4.1 Results Summary
Based on the work presented above, and very new data still being analyzed, we have four observations that are generally consistent with other recent work. First, in some users, the P300 BCI approaches relied largely on non-P300 activity. Second, the P300 BCI approaches were generally more effective than the MI approach. This, patients exhibited substantial within- and across- subject variability. Fourth, effective BCI performance was possible for persons with severe disabilities, but often required more data than with healthy users.
4.2 Future Directions
Because of the novelty of this research field and patient group, we see considerable opportunity for new research. Our highest priority is validation. We need to evaluate mindBEAGLE and emerging technologies with many more patients. The challenges with the MI approach can only be addressed through further research, along with several other future directions.
From an HCI perspective, DOC patients introduce several challenges that merit future study. Visual stimuli, which are vital in most interfaces, cannot provide reliable interaction. DOC patients have cognitive and/or attentional deficits that may create difficulty understanding the tasks, remembering instructions, maintaining attention to task demands, or focusing on the different cues. They may also have sensory deficits, such as difficulty hearing or feeling, which could render auditory and/or vibrotactile modalities ineffective. These challenges have long been recognized within the HCI community, such as within the design of assistive technologies (ATs) designed for persons with disabilities . However, DOC patients introduce another challenge that is unique within HCI. Many of these patients fade in and out of consciousness, with no a priori way to determine whether they are in an up or down state before beginning a session. Thus, we need to attempt assessment several times, since assessments that do not reveal indicators of consciousness may simply mean that we need to try again. Furthermore, we currently have no way to know how many assessments are appropriate before concluding that a patient could never communicate. Even if many prior assessments were unsuccessful, there is always a chance that the patient could pass an assessment if medical or research experts try one more time. This may create a heart wrenching dilemma: given the limited resources in a medical environment, when is it appropriate to give up with one patient and move on to the next one to possibly change his/her life?
This is a serious challenge, and we currently have no solution. Behaviorally, there is no known way to determine when patients may be sufficiently aware for a mindBEAGLE session. It may be possible to use EEG or other measures to identify indicators of awareness more quickly than we can right now, which is an important future direction we are now exploring. In some cases, administering medications may increase the chance that a patient will be aware prior to a session, although this solution introduces further ethical challenges.
Another unique challenge involves “BCI inefficiency”. This phenomenon, also called “BCI illiteracy”, means that a user is unable to use a particular BCI approach. A BCI used for communication should ideally provide communication for all users who need it. However, this is not always possible, just like conventional interfaces; some people cannot use keyboard or mice due to motor disabilities or other reasons. Moreover, given this unique patient group, it is not realistic to expect high literacy. Many patients who are diagnosed as unable to communicate really are unable to communicate. If they are not able to use a BCI for communication, this may not reflect a failure by the BCI designers. Rather, it may indicate that the BCI is performing as well as can be expected for a patient who is indeed unable to reliably produce EEG differences. Many patients diagnosed with DOC are indeed unable to understand instructions, maintain attention, develop and implement goal-directed behavior required to answer questions, etc. These patients are presumably unable to communicate with any known technology, and could not be helped without major advancements in medical technology. Nonetheless, we are exploring improved protocols and signal processing approaches that could improve BCI literacy. Additional data from DOC patients should help us identify relevant EPs and other EEG characteristics could lead to improved classification accuracy and literacy.
Furthermore, the MI approach usually requires training, and has been recognized within the BCI community as more prone to BCI inefficiency [13, 14]. While most healthy people can attain effective communication with the MI approach, MI BCIs are not able to detect reliably discriminable brain signals in a minority of users. The training requirements in most MI BCI approaches are more daunting for DOC patients for at least three reasons. First, training sessions are pointless if the patient cannot understand instructions and follow the task. With this target patient group, a P300 BCI may be needed to assess these mental capabilities. However, some healthy people can attain effective MI-based communication with fairly brief training , even with a limited electrode montage . Second, any session with this patient group is much less casual than with healthy persons. Third, it is unknown whether this target population is capable of producing clear differences between left and right hand MI that a BCI could detect. Research has shown that persons with late-stage ALS can produce MI signals adequate for BCI control ; therefore, the inability to produce certain movements, even for an extended time, does not necessarily prevent people from imagining movement in a way that an MI BCI could detect. Patients with DOC have different disabilities than late stage ALS patients, and thus this question remains open for further study.
So far, we have focused on EEG activity. Within the BCI community, there has been some attention to “hybrid” BCIs  that might combine EEG-based signals with other tools to image the brain (such as fMRI) or other biosignals (such as activity from the heart, eyes, or muscles). These methods may be less promising as tools to provide communication for persons diagnosed with DOC. Aside from high cost and low portability, fMRI imaging requires a very strong magnetic field that is difficult or impossible to use with the electronic and metal devices that DOC patients need. Deriving information from other biosignals can be helpful for diagnosis, but is not especially helpful when trying to provide communication to persons with little or no ability to voluntarily modulate these signals.
Like many BCI researchers, we are interested in improving software, including better HCI-based and user-centered software to interact with users. In addition to the unique challenges of working with DOC patients, we also need to present information to system operators in an engaging, informative manner. Figure 2 shows how system operators can view the EEG, brain maps, and ERPs in real-time. We need additional testing to determine whether this is indeed the most informative and effective way to present information to system operators. We also need to consider differences among system operators, who might be medical doctors, nurses, research professors, postdocs, graduate students, technicians, or other people with different backgrounds and skills. The most effective interface components may differ for different users based on their expertise.
Hardware development is another important future direction. Figures 1 and 2 show systems that used wired, active, gel-based electrodes embedded in a cap. A wireless system might be easier to use while eliminating the risk of snagging cables on equipment. Dry electrodes might provide adequate signal quality while reducing preparation and cleanup time. New electrode montages and mounting hardware could enable electrodes that do not require a cap, such as electrodes embedded in headphones, headbands, or other head-mounted devices.
In summary, the results obtained so far from our group and other groups have clearly shown that some patients diagnosed with DOC do exhibit indicators of consciousness with BCI technology and might be able to communicate, even though they do not exhibit indicators of consciousness on a behavioral level. On the other hand, the majority of DOC patients do not exhibit such indicators with BCIs. This is still a new research direction with many unanswered questions and very serious challenges, and we do not know which approaches will be most effective. Drawing on lessons from BCI research, the most effective approach will probably vary across different users, which underscores the importance of providing different approaches through different sensory modalities to increase the chance that at least one of them will be effective. We are excited about the prospect of helping some people within this unique patient population, and both hope and expect that future research will lead to more universal and effective solutions.