Recent Progress in Wearable Brain–Computer Interface (BCI) Devices Based on Electroencephalogram (EEG) for Medical Applications: A Review

Importance: Brain–computer interface (BCI) decodes and converts brain signals into machine instructions to interoperate with the external world. However, limited by the implantation risks of invasive BCIs and the operational complexity of conventional noninvasive BCIs, applications of BCIs are mainly used in laboratory or clinical environments, which are not conducive to the daily use of BCI devices. With the increasing demand for intelligent medical care, the development of wearable BCI systems is necessary. Highlights: Based on the scalp-electroencephalogram (EEG), forehead-EEG, and ear-EEG, the state-of-the-art wearable BCI devices for disease management and patient assistance are reviewed. This paper focuses on the EEG acquisition equipment of the novel wearable BCI devices and summarizes the development direction of wearable EEG-based BCI devices. Conclusions: BCI devices play an essential role in the medical field. This review briefly summarizes novel wearable EEG-based BCIs applied in the medical field and the latest progress in related technologies, emphasizing its potential to help doctors, patients, and caregivers better understand and utilize BCI devices.


Introduction
Brain-computer interface (BCI) establishes a connection bet ween the brain and the machine to replace, restore, supplement, or enhance brain functions.A BCI system has 4 functional mod ules: signal acquisition, data preprocessing, feature extraction and classification, and output equipment [1].According to the brain signal acquisition method, there are 3 main types of BCI: invasive BCI, semiinvasive BCI, and noninvasive BCI [2].Invasive BCIs realize external devices controlled by decoding neuronal activities inside the brain.Since the recording elec trodes are close to the brain signals' source and unaffected by the brain tissues' attenuation and filtering, invasive BCIs can record and analyze the information in neural signals with the highest temporal-spatial resolution and accuracy.At present, invasive BCIs have been applied in motion control [3][4][5][6], disease diagnosis and treatment [7][8][9], communication assistance [10][11][12][13], cursor control [14][15][16], and other aspects.However, invasive BCIs require the surgical implantation of signal acquisition devices into the brain, which may bring risks of damage or infec tion to the brain.In addition, the immune response surrounding the implanted electrode will decrease signal quality over time, degrading the performance and lifespan of invasive BCIs [2].To minimize the damage to the brain, semiinvasive BCIs have been proposed.The electrodes of semiinvasive BCIs are usually placed under the skull or below the dura mate to record electrocorti cography (ECoG).Compared with invasive BCIs, semiinvasive BCIs are less invasive to the brain.Invasive BCIs and semiinvasive BCIs can be collectively referred to as invasive BCIs.
To avoid the risk of infection and brain damage caused by electrode implantation, noninvasive BCIs have been proposed.Brain signal acquisition of noninvasive BCIs does not require surgical implantation of electrodes, which has a low risk of brain damage and can reduce the psychological and physical burden of the users.Standard noninvasive techniques include electroencephalogram (EEG), magnetoencephalogram (MEG), functional nearinfrared imaging (fNIR), and functional mag netic resonance imaging (fMRI) [2].EEG waves are clusters of electrical signals from neurons in the brain and are usually obtained through electrodes placed on the scalp, forehead, and behind the ears [17], while MEG, fMRI, and fNIR need to be collected with bulky specialized acquisition equipment.
For medical applications of BCIs, intermittent neurological diseases (epilepsy, migraine, etc.) require longterm continuous monitoring with BCI systems to diagnose and predict, and the assistive devices (prosthetics, robotic arms, wheelchairs, etc.) controlled by BCI systems must be used in real scenarios.Long term continuous signal monitoring and daily device control rely on the wearability of BCIs.It is of great importance to make BCIs wearable.Wearability is defined as a performance that can be carried out in daily activities without notably obstructing the subject and affecting the devices' operation.For ease of use, wearable devices are mainly placed outside the user's body and can be worn and taken off by users independently.Invasive BCIs require surgical implantation of signal acquisition electrodes and are normally defined as implantable devices.Although some wearable BCI systems are composed of semiinvasive brain sig nal acquisition devices and wearable actuators [18][19][20], the rel evant work is limited.Noninvasive BCIs based on MEG, fMRI, and fNIR are limited by large signal acquisition devices and can only be used in hospitals and other specific places, which cannot meet the requirements for daily wearable use.On the contrary, EEGbased BCI devices do not need electrode implantation for signal acquisition and the use is not restricted by the venue, making EEGbased BCIs suitable for daily wearable applications.Hence, EEGbased BCIs for wearable medical applications are mainly focused in this review.
Nowadays, wearable EEGbased BCIs mainly rely on the wet electrode EEG caps for stable and reliable EEG signal acquisition and have been applied to the diagnosis and auxiliary treatment of various diseases, such as disorders of consciousness [21][22][23], Parkinson's disease [24], paralysis [25][26][27], stroke [28][29][30], de pression [31], autism [32], and sleep disorders [33].Besides, EEGbased BCIs are also used for wheelchair control [34][35][36].However, wet electrodes require skin preparation and the appli cation of conductive gels for low electrodeskin impedances and highquality EEG signals are obtained.Because the conductivity of the gel decreases with time, it is hard to realize longterm EEG signal acquisition with the wet electrode caps.Moreover, con ventional EEGbased BCI systems also have limitations such as bulky size, complex wires, and dependence on large apparatus (e.g., the EEG amplifier), which increase the discomfort of wear ing and limit the daily use of EEGbased BCIs.
For wearable BCI systems, brain signal acquisition equip ment is of great importance not only because the quality and reliability of brain signal acquisition will affect the performance of the BCI systems but also the signalacquiring methods will influence the wearability of the BCIs to some extent.Therefore, this topic review introduces wearable EEGbased BCIs from the perspective of EEG acquisition equipment and focuses on their applications in the medical field.According to the distri bution of recording electrodes, EEGbased BCIs reviewed in this paper are divided into scalpEEGbased BCIs, forehead EEGbased BCIs, and earEEGbased BCIs, as shown in Fig. 1.The applications of wearable EEGbased BCIs in the medical field are summarized as disease management (including disease prevention and diagnosis), rehabilitation therapy, health mon itoring, communication assistance, and equipment control.

Scalp-EEG-Based BCI
Brain signals of EEGbased BCIs are usually obtained from the scalp.Compared with the forehead and behind/inear areas, the scalp offers more sampling space, which is conducive to arranging more EEG signal recording points.The number of signalrecording electrodes ranges from a few (for targeted BCI applications) to 256 [2].In the past 2 decades, many scalpEEG based BCIs with conventional EEG caps have been applied in medical, education, military, and other fields.However, due to inconvenience of skin preparation and conductive gels, and the complexity of wire connection, the EEGbased BCIs with con ventional EEG caps cannot meet the requirements of daily use.
ScalpEEG recording has experienced a long development history.To improve the wearability of the scalpEEGbased BCIs, scalpEEG acquisition devices with unique structures and mate rials have been proposed.In this section, 3 commercial scalp EEG headsets, Emotiv EPOC (Fig. 2A), DSI24 (Fig. 2B), and OpenBCI Ultracortex "Mark IV" (Fig. 2C), and their wearable BCI applications in the medical field are introduced in detail.

Emotiv EPOC
Emotiv EPOC is a typical consumeroriented EEG equipment released by Emotiv Systems Company of the United States.Emotiv EPOC is an octopus structure that gives users good flex ibility and movement.The device has 14 salinebased recording electrodes and 2 reference electrodes [37].The headset can wire lessly connect to computer or mobile devices and continuously work for up to 12 h.Emotiv EPOC has been widely used in research, since its development [38].
Here, we review some of the representative wearable BCI work by using Emotiv EPOC.For the neurorehabilitation of stroke patients, Jure et al. [39] presented a functional electrical stimulation (FES)based BCI system (Fig. 2D), which was made up of an Emotiv EPOC headset for EEG signal recording and a 2channel controlled stimulator for neuromuscular system stimulation.When the BCI system detected cerebral activities related to motor imagery (MI), the electrical stimulator would be activated to realize therapeutic intervention.Tabernig et al. [40] used the proposed BCIFES system to perform neuro rehabilitation therapy for patients with sequelae of ischemic stroke and evaluated the effects.The cerebral cortex activation during the presence of MI and the sensory feedback produced by the movement were used to facilitate neuroplasticity.Before and after the intervention, the upper limb was assessed by the Fugl-Meyer score, and marked posttreatment improvement was detected (Fig. 2E), suggesting that the proposed therapy could benefit stroke individuals' neurorehabilitation.  [43].©2016 EDP Sciences.Reprinted with permission from Swee et al. [43].(B) DSI-24 headset (https://www.neurospec.com/Products/Details/1079/dsi-24).(C) OpenBCI Ultracortex "Mark IV" headset (https://docs.openbci.com/AddOns/Headwear/MarkIV/).(D) Block diagram of the BCI-FES system [39].©2016 IOPscience.Reprinted with permission from Jure et al. [39].(E) Scores of the quality of movement (left) and quality of life (right) were measured for each stroke patient [40].Used with permission from Tabernig et al. [40]; permission conveyed through Copyright Clearance Center Inc. (F) BCI-NFB system (left) and power spectrum density in one representative participant (right) [41].©2019 BMC.Reprinted with permission from Al-Taleb et al. [41].(G) User with Emotiv EPOC headset, BCI program, and FES electrodes on the arm [42].©2021 BMC.Reprinted with permission from Zulauf-Czaja et al. [42].(H) The percentage of true positive activation (left) and time to activate FES (right) out of all attempted trials [42].©2021 BMC.Reprinted with permission from Zulauf-Czaja et al. [42].(I) Illustration of mu suppression in affected and unaffected hemispheres [52].©2019 MDPI.Reprinted with permission from Choi et al. [52].
Besides the neurological rehabilitation of stroke patients, wearable BCI with the Emotiv EPOC can also be used to detect central neuropathic pain (CNP), which is a frequent chronic condition in spinal cord injury (SCI) patients.AlTaleb et al. [41] designed a BCI system based on the Emotiv EPOC device (Fig. 2F, left) for selfmanaged neurofeedback (NFB) treatment of people with chronic SCI.According to the visual feedback of selected frequency EEG band power, users had to selfregu late their primary motor cortex brain activity with the BCINFB system.Results showed that users had successfully regulated their brainwaves in a frequencyspecific manner (Fig. 2F, right).The reduction in pain experienced was clinically obvious (greater than 30%) in 8 participants, demonstrating that the BCINFB system could reduce CNP in people with SCI.In 2021, the same research team presented a BCIFES system based on an Emotiv EPOC headset (Fig. 2G) for hand function rehabilitation and evaluated its usability [42].Hand therapy was performed by producing the attempted movement of one hand to lower the 8 to 12Hz frequency band power and activate FES to induce wrist flexion and extension.The system obtained an accuracy of 70 to 90%, and the median activation time of FES remained constant across sessions (Fig. 2H).
BCIs based on the Emotiv EPOC headset also play an impor tant role in patient assistance.To meet the needs of some para lyzed patients for wheelchair control without joysticks, different BCIcontrolled wheelchairs have been developed.In 2016, Swee et al. [43] developed a brainwavecontrolled wheelchair with an Emotiv EPOC headset.Based on the EEG signal, the electric wheelchair performed the desired movement and achieved up to 90% accuracy [44].Voznenko et al. [45] developed a robotic wheelchair controlled by the onboard computer that received commands from the extended Emotiv EPOC BCI.To improve system control accuracy, the robotic wheelchair could also be controlled by voice and gestures.Zgallai et al. [46] designed a smart wheelchair for paralyzed people who are unable to con trol their bodies.The wheelchair BCI system completed 4 kinds of command recognition and achieved an accuracy rate of up to 96%.Shahin et al. [47] proposed a wheelchair control BCI system that could switch between the automatic control mode and the manual control mode.Three types of input, EEG sig nals, head gestures, and facial expressions, were collected and translated into 4 control instructions.
Besides direction control, Bousseta et al. [48] designed a novel BCI system with an Emotiv EPOC headset for robotic arm control.Participants were instructed to imagine the exe cution of hand or foot movements.After command translation, subjects achieved control of the robotic arm in 4 directions and obtained an average accuracy of 85.45%, which had the potential to provide a helpful aid for the disabled.

DSI-24
DSI24 is a wireless EEG headset developed by the Wearable Sensing Company of the United States.DSI24 has 21 signal recording dry electrodes, including 19 electrodes on the head for full head coverage, 2 ear clip electrodes, and 3 auxiliary sen sors [49].The electrodes are springloaded to provide constant and comfortable pressure, which enhances the contact between electrodes and the skin and, at the same time, reduces motion artifacts.With the unique structure, DSI24 can obtain the EEG signal with a quality comparable to that of wet electrodes even with dry electrodes.The electrodes use active/passive shielding technology to prevent electromagnetic interference.
Eldeeb et al. [50] developed an EEGbased BCI system to analyze the distress effect on the brain activity of autism spec trum disorder (ADS) individuals [51].Based on the affective Posner task, the proposed BCI identified the patterns associated with emotion regulation.The EEG signals were obtained from 21 ADS with DSI24 EEG headsets.Choi et al. [52] designed an action observation BCI system based on the DSI24 EEG head set and detected the participants' attention level by analyzing Mu rhythm (i.e., alpha wave, 8 to 13 Hz) power when watching a video of repetitive grasping actions.The system provided the steadystate visual evoked potentials (SSVEPs) as the feedback.Results showed that, compared with conventional action obser vation (AO), the proposed BCIAO suppressed stroke patients' Mu rhythm more (Fig. 2I), suggesting that the proposed para digm was an effective tool for stroke patients' rehabilitation.Kim et al. [53] combined BCIAO with peripheral electrical stimu lation (PES) and assessed the effect of the system on corticos pinal plasticity for motor recovery.In this work, participants watched a video of repeated gripping movements under 4 dif ferent tasks, and visual feedback was provided by BCI.Motor evoked potentials (MEPs) were measured during the task.According to the results, 4 tasks all realized that the MEP latency decreased and had the potential in promoting corticospinal plasticity in stroke patients.
Zhang et al. [54] designed an SSVEPBCI system for robotic arm control in patient assistance.To improve the system oper ation effectiveness, an adaptive decoding FBCCA algorithm was adopted, which could adapt to individual differences.In this work, an average recognition success rate of 95.5% was obtained, proving that the proposed system allowed the hand icapped to grasp objects by controlling the mechanical arm through the brain in daily life.

OpenBCI Ultracortex "Mark IV"
Ultracortex "Mark IV" is an EEG headset designed to work with OpenBCI boards.The headset can record researchgrade EEG signals and sample up to 16 channels from all 35 recording locations [55].Spikey (for the scalp with hair) and nonspikey (for the scalp without hair) dry electrodes could be screwed at the locations.Since the Ultracortex "Mark IV" headset could be 3Dprinted in different sizes, it is adjustable for different head shapes and sizes and enables all electrodes to contact the scalp closely.Compared with other consumergrade EEG head sets, the OpenBCI "Mark IV" is costeffective [56].The BCIs based on the Ultracortex "Mark IV" EEG headset are mainly used for emotion recognition [57][58][59] and assistance technol ogy [60][61][62][63].
To help disabled persons control a robotic arm, Lim and Quan [60] presented an EEGbased BCI system for robotic arm con trol.EEG signals were classified into 8 mental commands with convolutional neural network (CNN) model to realize 6degree offreedom operations.The recall rate and precision of the sys tem were 91.9% and 92%.Saragih et al. [61] applied CNN and long shortterm memory (LSTM) networks in an EEGbased BCI system for effectively controlling the artificial hand.In the hand operations classification, the accuracy of the CNN model was 95.45%, while that of the LSTM model was 93.64%.Bolaños et al. [62] created a room prototype that allowed people with motor disabilities to realize the control of the light and bed, and request assistance with a buzzer through an EEGbased BCI.EEG data recorded with Ultracortex "Mark IV" headset were filtered in the alpha band to train a onedimensional (1D) CNN model.The proposed system achieved a realtime classification accuracy of 78.75%.
To enhance cognitive control, Dutta et al. [63] developed a BCIbased application with interactive media.Corresponding alpha and beta waves were obtained from the recorded EEG signals and regarded as parameters of attention to move the object in the proposed game.This work provided doctors with a new choice for psychological disorder treatment.
Although the wearable BCIs based on scalpEEG have been widely used, the influence of hair on EEG signal acquisition cannot be ignored.Obtaining highquality EEG signals on the scalp is challenging without the aid of conductive agents.Even if the spring pressurized structure is applied in scalpEEG head sets to improve the contact between the electrodes and scalp, the contact impedance is still high, which cannot guarantee longterm stable EEG recording and is easily affected by motion artifacts.

Forehead-EEG-Based BCI
ForeheadEEGbased BCI converts the EEG signals recorded from the forehead area into commands to realize actuators' con trol.The forehead is an ideal place for electrodes attaching as a nonhairbearing area.Recording EEG on the forehead requires no complex preparation work prior to signal acquisition and avoids the interference from hair.Besides, foreheadEEG includes rich information associated with cognitive abilities and dysfunc tions [64].Given this unique feature, the research on wearable foreheadEEGbased BCIs has gained wide attention in recent years, and foreheadEEG BCIs are often used for attention mon itoring and then to realize functions such as sleep monitoring, cursor control, and external device control.
Nowadays, many existing consumergrade wearable BCIs and clinical applications rely on foreheadEEG.This section introduces wearable BCI systems' development and medical application based on foreheadEEG acquisition devices.Three foreheadEEG acquisition devices introduced in this section are shown in Fig. 3.

Mindo EEG device
The Mindo 4S EEG device developed by the Brain Research Center of National Chiao Tung University consists of 4channel dry recording electrodes and 2 extra reference channels (Fig. 3A).To realize appropriate electrodeskin contact impedance, springloaded dry electrodes were adopted in the Mindo 4S EEG device, whose wearing position can be adjusted [64].The flexible substrate acting as a buffer can eliminate pain when a force is applied, which is conducive to wear comfort and long term EEG acquisition.The sampling rate of Mindo 4S is up to 512 Hz and can be adjusted according to the system's require ments.The system uses Bluetooth to realize wireless data trans mission and can work continuously for 20 h [65].
In 2017, Lin et al. [64] proposed a siliconbased dry sensor for foreheadEEG acquisition, the Mindo 4S EEG device pro totype.The proposed system successfully realized 5 sleep stages identification, headache prevention, and a rapid antidepressant agent assessment.In 2018, Cao et al. [66] from the same research team used the Mindo 4S EEG device to assess the ketamine effect in patients with treatmentresistant depression.The band power and asymmetry of the alpha band as the feedback of the BCI system were detected.Results showed that the BCI system classified the responders and nonresponders with 81.3 ± 9.5% accuracy based on the support vector machine with radial basis function (SVMRBF) predictor.Compared with the previous work, the accuracy improved.
Chiu et al. [67] built an SSVEPbased BCI eating assistive system (Fig. 3D).In this system, different frequencies of visual stimulation caused EEG waves with specific characteristics so as to achieve different functional selections.Users can also use this system for subjective training to optimize the current user's model.The SSVEPbased BCI enabled the disabled to have meals by themselves, obtaining 91.35% average accuracy and 20.69 bits per minute information transfer rate.To further improve the practicability of the system, Lin et al. [65] integrated more enter tainment and interaction functions into the system and gained a 90.91% average accuracy.

BrainLink Lite
BrainLink Lite is a headmounted EEG sensor developed by Shenzhen Macrotellect Company for iOS and Android systems.It has 3 goldplated copper dry electrodes, including an EEG recording electrode, a ground electrode, and a reference electrode [68], as the Fig. 3B shows.BrainLink Lite can be easily worn on the forehead.Because the sensor connects to the smart terminals via Bluetooth, it avoids complicated leads and is convenient to acquire EEG signals in nonlaboratory environments.
Japaridze et al. [69] used the BCI based on the BrainLink Lite to detect absence seizures.The BrainLink Lite was used for fore headEEG acquisition, and a predefined algorithm was used for the automated detection of absence seizures in real time.This work obtained an average sensitivity of absence seizure detection of 78.83% with the proposed BCI system, which showed the potential to detect absence seizures with wearable BCIs in every day life.For children with autism NFB training, Mercado et al. [70] designed a BCI video game named FarmerKeeper.The attention of children with autism was read by BrainLink Lite and used to control a runner in the BCI video game.The results showed that the proposed BCI video game improved attention and reduced the anxiety of children with autism.
To restore motor function through robotassisted rehabilita tion therapy, Li et al. [71] introduced a BCI system for wrist rehabilitation (Fig. 3E).In the system, the attention level was measured to activate a flexible wrist exoskeleton for wrist reha bilitation training.The overall actuation success rate was 95%, proving the feasibility of attentionbased control.

MindWave Mobile
MindWave Mobile developed by the NeuroSky Company is made up of a headset, an ear clip, and a sensor arm (Fig. 3C).The only EEG recording electrode of the MindWave Mobile is located on the forehead above the eye (FP1 position accord ing to the 1020 International System) with a reference elec trode and grounding electrode inside the ear clip.Therefore, the MindWave Mobile EEG headset is commonly used for eye blink detection [72][73][74][75].The device outputs 12bit 3 to 100Hz original brainwaves with a sampling rate of 512 Hz and can achieve an 8h battery run time.With only one recording elec trode, MindWave Mobile is more portable and more accessible to wear than multichannel EEG sensors.
For neurotherapy, Mercado et al. [76] used MindWave Mobile for attention detection and used FarmerKeeper [70] as an assistant tool.Results showed that participants who used FarmerKeeper were more focused during NFB sessions.The pre and post assessment indicated that all autistic children improved their attention.This technique helps researchers real ize realtime attention monitoring and regulation through wearable BCIs.
To help the disabled communicate with the outside world, Salih and Abdal [77] designed a BCIbased visual keyboard using the MindWave Mobile headset.Participants were demanded to write "Help" words for 9 sessions on visual keyboards.For print ing proposes, voluntary blinks and attention were detected using EEG signals.With 2 designed visual keyboards, an average text entry speed of about 1.55 to 1.8 words per minute and an error rate of 5 to 5.25% were obtained (Fig. 3F).
However, because only a single recording electrode is placed on the forehead, the richness and accuracy of the EEG signals recorded with the MindWave Mobile headset are limited, and less electrode recording may be more obvious interference by artifacts.The accuracy of the wheelchair control studies [78][79][80] that relied only on EEG signals collected with the MindWave (D) Eating assistive BCI system based on SSVEP [67].©2017 IEEE.Reprinted with permission from Chiu et al. [67].(E) Diagrammatic sketch of brain-controlled wrist rehabilitation BCI [71].Used with permission from Li et al. [71]; permission conveyed through Copyright Clearance Center Inc. (F) Text error rate and entry speed of QWERTY virtual keyboard (left) and ABC virtual keyboard (right) [77].©2020 UAD.Reprinted with permission from Salih et al. [77].
Mobile headset was lower than that of other BCIcontrolled wheelchair studies.To improve the control accuracy, Girase and Deshmukh [81] used not only the EEG signals but also the blinking eye signal as a control reference in the BCIcontrolled wheelchair system, and a control accuracy of about 95% was obtained.
Although the acquisition of foreheadEEG is relatively sim ple because there is no need for complex skin processing, the application of foreheadEEG in wearable BCI may be proble matic since the foreheadEEG is easy to be affected by ocular artifacts and facial muscle artifacts, which will influence the quality of signal acquisition and the accuracy of the whole BCI system.As the forehead region is far from the occipital region and temporal lobe, the characteristic EEG signals of the relevant brain regions are difficult to collect.Moreover, for everyday use, most foreheadEEG acquisition devices are conspicuous and unsuitable for daily wear.

Ear-EEG-Based BCI
EEG signal recorded by earEEGbased BCIs comes from the ear canals or the area behind the ear.The areas in and behind the ears are favorable positions for EEG signal acquisition for nonhair bearing.Compared to other EEG acquisition devices, earEEG acquisition devices are more miniaturized, and only a small part of the head is covered with electrodes, which is more comfortable to wear and is suitable for daily use.In addition, earEEG can map more brain regions than the foreheadEEG.
Recently, many research teams have developed different ear EEG sensors, studied the information contained in earEEG, and designed corresponding wearable BCI systems.According to the position where the earEEG sensor is worn, there are 2 categories earEEGbased BCIs: (a) inear EEGbased BCIs with recording electrodes placed in the ear canals, and (b) behind the ear EEG based BCIs with recording electrodes placed behind the ear.The corresponding 2 types of earEEG acquisition devices are shown in Fig. 4.

In-ear EEG-based BCI
In 2011, Looney et al. [82] from Imperial College London designed an intheear (ITE) electrode (Fig. 4A) for wearable earEEG recording.The proposed ITE system used 2 or more electrodes embedded into an earplug to record EEG signals.
The earplug was produced based on the 3D printing of the ear canal, and the mounted electrodes were made from silver/silver chloride (Ag/AgCl).ITE electrode showed excellent correlation and coherence with onscalp electrodes and was proven to extract several key EEG features, including the auditory steady state response (ASSR), alpha attenuation response (AAR), and P300 paradigms, which illustrated the potential of earEEG in BCI application [83,84].A further study demonstrated that the signaltonoise ratio of the earEEG signal was comparable to that of the EEG recorded in the temporal region [85].
Since ITE earphones need custom earpieces, the cost of time and money is relatively high.In 2015, Goverdovsky et al. from the same laboratory proposed a novel inear sensor (Fig. 4B) for highquality longterm EEG monitoring [86].The inear sensor used viscoelastic substrates and conductive clothes to realize stable electrical contact between the sensors and the ear canals.To establish a lowimpedance contact, only saline solution was required.The inear sensor was proved to be useful in capturing a wide frequency range of EEG and all of the standard EEG responses.Zibrandtsen et al. [87] compared the inear EEG and scalpEEG recorded from patients with suspected temporal lobe epilepsy.Results suggested that earEEG was a reliable signal source to detect EEG patterns associated with focal temporal lobe seizures.
Although there are some feasibility studies on inear EEG for BCI applications [88][89][90], inear EEG has not been widely applied in wearable medical BCIs.

Behind the ear EEG-based BCI
In 2015, based on funding from TSMi (Oldenzaal, The Netherlands), Debener et al. [91] from the University of Oldenburg designed a flexible cshape earEEG acquisition sensor (cEEGrid), as shown in Fig. 4C, which is made up of 10 printed Ag/AgCl electrodes and placed behind the ear.To realize signal recording, the electrode gel was applied to the cEEGrid's electrodes.Researchers proved that the cEEGrid electrodes array could record reliable EEG data for the first time.One year later, the same research team found clear attentionmodulated eventrelated potential (ERP) effects in EEG signals recorded with cEEGrid sensors, which agreed with the sig nals recorded with classical EEG caps in morphology and effect size.The discovery demonstrated that the cEEGrid sensor could measure welldescribed ERPs and be expected to replace classical EEGcap in auditory attention monitoring [92].In a further study, the visual and cognitive ERPs (N1, P1, P300) and eventrelated lateralizations (ERLs) were recorded with cEEGrid, while motor related cortical potentials were not well measured [93,94].
For the hearing impaired, the function of target speaker iden tification in assistive devices is of great importance.In 2016, the research team used EEG signals recorded by the cEEGrid sensor and EEG caps to identify the attended speaker [95].Based on the earEEG, the positive correlation of the performance scores was evident (Fig. 4E).The decoding accuracy of earEEG was 69.3%, suggesting that the cEEGrid sensor has the potential to apply for BCI control of hearing aids.To evaluate the cEEGrid in attention selection, normalhearing and cochlear implant participants were recruited [96].For the cEEGrid data, only half of both the cochlear implant and normalhearing users obtained decoding accuracies above the chance level (Fig. 4F).
Other research teams have also carried out a series of stud ies on wearable BCIs based on the behind the ear EEG.Millard et al. [97] recorded individual peak alpha frequency (PAF) with cEEGrid sensors to predict future pain severity in patients undergoing thoracotomy.Segaert et al. [98] took use of the cEEGrid setups for early language comprehension impairment detection in mild cognitive impairment (MCI).The proposed system showed outstanding classification ability when detect ing patients with MCI from the healthy controls.
In addition to pain prediction and language comprehension impairment detection, behind the ear EEG signal is also used for seizure detection.However, most studies rely on experi enced epileptologists to analyze and annotate the earEEG data, rather than using a complete BCI system [99,100].Swinnen et al. [100] developed a semiautomatic absence seizure detec tion algorithm to label earEEG, which could reduce the con sumption of the recording review time and the workload of epileptologists to some extent.
For home appliance control, Kaongoen et al. [101] designed a novel online BCI system.The system used speech imagery (SI) and earEEG recorded with a custommade wearable BCI headphone (Fig. 4D).Each side of the headphone has 4 record ing semidry electrodes distributed behind the ear.During the online experiments, a few users could control the home appli ance freely with the earEEG SIbased BCI system.
Compared with scalpEEGbased BCIs and foreheadEEG based BCIs, the wearable earEEGbased BCI is still in the stage of practical exploration and improvement and the medical applications are relatively insufficient.The possible limitations and challenges of earEEGbased BCIs are listed as follows.
a. EarEEG signal is weak and can be easily affected by muscle artifacts when the jaw moves, such as speaking and swallowing.For wearable BCI applications, the hardware and software of signal processing need to be further improved.
b.The signals from different brain regions overlap at the ear area, resulting in the difficulty of extracting the characteristic signals.The current methods and technologies for feature extrac tion and classification are still not efficient or accurate enough.
c. Compared with the scalpEEG, the earEEG mapping model to the brain region is not perfect and the algorithm is not yet mature.Perfecting the model and algorithm is the key to expanding the wearable application of the earEEGbased BCIs.

Limitations and Challenges of Wearable EEG-Based BCIs
Currently, wearable BCIs based on scalpEEG, foreheadEEG, and earEEG have played an important role in medical applica tions such as disease diagnosis and prediction, neurological rehabilitation, health monitoring, and auxiliary equipment con trol, providing a new form of medical treatment.However, wear able BCI technology is still in the research stage.At present, the volume of BCI is large and brain signal processing relies on back end equipment such as a computer.Due to the individual differ ences in EEG signal, for each new user, a lot of pretraining and professional guidance are indispensable to achieve a relatively stable BCI performance.Besides, current wearable BCI systems are not robust enough to overcome the interference caused by users and the environment.The high cost of time and money also makes wearable BCIs hard to be widely used.To meet the needs of further application, wearable BCI devices still have a lot of room for improvement in highquality signal acquisition, wireless transmission, information security, wearing comfort, and so on.Various research teams are active in the frontier research of wearable EEGbased BCIs, and the latest research progress is briefly introduced in this section.
Acquisition of EEG signals is an important part of BCI.The quality of signal acquisition directly influences the performance of the BCI system.Since wearable BCIs are commonly used for a relatively long period, longterm stable and reliable EEG signal recording is required.To realize highquality EEG signal acqui sition, different kinds of electrodes have been proposed.For EEG signal recording on a hairy scalp, hair is a natural barrier.Standard Ag/AgCl wet electrode uses conductive gel to enhance the electrical contact between the electrodes and the scalp.However, the properties of the conductive gel decline with time, which is difficult to meet the requirement of longterm EEG recording.To overcome this shortage, novel hydrogel electrodes for longterm stable EEG recording were designed [102][103][104].The downside of hydrogel electrodes is that hair washing is required after the use of hydrogel electrodes, which is not con ducive to the daily use of wearable EEGbased BCIs.Therefore, dry electrode research has gained more and more attention.Combshaped active dry electrodes [105][106][107][108], annularshaped dry microneedle array electrodes [109], and dry microneedle electrodes with soft circuits [110] were proposed for EEG signal acquisition in the hairy scalp.Nevertheless, without the aid of conductive agents, the contact impedance between the dry elec trodes and hairy scalp was generally higher than that between the wet electrode and hairy scalp.The proposal of the semidry electrodes [111,112] has made a compromise between wearing comfort and contact impedance.
For nonhairbearing areas like the scalp without hair, fore head, and behind/intheear, there are more kinds of electrodes to choose from.Increasing the number of electrodes is beneficial to obtain a richer EEG signal, so G.tec Company designed a highdensity pangolin wet electrode system, g.PANGOLIN, and up to 1024 EEG channels were recorded on the head to achieve highresolution signal acquisition.For improving the quality of signal acquisition, various flexible electrodes have also been applied for wearable EEGbased BCIs [113][114][115].In addition, flexible electrodes can also reduce motion artifacts and provide better wearing comfort.For wearable BCIs based on earEEG, the morphology of the ear canal and auricle should also be con sidered for electrode designing [116,117].
Artifact is another key factor affecting EEG signal quality.For wearable EEGbased BCIs, random noise and unexpected signal artifacts, such as muscle artifacts, ocular artifacts, cardiac artifacts, and extrinsic artifacts, will affect EEG signal acquisition during daily activities [118].The main artifact elimination meth ods can be divided into 2 categories.One approach is to opti mizing the EEG electrodes to realize low electrodeskin contact impedance, which can apparently reduce the artifact coupling.The other way is setting the electrocardiogram (ECG), electrooc ulogram (EOG), and electromyogram (EMG) leads to record the 3 main sources of the artifacts synchronously with EEG and then remove the artifacts in EEG signals with the algorithm for artifact elimination.Different artifact removal algorithms have been proposed to minimize the influence of artifacts in recent years.For instance, Lee et al. [119] proposed a method called constrained independent component analysis with online learn ing (cIOL) to search and reject the movement artifacts in EEG signals.Chang et al. [120] optimized the parameters of the arti fact subspace reconstruction (ASR) artifacts removal approach, demonstrating that ASR successfully removed ocular and muscle artifact components.Egambaram et al. [121] proposed unsu pervised eye blink artifact detection algorithms and achieved an over 90% average artifact removal accuracy.
The complicated lead connection increases the installation complexity and system weight, which brings inconvenience to the wearer's daily activities.In order to improve the wearability of BCI devices, it is important to realize low power consump tion, high bandwidth, and highprecision wireless data trans mission.At present, because of the mature technology, stable performance, and low cost, Bluetooth and WiFi modules are commonly used wireless signal transmission methods.Recently, novel wireless transmission methods have been proposed.For instance, Qiu et al. [122] designed a lowpower data acquisi tion frontend solution based on the field programmable gate array (FPGA), and realized scalpEEG signal realtime wireless transmission.
In addition to the use of wireless transmission, improving the flexibility, minimization, and lightweight of the wearable BCI system is also conducive to increasing the wearing comfort of users.With the development of technology, multisource acquisition and multifunction integrated wearable BCI sys tems have become a new development trend [123].However, with the use of wireless transmission technology and the wide spread application of wearable BCIs, a large amount of user information has been collected, and data security needs to be paid more attention to.

Conclusion
EEG acquisition is an important component of wearable BCI systems.In this paper, novel wearable EEGbased BCIs designed for medical applications are reviewed from the perspective of EEG signal recording.According to the recording electrode position, wearable EEGbased BCIs are divided into 3 categories: scalpEEGbased BCI, foreheadEEGbased BCI, and earEEG based BCI.Table 1 summarizes the wearable EEGbased BCI devices reviewed in this paper and their related applications in the medical field.Table 2 lists the work of wearable EEGbased BCIs for equipment control and the corresponding accuracies.
As can be seen from Tables 1 and 2, the wearable BCIs based on scalpEEG and foreheadEEG are relatively mature and have been successfully commercialized, while the earEEGbased wearable BCI has not been sufficiently applied in the medical field yet.
The characteristics of the 3 types of EEG are discussed in the corresponding sections, and each type of EEG has its own strengths and weaknesses for wearable BCI applications.Scalp EEG contains relatively rich brain information compared with the foreheadEEG and earEEG.As the mainstream wearable BCI solution, both hardware and software of scalpEEGbased BCI technology are the most mature.Nevertheless, the scalp EEG acquisition equipment is complicated to wear and signal recording can be easily interfered with hair.Both foreheadEEG and earEEG are recorded from hairless or less hairy areas, which reduces for the inconvenience of wearing to some extent.However, foreheadEEG is easily affected by artifacts, and hard  [95,96] Disease management [97,98] Kaongoen et al. [101] Custom-made headphone Ear-EEG Semi-dry electrode 8 Equipment control [101] to record the brain signals produced by the regions far from the frontal head, while the earEEG is hard to accurately extract and classify the characteristic signal because of the low intensity and high signal overlap.With the gradual improvement of the earEEG signal mapping model and the data processing algo rithm, wearable earEEGbased BCI is expected to become the mainstream technology in the next generation.
In order to meet the needs of daily use, wearable EEGbased BCI still needs to be further improved to meet the requirements of highquality brain signal acquisition, efficient and stable wireless data transmission, longterm wearing comfort, and high control accuracy.In the foreseeable future, owing to the advanced electrode technology, the optimized montage con figuration, and the powerful algorithm, wearable EEGbased BCI devices and systems will continue to develop in the form of simplicity, comfort, high performance, and automationreasons to believe that wearable EEGbased BCIs will bring an innovation of medical technology, fully employed in disease prediction, diagnosis, treatment, auxiliary equipment control, and other broad medical applications.[45] Scalp-EEG Emotiv EPOC Wheelchair / Zgallai et al. [46] Scalp-EEG Emotiv EPOC Wheelchair 70% (raw EEG data) 96% (spectrum of EEG) Shahin et al. [47] Scalp-EEG Emotiv EPOC Wheelchair / Bousseta et al. [48] Scalp-EEG Emotiv EPOC Robotic arm 85.45% Zhang et al. [54] Scalp-EEG DSI-24 Robotic arm 95.5% Lim and Quan [60] Scalp-EEG Ultracortex "Mark IV" Robotic arm 92% Saragih et al. [61] Scalp

Fig. 1 .
Fig. 1.Categories of novel wearable EEG-based BCIs and the primary medical applications.

Table 1 .
Summary of wearable EEG devices and corresponding BCI applications in medical field

Table 2 .
Summary of applications of wearable EEG-based BCIs in equipment control