Due to the presence of artifacts and contamination, EEG data arestill not being used for large-scale BCI studies (Pedroni et al., 2019). Even though some deep learning studies for EEG-based BCI say they did not use any preprocessing steps, most of the time, preprocessing steps are very important. Some research works combine the preprocessing steps in their deep learning pipeline and call it as end-to-end framework (Antoniades et al., 2018; Aznan et al., 2018; Zhang et al., 2021). Moreover, an additional CNN layer has been used for the preprocessing in some cases (Amin et al., 2019a; Tang et al., 2019).

Most of the time, frequency domain filters were used in research to limit the bandwidth of the EEG signal. This is useful when there is a specific frequency range of interest so that the rest can be safely ignored (Islam et al., 2016; Kilicarslan et al., 2016). In 30% of the studies, a signal below 45 Hz, or below a typical low gamma band, was low-pass filtered. The filtered frequency ranges were grouped by task type and artifact reduction methods. It shows that most research used a technique to get rid of artifacts along with lowering the frequency ranges that were studied.

From our observation, 20% of the studies manually eliminated artifacts (Rammy et al., 2020; Atilla and Alimardani, 2021; Sundaresan et al., 2021). It is easy to see unexpected outliers visually, such as when data are missing or significant EEG artifacts are evident. But it might be hard to tell the difference between noisy channels that are always on and noisy channels that are only noisy sometimes. Furthermore, since the way the data are processed is very random, it is hard for other researchers to copy the methods. In addition to this, 10% of the studies did not routinely eliminate EEG artifacts. Independent component analysis (ICA) (Delorme et al., 2007) and discrete wavelet transformation (DWT) were the most common artifact removal methods that were utilized in the remaining two-thirds of the analyzed research (Kline et al., 2015).

The EEG electrodes also take up undesired electrical physiological signals from eye blinks and neck muscles (Crespo-Garcia et al., 2008; Amin et al., 2019b). Additionally, there are issues with motion artifacts brought on by cable motion and electrode displacement while the individual is moving (Arnau-González et al., 2017; Chen et al., 2019a; Gao et al., 2019). There have been a lot of studies performed on how to find and remove EEG artifacts (Nathan and Contreras-Vidal, 2016), but it is not the primary focus of our review work. In summary, one of the three methods (i.e., manual process, automated process, or no removal of artifact) is considered in each study to conduct the artifact removal procedure.

One of the main limitations of the classical EEG-based BCI is the number of subjects who participated in this study. Within the course of this review, EEG-based datasets were covered. This scope was taken into account as keywords to find the right research articles on the Google Scholar and Research Gate websites. For this literature review, more than 100 research studies were found on these two websites by using the above criteria. Among these, around 47% of research has been conducted based on the BCI competition dataset. Moreover, 9%, 16%, and 7% of the studies have been conducted on DRYAD, SEED-VIG, and EEG MI datasets, respectively (Figure 5).

Deep learning has enabled larger datasets and more rigorous experiments in BCI. “How much data is enough data?” remains a significant question when using DL with EEG data. We looked at numerous descriptive dimensions to investigate this question: the number of participants, the amount of EEG data collected, and the task of the datasets. There are few studies that make use of their own collected datasets (Tang et al., 2017; Vilamala et al., 2017; Antoniades et al., 2018; Aznan et al., 2018; Behncke et al., 2018; El-Fiqi et al., 2018; Nguyen and Chung, 2018; Alazrai et al., 2019; Chen et al., 2019b; Fahimi et al., 2019; Hussein et al., 2019; Zgallai et al., 2019; Gao et al., 2020b; León et al., 2020; Maiorana, 2020; Penchina et al., 2020; Tortora et al., 2020; Atilla and Alimardani, 2021; Cai et al., 2021; Cho et al., 2021; Mai et al., 2021; Mammone et al., 2021; Petoku and Capi, 2021; Reddy et al., 2021; Shoeibi et al., 2021; Sundaresan et al., 2021; Ak et al., 2022). However, most of the deep learning studies have been conducted based on publicly available EEG datasets, such as:

Deep Neural Networks (DNNs) are highly structured and therefore can learn features from unrefined or modestly heavily processed data, minimizing the need for domain-specific processing and feature extraction processes. Furthermore, DNN-learned attributes may be even more proficient or evocative than human-designed attributes. Second, as in many realms where DL has surpassed the previous condition, it has the potential to improve the effectiveness of other analyses and classifications. Third, DL makes it easier to make tasks such as conceptual sculpting and domain acclimation, which are not tried as often and fail less often when using EEG data. Deep learning techniques have made it feasible to synthesize high-dimensional structured data, such as images and audio.

Deep learning-based methods have been used to sum up high-dimensional, well-organized content such as images and speech. Computational methods could be used by readers to grasp transitional depictions or complement data. Deep neural networks combined with techniques such as linkage synchronization make it easier to learn representations that do not depend on the domain, while keeping information about the task for domain adaptation. Similar methods can be implemented with EEG data to obtain more accurate depictions, and as a result, improve the performance of EEG-based models across a wide range of subjects and tasks.

Various deep learning algorithms have been employed in EEG-based BCI applications, whereas CNN is clearly the most frequent one. For example, Arnau-González et al. (2017), Tang et al. (2017), Vilamala et al. (2017), Antoniades et al. (2018), Aznan et al. (2018), Behncke et al. (2018), Dose et al. (2018), El-Fiqi et al. (2018), Nguyen and Chung (2018), Völker et al. (2018), Alazrai et al. (2019), Amber et al. (2019), Amin et al. (2019b), Chen et al. (2019a,b), Fahimi et al. (2019), Gao et al. (2019), Olivas-Padilla and Chacon-Murguia (2019), Ozdemir et al. (2019), Roy et al. (2019), Song et al. (2019), Tayeb et al. (2019), Zgallai et al. (2019), Zhao et al. (2019), Aldayel et al. (2020), Gao et al. (2020a,b), Hwang et al. (2020), Ko et al. (2020), Li Y. et al. (2020), Liu J. et al. (2020), Miao et al. (2020), Oh et al. (2020), Polat and Özerdem (2020), Atilla and Alimardani (2021), Cai et al. (2021), Dang et al. (2021), Deng et al. (2021), Huang et al. (2021), Ieracitano et al. (2021), Mai et al. (2021), Mammone et al. (2021), Petoku and Capi (2021), Reddy et al. (2021), Tiwari et al. (2021), Zhang et al. (2021), Ak et al. (2022), and, Huang et al. (2022) have explored deep learning-based algorithms. However, more recent BCI studies have implemented other deep learning modalities including,

Different classification algorithms give different maximum accuracy values for different datasets, as shown in Table 3. The LSTM algorithm gave the highest accuracy, which was based on the BCI competition dataset. All researchers achieved an accuracy of over 80% for this dataset, that is, this dataset has the highest accuracy so far. We have found the highest classification accuracy for any algorithm on the BCI competition dataset from various studies, as shown in Figure 6.

For the DEAP dataset (Koelstra et al., 2011), all researchers achieved an accuracy of roughly over 90% (Figure 7), that is, this dataset has the highest reliability so far. Unlike the previous dataset, this one has received little attention in terms of deep learning applications. As with the previous two datasets, there are a few works on the SEED dataset. However, the published works have achieved over 90% accuracy based on CNN or CNN+SAE (Gao et al., 2020a; Hwang et al., 2020; Liu J. et al., 2020). We can apply smarter algorithms to this dataset to explore further.

Due to insufficient work on the rest of the datasets shown in Figure 8, we cannot comment on them. However, we think that whether the accuracy can be increased on the rest of the dataset, it can be worked on in future.

Among the 110 publications that have been studied in this study, discriminative models, particularly CNN, are utilized most frequently. This is right since almost all BCI problems can be put into the category of classification problems. More than 75% of the models are powered by CNN algorithms, and we can summarize them as follows: (i) CNN can use EEG data to find hidden features and spatial correlations that can be used to classify something. As a result, CNN structures are used for classification in certain research while features are engineered in others; (ii) CNN has had considerable success in various research areas (especially in imaging and computer vision domains), making it exceedingly well-known and simple to use (through the available public code). Surprisingly, several BCI techniques naturally produce two-dimensional visuals that can be processed by CNN, and EEG data could be converted into two-dimensional images in the meantime for additional processing by CNN.

On the contrary, only 15% of the model-based articles used a recurrent neural network (RNN), even though RNN is capable of predicting temporal feature learning. One likely reason for this is that it takes time for an RNN to process a long sequence, and EEG signals are often long sequences. EEG signals, for example, are typically divided into 30-s segments with 2,500 time points at a 120 Hz sampling rate. Moreover, RNN takes more than 25 times as long to train as CNN for a sequence of 2,500 items.

Furthermore, among the typical models, the deep belief network (DBN), particularly the DBN-restricted Boltzmann machine (RBM), is the most often used model for feature extraction. DBN is commonly utilized in BCI for two reasons: 1) It is an efficient way to learn the top-down generative parameters that show how variables in one layer depend on variables in the layer above. 2) The values of the latent variables in each layer can be guessed by a single bottom-up pass that starts with an observed data vector in the bottom layer and uses the generative weights in the opposite direction. But most of the work that used the DBN-RBM model was published before 2018, which shows that DBN is not popular right now. Before 2018, researchers used DBN to learn about features, and then a classifier that did not use deep learning. Now, more and more studies use CNN or hybrid models for both learning about features and classifying them.

Finally, there are nine articles suggesting hybrid models for BCI research. Combinations of RNN and CNN account for approximately a third. It is logical to integrate RNN and CNN for both temporal and spatial feature learning, given that RNN and CNN are renowned for their exceptional temporal and spatial feature extraction capabilities. Combining representative and discriminative models is yet another sort of hybrid model. This is easy to understand since the first is used to pull out features and the second is used to put things into groups. There are nine articles that use this form of hybrid deep learning model, which encompasses almost all types of BCI signals. In addition, 12 studies have presented alternative hybrid models, including two discriminative ones. Several research, for instance, have advocated the combination of CNN with MLP in which the CNN structure is utilized to extract spatial data that are then given to an MLP for classification.

Deep learning-based BCI systems are mostly used in the healthcare industry to identify and diagnose mental illnesses, including epilepsy, Alzheimer's disease, and other disorders (Dose et al., 2018). First, research focusing on sleep-stage recognition based on sleeping spontaneous EEG is utilized to identify sleeping disorders (Vallabhaneni et al., 2021). As a result, the researchers do not need to seek out patients with sleeping issues since it is simple to gather the sleeping EEG signals from healthy people in this condition. The diagnosis of epileptic seizures has also garnered a great deal of interest. The majority of seizure detection is dependent on spontaneous EEG and mental illness signs (Antoniades et al., 2018; Hussein et al., 2019; Dang et al., 2021; Shoeibi et al., 2021). CNN and RNN are common deep learning models in this context, as are hybrid models that combine RNN and CNN. Several methods (Turner et al., 2014) combined deep learning models for feature extraction with classical classifiers for detection. To diagnose seizures, researchers used a VAE in feature engineering followed by SVM.

Smart environments are a possible future application scenario for BCI. With the rise of the Internet of Things (IoT), BCI can be linked to a growing number of smart settings. For instance, an aiding robot may be used in a smart house (Zhang et al., 2018c) in which the robot is controlled by brain impulses. In addition, Behncke et al. (2018) examined how to operate a robot using visually stimulated spontaneous EEG and fNIRS data. BCI-controlled exoskeletons might assist individuals with compromised lower limb motor control in walking and everyday activities (Kwak et al., 2017). In future, research on brain-controlled equipment may be useful for developing smart homes and smart hospitals for the elderly and the crippled.

In comparison to other human–machine interface approaches, the greatest benefit of BCI is that it allows patients who have lost most motor functions, such as speaking, to interact with the outside world (Nguyen and Chung, 2018). Deep learning technology has considerably enhanced the efficiency of the brain's signal-based communications. The P300 speller is a common paradigm that allows individuals to type without a motor system, which can turn the user's intent into text (Cecotti and Graser, 2010). In addition, Zhang et al. (2018b) suggested a hybrid model that combines RNN, CNN, and AE to extract relevant characteristics from MI EEG to detect the letter the user intends to write. The suggested interface consists of 27 characters (26 English alphabets and the space bar) split into three character blocks (each block containing nine characters) in the first interface. There are three possible choices, and each one leads to a separate sub-interface with nine characters.

A prominent topic of interest for BCI researchers is the security industry. A security issue may be broken down into authentication (also known as “verification”) and identity (also known as “recognition”) components (Arnau-González et al., 2017; El-Fiqi et al., 2018; Chen et al., 2019b; Puengdang et al., 2019; Maiorana, 2020). The goal of the former, which is often a multi-class classification task, is to identify the test subject (Zhang et al., 2017). This is usually a simple yes-or-no question that only looks at whether the test subject is allowed or not. Existing biometric identification/authentication systems rely primarily on the unique inherent physiological characteristics of people (e.g., face, iris, retina, voice, and fingerprint). Anti-surveillance prosthetic masks that may defy face recognition, contact lenses that can fool iris detection, vocoders that can compromise speech identification, and fingerprint films that can fool fingerprint sensors are all vulnerable. Due to their great attack resilience, EEG-based biometric person identification systems are emerging as attractive alternatives. Individual EEG waves are almost impossible for an impostor to replicate, making this method extremely resistant to spoofing assaults faced by other identification methods. Deep neural networks were used by Mao et al. (2017) to identify the user's ID based on EEG signals, and CNN was used for personal identification. Zhang et al. (2017) presented and analyzed an attention-based LSTM model on both public and local datasets. The researchers (Zhang et al., 2018a) subsequently merged EEG signals with gait data to develop a dual-authentication system using a hybrid deep learning model.

Several articles simply aim to categorize the user's emotional state as a binary (positive/negative) or three-category (positive, neutral, and negative) issue using deep learning algorithms (Chen et al., 2019b; Ozdemir et al., 2019; Gao et al., 2020a,b; Hwang et al., 2020; Liu J. et al., 2020; Liu Y. et al., 2020; Sundaresan et al., 2021). Diverse articles employed CNN and its modifications to identify emotional EEG data (Li et al., 2016) and lie detection (Amber et al., 2019). Most of the time, the CNN-RNN deep learning model is used to find hidden traits in spontaneous emotional EEG. Using EEG data, Xu and Plataniotis (2016) employed a deep belief network (DBN) as a particular feature extractor for the emotional state categorization task. Moreover, on a more basic level, some studies seek to identify a positive/negative condition for each emotional dimension. For identifying emotions, Yin et al. (2017) suggested a multiple-fusion-layer-based ensemble classifier of AE. Each AE is made up of three hidden layers that remove unwanted noise from the physiological data and give accurate representations of the features.

For traffic safety to be assured, a driver must be able to keep up their best performance and pay close attention. It has been shown that EEG signals may be beneficial in assessing people's cognitive status while doing certain activities (Almogbel et al., 2018). A motorist is often considered alert if their response time is less than or equal to 0.7 s and weary if their reaction time is more than or equal to 2.1 s. By extracting the distinctive elements from the EEG data, Hajinoroozi et al. (2015) investigated the prediction of a driver's weariness. They investigated a DBN-based dimensionality reduction strategy. It is important to be able to tell when a driver is tired since that can make accidents more likely. Furthermore, it is practical to identify driver weariness in daily life. The technology that is used to record EEG data is easy to find and small enough to use in a car. In addition, the cost of an EEG headset is reasonable for the majority of individuals. Deep learning algorithms have greatly improved the accuracy of tiredness detection. In conclusion, driving sleepiness based on EEG may be identified with excellent accuracy (83–98%) (Fahimi et al., 2019; Ko et al., 2020; Atilla and Alimardani, 2021; Cai et al., 2021). The self-driving situation is where driver fatigue monitoring will likely be used in future. Since the human driver is often expected to react correctly to a request to intervene in most self-driving scenarios, the driver must always be aware. As a result, we think that using BCI-based drive fatigue detection can help the development of autonomous vehicles.

Human operators play an important role in automation systems for decision-making and strategy formulation. Human functional states, unlike those of machines or computers, cannot always meet the needs of a task because working memory is limited, and psycho-physiological experience changes over time. A lot of researchers have concentrated on this subject. The mental effort may be calculated using spontaneous EEG. Bashivan et al. (2015) introduced a DBN model, a statistical technique for predicting cognitive load from single trial EEG.

However, there are still plenty of deep learning premises and domains to be used in EEG-based BCI, which will not only improve the performance but also make them more generalizable. Here are a few suggestions for future researchers regarding where they can uncover novelty utilizing deep learning.