuracy.

According to Meghdadi et al. [74], patients with severe Alzheimer’s show frequent changes in the alpha frequency band, while those with mild Alzheimer’s do not.

Another criterion for diagnosing Alzheimer’s disease is the presence of biomarkers. Una Smailovic et al. [75] have been able to find a dynamic relationship in neural signaling by analyzing brain activity at rest despite the lack of a strong and functional relationship between these indicators. This method promises to be low-cost and non-invasive for the diagnosis of Alzheimer’s disease.

Nerve activity frequency changes are considered a method of diagnosing disease. Oscillatory activity changes with frequency in both healthy and pathological aging. Benwell et al. [76] used this method to examine the relationship between healthy aging and EEG signals in Alzheimer’s disease and type 2 diabetes. In type 2 diabetes patients, the majority of them perform weaker on learning and cognitive tests and are disturbed.

According to Smailovic et al. [77], studying synaptic function is one of the early diagnosis methods for Alzheimer’s disease. According to studies, CFS (conventional cerebrospinal fluid) biomarkers show a significant correlation with Alzheimer’s disease biomarkers.

4.3 Parkinson Diseases

Unique spatial microstates can reveal differences between a healthy person and someone with Parkinson’s (PD) that may be related to brain dysfunction. Chu et al. [78] concluded that people with Parkinson’s who don’t take drugs have regular changes in their temporal microstates. A diagnosis of Parkinson’s disease can be made for them.

According to Galves et al. [79], their study investigating the effects of earbeats on EEG power and walking and Parkinson’s disease disorders found that hearing stimulation did not affect patients’ walking significantly. Parkinson’s patients can, however, benefit from it by improving their working memory.

Using the flexible analytic wavelet transform (FAWT), Chawla et al. [80] were able to detect entropy parameters with 99% accuracy. Parkinson’s syndrome can be automatically diagnosed using this decision support system.

According to Shah et al. [81], Parkinson’s disease is a method that has been investigated. Results show that there is a strong relationship between high-frequency components’ amplitudes and phases. The neural network designed during this research achieved a classification accuracy of 99.2%.

Anjum et al. [82] investigated the use of predictive-coding EEG algorithms to diagnose Parkinson’s disease. For Parkinson’s disease diagnosis, this method converts the temporal features of EEG into coded features.

Bhat et al. [83] investigated Parkinson’s disease causes and found that gene mutation and aging were involved. In addition to analyzing brain signals and brain imaging, gene therapy methods can be used to study nerve protection.

4.4 Multiple Sclerosis

Symptoms of multiple sclerosis (MS) include problems with vision, arm movement, sensation, balance, and sensation, among other things. Occasionally, it can cause mild disability, but it is usually a lifelong condition. An important component of brain signal analysis is the recognition of emotions. Using the domain adaptation (DA) approach, emotions have been detected in some research, but they are not accurate enough. In order to extract three or four emotions from brain signals, Chen et al. [84] introduced a multi-source marginal distribution adaptation (MS-MDA) approach. By using this method, brain signals can be taken into account both invariant and domain-specific features.

The use of electrical stimulation of the brain is another method. In the study conducted by Fiene et al. [85]; fatigue has been shown to be one of the most common and debilitating factors in MS, and nerve stimulation can reduce this fatigue during experiments, and the results have proven to be effective for treating MS.

With virtual reality (VR), movement behaviors can be simulated and a behavioral pattern of the patient’s movement can be obtained. A virtual reality study conducted by Recenti et al. [86] examined the movement patterns of muscles, the heart, and the brain. MS has been able to achieve an accuracy of 74.7%.

Neurological diseases are commonly diagnosed using computer-aided diagnosis (CAD) techniques. In their studies, Ahmadi et al. [87] have introduced a new CAD system that can diagnose diseases. Approximately 90% of MS’s predictions are accurate.

McMackin et al. [88] investigated how the brain behaves in people with MS by measuring network disturbances, receiving brain images, and analyzing EEG signals.

Michel et al. [89] investigated the use of EEG microstates as a tool for analyzing brain systems. It is possible to analyze brain signals for diseases such as epilepsy and MS by examining these microstates. These microstates are defined as consecutive short periods of time.

4.5 Sleep Disorder

Patients with sleep disorders may experience difficulty sleeping due to disrupted sleep patterns. The traditional method of identifying sleep disorders is time-consuming and boring, so Sharma et al. [90] have developed an automatic method that can identify six sleep disorders and reach 91.3% accuracy.

Because of the emergence of Coronavirus (COVID-19) and its epidemic, people in society are suffering from a lot of stress and sleep problems; therefore, Semyachkina-Glushkovskaya et al. [91] found that sleep disorders can be a new marker for Corona disease during their studies.

Apnea is a sleep disorder that causes prolonged waking during sleep. The EEG microstructure during neurological disorders after prolonged awakening has been examined by Mullins et al. [92]. There is a decrease in the correlation between the nervous system and the nervous function in these people, indicating a slower decision-making process.

Sleep is characterized by cycles of alternating sleep (CAP). Sharma et al. [93] were able to accurately diagnose sleep disorders by examining this pattern and recognizing the average sleep pattern 78% of the time.

As described by Lai et al. [94], teeth grinding is one of the problems caused by the sleep cycle. According to them, sleep disorders can be predicted by analyzing people’s teeth-grinding patterns and their brain signals.

Using gender differences as an effective tool for treating sleep disorders was found in a study conducted by Baker et al. [95]. According to their research, women are more susceptible to sleep disorders than men, and women have objectively better sleep than men. Women’s sleep health differs from men’s because of hormonal changes.

4.6 Autism Spectrum Disorder

In autism spectrum disorders (ASD), abnormal neural connections in the brain are responsible for causing the condition. Using EEG signals taken from autistic children in a resting state, Kang et al. [96] introduced four entropy methods.

Identifying multivariate analytical methods is one of the challenges in autism diagnosis. It is difficult to find these methods due to the nonlinear nature of EEG signals. Accordingly, Heunis et al. [97] have examined different analytical methods to address this challenge.

As shown by Sundaresan et al. [98], psychological stress can worsen autism symptoms. For the first time, they were able to classify stress states in autistic children using the BCI system in order to determine the level of mental stress and severity of autism.

Gabard-Durnam et al. [99] discovered that autistic children can be diagnosed based on their EEG signal strength at birth. As a result of this method, new biomarkers for early diagnosis have been introduced.

Brain abnormalities can also be measured using this method. Using inter-brain communication abnormalities as a biomarker for autism diagnosis, Lbrahim et al. [100] were able to diagnose autism with 94.6% accuracy.

Jia et al. [101] also suggest that the analysis of unique spatial microstates of the brain can be used to identify deviant brain functions.

5 Conclusion

Liu et al. [102] have introduced a dynamic correlation interactive network (DCENet) for short-term traffic prediction that obtains a new node structure without creating a new graph structure. For neural messages, an encoder and encoder module can be designed using this system. Nowadays, 3D reconstructions of neural networks are of great interest. The Encapsulated Attention Encoder-Decoder Network (EA-EDNet) was introduced by Deng et al. [103], which allows reconstruction in low-light environments. The system can be used for internal imaging and brain recognition. In recent years, graph features have been used to analyze brain networks. Hossieni et al. [104] introduced a system that could decode hand movements by identifying patterns in six frequency bands. The BCI system offers potential for neurological disorders treatment. However, its widespread implementation is impeded by technological and ethical challenges. When considering technological challenges, the crucial factors to consider are speed and precision. Apart from inequality, the ethical dilemmas encompass concerns regarding privacy, the consent of locked-in patients, individual autonomy, and various other societal predicaments. Prior to the widespread adoption of BCI technologies, it is imperative to engage in comprehensive discussions regarding these matters. BCI systems can be improved in the future with the development of this system. BCI challenges include decoding continuous hand movements. It was Hossieni et al. [105] who introduced the ability to do such a task by integrating a continuous decoder with a discrete decoder. It is estimated that this system is capable of classifying 97.1% of the data. Hopefully, this accuracy will increase with the development of this system in the future.