The aim of this research was to better understand the effect of different BCI paradigms on self-reported fatigue and EEG biomarkers in children. In all conditions, there was an increase in self-reported fatigue post-task compared to pre-task. The difference was small, roughly a 1.5 pt difference on a 10 pt scale. There was also a small increase in alpha band power in all conditions. Contrary to our hypothesis, it was not clear whether the increase in self-reported fatigue or RS EEG alpha band power related to BCI tasks was greater as compared to watching a video or if changes differed between P300 and MI tasks. There was also no correlation between the change in self-reported fatigue and alpha band power as described in adults. We demonstrated that short periods of both MI and P300 BCI operation increase self-reported fatigue in children, but our results do not support the use of EEG generic alpha band power as a sensitive biomarker for fatigue.
The increase in VASF during the P300 session was greater than the increase during the video session. There were also significant increases in VASF from pre- to post-task in the MI and P300 sessions but not the video session (Fig. 3). It is possible that failure to clearly detect an effect for our main research question is due to this study being underpowered for this modest effect size. Prior adult BCI studies looking at self-reported fatigue during BCI use have found increases of above 3 pts on a 10 pt scale [15, 39], and between 1–3 pt on a 20 pt scale [16]. The standard error of measurement of the VASF scale for our data was 1.1 pts. Calculated with this error, the minimum detectable change is predicted at 3.2 pts on the 10 pt scale, much larger than what we observed. Our hypothesis regarding self-reported fatigue for this study was built off unpublished work in our lab (Kelly et al. in progress). This work demonstrated variable fatigue with multiple BCI paradigms and MI-BCI was more fatiguing than P300-BCI with a roughly 1.5 pt VASF increase. In contrast we found that P300-BCI sessions had the greatest increase in VASF, P300-BCI was not associated with any increase in VASF in this unpublished work (Kelly et al. in progress).
This previous pediatric BCI study was a different experimental set up than the present study, and any number of differences may have impacted user experience and fatigue. P300 and MI tasks were completed on the same day in a randomized order and tasks were 8–20 minutes in length. The training, the task, and the BCI hardware/software were also not consistent between studies. In the prior study, the age of participants was slightly higher with a mean age of 11.3. The primary research question was about BCI control in children, with self-reported fatigue as a secondary outcome. Even though the task length was longer in the current study, results from self-reported fatigue indicate P300 was fatiguing even at 10 and 15 minutes. Similar numbers of participants had reported an increase by one or more points on the VASF in the P300-task and the MI-task by these time points (10 minutes: 8 participants in P300, 6 in MI; 15 minutes: 13 in P300, 12 in MI). Particularly in the P300 session, however, younger participants had a VASF increase of over two points while older participants did not even have a one-point increase at this time. The slightly older participants in the prior study may have impacted the difference in P300 observed in these studies. The VASF score change observed across time in all sessions was greater for the younger children in this study. The older children and adolescents were better able to handle 30-mintues of a BCI task without significant fatigue and generally did not find the video session fatiguing (supplemental Fig. 1, Additional file 1). It is also likely that having the tasks in the same study session had a different impact on self-reported fatigue than having the tasks on separate days. Additional pediatric BCI fatigue studies considering these many variables are required.
Day-to-day chronic fatigue was assessed using the PedsQLTM-MFS before each session in the present study. As expected, children with lower MFS scores (ie. higher fatigue) reported higher VASF values pre- and post-task. The impact of more chronic and pathological fatigue will be an important consideration for understanding BCI fatigue in children with CP, who typically have higher chronic fatigue [19]. Further pediatric work is needed to draw conclusions on the impact of more significant fatigue at baseline and its impact on BCI performance and behavioural changes. We suggest that simple, validated measures like the PedsQLTM-MFS, for which CP-specific versions are also available, be employed prospectively in BCI studies in such clinical populations to better characterize the role of fatigue in performance and other outcomes.
To our knowledge, this was the first trial to investigate EEG biomarkers of fatigue in children during BCI use. While adult studies show promise in tracking these biomarkers and associating them with self-reported fatigue, more work may be needed in children to refine these measures. During analysis some electrodes were excluded due to signal artifacts or noise. This was either from sub-optimal connection at known specified electrode sites or from participants touching an electrode or moving the headset forehead strap. Identification of individualized alpha bands was unsuccessful with existing python algorithms for use with eyes-open RS. Pipelines were applied but resulted in non-physiological interparticipant and intraparticipant variability. Comparison of alpha band power results from the present work to past literature was also difficult due to variable units, use of undefined units, lack of values given, and use of both relative and absolute metrics. From our data, the standard error of measurement of this band power was 4.8 µV2/Hz. Calculated with this error, the minimum detectable change is predicted at 13.3 µV2/Hz. A change near this magnitude was not detected. Change in alpha band power magnitude in many prior adult studies is significantly larger and often correlates with self-reported fatigue [15, 16, 23]. Prior EEG studies have found that changes in EEG are not apparent until an individual is highly fatigued [40] and perhaps our interventions were not long or hard enough or lacked appropriate difficulty to reach such a threshold. Age-dependent developmental differences in EEG neurophysiology may also have affected our ability to detect fatigue-related changes.
Despite a broad acceptance of EEG band metrics as fatigue biomarkers, there is still inconsistency in the literature. Particularly in research looking at driver fatigue, several studies have found no change in alpha band power or even a decrease in alpha band power [41]. As an alternative metric, EEG entropy has been calculated as a biomarker for fatigue including during sustained attention cognitive tasks [42, 43] and high stress cognitive tasks [44]. In a driving simulator study, changes in sample entropy were more consistent during fatigued states than band metrics [45]. Peng and Colleagues also found that a multiscale entropy metric better distinguish fatigue states than the classic bands during a steady-state visually evoked potential BCI task [41]. Entropy may therefore be a useful measure in children to help avoid difficult individual frequency band calculations or use of generalized bands which may be less accurate [46–49]. Evaluation and comparison of pre- and post-task RS is complex as post-task RS brain activity can be modulated by other elements of the prior task outside of just fatigue [50]. A combination of band metrics, from the frequency domain, and entropy, from the time or the time-frequency domain [51], may be more useful for a deeper understanding of these complexities [50]. Such dual metric analysis should be investigated in children using BCIs.
Performance varied in both the P300 game and the MI game from 100% or nearly 100% to essentially no control at all considering chance accuracies for each paradigm (< 11% in P300; <50% in MI). Interestingly, those who performed better tended to report feeling that they did worse than those who performed poorly although this was purely correlational. Performance did not decline across the BCI sessions. The larger their change in fatigue, the lower participants rated their session enjoyment. Correlational analysis also revealed that children who felt the tasks were more mentally, physically, and temporally demanding also reported being more frustrated. A link between frustration and workload was also noted in end users with motor impairments during BCI gaming [52]. This study had a small sample of four individuals, but interestingly, this association existed for those who had less severe motor impairment [52]. Due to potential EEG changes with frustration, it has been suggested that EEG signal, BCI classification, and performance may be impacted [53]. However, a previous study looking at the performance-frustration relationship did not report significant influences of frustration on performance [54]. Anecdotal reports from an additional study with Amyotrophic Lateral Sclerosis patients also noted that frustration did not seem to impact motivation [55], but we found a negative correlation between frustration and motivation. Psychological factors such as frustration and motivation are clearly influenced by age and development and need to be considered carefully in pediatric populations.
Limitations
Our calculated but modest sample size is a potential limitation for this work. The present study was not adequately powered to draw conclusions about differences between sessions since changes in both primary outcomes were minimal. The broad age range (7–16), with inconsistent samples at each age, also presented additional challenges to age-base analysis that was overcome by splitting the sample into just an upper and a lower half by date of birth. Additional limitations include low headset tolerability that reduced the number of our participants completing the full study protocol. Time on task is accepted as a key predictor of fatigue for physical and cognitive tasks [56–59], and this inconsistency introduced more variability into the study. There is a relationship with fatigue and pain or discomfort and while self-reported pain and comfort ratings for the headset were controlled in the statistical models, discomfort is another potentially confounding variable.
Future studies should continue to evolve our understanding of BCI fatigue in children using larger samples, and/or smaller age ranges to overcome challenges of a highly variable population. Studies should also investigate longer durations of BCI use with a system that will be more broadly tolerated by children. It will be important to consider BCI experience and performance as not all children have good BCI control in one session with no prior training. An initial training session may be beneficial to ensure only those with adequate BCI control go forward and are involved with research questions surrounding fatigue. For those with lower BCI performance, future studies will also be needed to investigate predictors of performance as well as strategies to promote BCI learning. Finally, and most importantly, studies on typically developing children should guide the development of similar studies in children with CP, the intended end users of this technology.