Prior to engaging in communication, a preliminary step for application of BCIs to ICU patients involves detection of wakefulness. In ICUs, severe brain injuries often pose a challenge to assessing consciousness and predicting long-term recovery. Measuring consciousness is complicated by its subjective nature and its difficulty to quantify. Several studies have been conducted to assess the consciousness of patients using EEG-based BCIs.

In previous studies, early restoration of consciousness was deemed a pivotal factor in forecasting long-term improvement in functionality [20]. Some patients with brain damage may retain covert consciousness, potentially possessing higher levels of consciousness than what is assessed through behavioral evaluation. Claassen et al. [21] reported that brain activity in response to verbal motor commands, inferred from power spectral density analysis of the EEG, was detectable in 15% of a group of unresponsive patients, indicating a certain level of cognitive function despite apparent unresponsiveness. This finding suggests that adoption of EEG-based BCIs combined with machine learning in ICU environments can enhance the identification of cognitive motor dissociation patients who have potential to recover. Thus, BCI can guide more tailored treatment approaches and prognosis estimation [22]. Edlow et al. [20] advocated for implementation of transcranial magnetic stimulation electroencephalography in ICUs as a revolutionary approach to assess consciousness in severely brain-injured patients. That approach overcomes the limitations of measuring only EEG signals and more effectively guides ethical decision-making in patient care.

Despite the development of neurological assessments to provide a more comprehensive understanding of a patient's condition, reliable assessment of consciousness in ICUs remains an ongoing challenge owing to the need for further validation. Fine-tuned EEG-based BCIs with abundant data, along with advanced machine-learning algorithms, may provide a useful means for consciousness assessment in ICUs.

For direct communication using BCIs, various features and paradigms have been proposed. EEG-based BCIs can be categorized into two broad types: synchronous and asynchronous. Synchronous BCI, also known as the exogenous type, consists of a cue-based paradigm. This type of BCI allows patients to select one of many available messages for communication. Since it relies on EEG responses, such as steady state visually evoked potential (SSVEP) and event-related potential (ERP) triggered by a sensory cue, the cue stimulus is crucial in the system design. However, the method has the limitation of being passive. In contrast, asynchronous BCI enables self-paced control without a cue, utilizing specific brain wave patterns that reflect the user's motor intentions or motor imagery. Additionally, game-type tasks are commonly employed to detect EEG features, encouraging active user engagement in the task. Further details on the EEG features and tasks used in previous studies will be introduced below.

SSVEP occurs when flickering visual stimuli are presented, showing peaks at the frequencies of the stimuli and its harmonics (Figure 2). The physiological mechanism behind SSVEP is not fully understood; however, its amplitude is related to increased synaptic activity. SSVEP-based BCIs have demonstrated high accuracy in detecting a user's intended visual targets. Accuracy rates of 90% or higher have been reported in many studies with healthy volunteers [23-25]. However, the actual accuracy can vary depending on the specific experimental setup, the number of targets to choose from, and the subject's familiarity and fatigue with the BCI system. In a recent study, SSVEP-based BCI was used by two volunteers with brain injuries, achieving a mean accuracy of 57% and 65% for each. It was additionally reported that patients preferred employing a speller tool with SSVEP rather than eye tracking. However, visual fatigue was a significant factor affecting performance using SSVEP [26]. To overcome this issue, recent investigations have experimented with innovative visual stimuli, such as quick response (QR) codes, which have outperformed traditional checkerboard patterns [27]. Furthermore, researchers have adapted SSVEP-based communication to the unique constraints of the ICU environment [23].

The P300 ERP component is a positive deflection in the EEG waveform that occurs approximately 300 ms after a stimulus (Figure 3A). A notable application is the P300 speller, a BCI system that highlights various columns and rows within an alphabet grid, inducing a response when the patient highlights the specific row or column containing the desired letter—called the row-column paradigm (RCP) (Figure 3B). Early research that introduced the P300 speller with RCP demonstrated a typing speed and peak accuracy of 12 bits/min and 95%, respectively [28]. This type of task engagement has been developed as a single display speller that can reduce the error rate up to 80% and cause less fatigue compared to RCP [29-31]. In a clinical study, Guy et al. [32] demonstrated that 65% of participants, including those with amyotrophic lateral sclerosis (ALS), accurately chose more than 95% of relayed symbols, achieving a typing speed of 3.6/min (without word prediction) and 5.04/min (with word prediction). Additionally, Miao et al. [33] confirmed that ERP-based BCIs are applicable for use by patients with ALS. Other studies attempted to achieve faster and more accurate BCIs by combining other EEG features, such as SSVEP [34,35]. Since the P300 component can be elicited by different modalities of stimuli, including visual, auditory, and tactile, BCI studies using multisensory stimuli [36,37] or multiple intelligent techniques such as 3D interfaces with virtual reality [38,39] have been demonstrated. P300 can be employed not only to send messages, but also to verify consciousness. Li et al. [40] suggested that the P300 can reflect the conscious state of an unresponsive patient, highlighting its usefulness across a broad spectrum of applications.

Brain activity that reflects motor imagery is another commonly used method to detect the intentions of patients who have limited mobility. It involves an EEG signal that can be independently generated by the subject without the need for external stimuli [41]. In practical settings, asynchronous control enhances the user experience and system autonomy by allowing control that is self-initiated. Therefore, it marks a significant trend in the advancement of BCI studies [42]. Liu et al. [43] suggested the self-paced BCI system, which integrates four types of daily assistance tasks in the hospital environment: medical calls, service calls, appliance control, and catering services. The online experiment conducted in that research highlighted the system's efficiency, with a rapid response time of 3.38 seconds, an accuracy rate of 89.2%, and utility in enhancing information processing, easing cognitive load and aiding daily activities for individuals with significant motor restrictions. Motor-imagery-related brain waves can also provide effective feedback by mirroring imagined actions, making it suitable for rehabilitation purposes. Furthermore, numerous studies have focused on rehabilitation BCI systems that utilize brain waves corresponding to motor imagery and intention [13,45,46]. In this regard, motor-imagery-based BCI can also be implemented in ICUs to facilitate active communication and motor function recovery, especially for ICU-AW patients. To date, several studies have attempted to employ motor-imagery-based BCIs not only for consciousness assessment, but also as a rehabilitation tool and communication method in hospitals for acute brain-injury patients, including locked-in-syndrome patients [43,47,48]. The specific EEG features for these applications are detailed in the subsequent section.

The detection of motor intentions plays a pivotal role in rehabilitation and responsive neural prosthetics. This process hinges on two key neural elements derived from EEG data: MRCP and SMR. MRCP represents temporal changes in brainwave potentials within the motor cortex when motor intentions are initiated [56,57] . Observable in the time domain, this waveform is characterized by a negative deflection starting at approximately 1.5 seconds prior to the onset of movement [58,59]. Its prevalence is noted primarily in the sensorimotor area. Many studies have reported that MRCP negativity peak amplitude increases with repetitive movements or increased weights owing to peripheral fatigue [60-62]. Furthermore, several studies have indicated a positive correlation between the energy required for movement and the amplitude of the MRCP negativity peak. For instance, stroke patients, who typically exert greater effort to initiate movement, exhibit a larger amplitude in their MRCP negativity peaks [63]. In the context of motor learning, an increase in MRCP amplitude has been observed following motor task training, which may reflect higher cortical loads required to complete the task. Conversely, as motor skills improve through learning, a corresponding reduction is evident in the amplitudes of these peaks [64,65]. This correlation may be observed in patients with ICU-AW or motor impairments who must exert greater energy for movement compared to healthy individuals. The distinction of MRCP in patients not only establishes it as a potential indicator, but also emphasizes the need for careful consideration of these traits to engender critical insights for patient-specific approaches in rehabilitation [66].

Conversely, SMR embodies rhythmic oscillations predominantly in the alpha (8–12 Hz) and beta (13–30 Hz) frequency bands, originating from the motor and somatosensory cortices [67-69]. The alpha-band SMR, often referred to as the Rolandic mu rhythm [70], varies in amplitude in response to movement or the motor imagery of specific limbs. A power decrement occurs just before the movement begins, which is known as event-related desynchronization (ERD). A subsequent power increment occurs post-movement, which is termed event-related synchronization (ERS) [71,72]. Individuals with motor cortex impairment have shown abnormal SMR patterns, such as reduced mu ERD or delayed ERD. Höhne et al. [48] developed a motor-imagery-based BCI system for a locked-in syndrome patient, utilizing oscillatory features, including ERD and ERS, as a class-discriminant feature. They consequently achieved successful communication through left-hand and right-hand motor imagery despite challenges with artifacts and involuntary movements. This study validated the effectiveness of a customized motor-imagery-based BCI system for severely impaired individuals, significantly improving communication capabilities and potentially enhancing their quality of life.

Current BCI systems that capture motor intentions usually focus on movement initiation, whereas the intention of stopping movement has been rarely studied with respect to BCIs. Movement cancellation, or stop, entails the capability of ceasing ongoing or imminent movements as required. This function assumes critical importance in rehabilitation that involves external devices owing to its implications for safety and sophisticated feedback [73-76]. Instances may occur in rehabilitation tasks or everyday activities in which the subject may need to suddenly discontinue their actions. Unintended movements cause a risk of injury, potentially resulting from muscular overexertion. Furthermore, in the context of neuroplasticity, incorporating refined "stop" feedback may enhance recovery processes, presenting an advantage over conventional rehabilitation BCI systems that solely emphasize initiation. Therefore, detection of movement cancellation emerges as a significant consideration in the field of rehabilitation [74]. However, most motor-imagery-based BCIs implement a stop function that relies on a rest state rather than an authentic stop intention. Alternatively, some systems use a combination of specific motor-imagery commands, such as initiating a left-hand command immediately after a right-hand one to signal a stop command [77,78]. Nevertheless, more intuitive methods should be investigated.