Competing at the Cybathlon Championship for Athletes With Disabilities: Long-Term Motor Imagery Brain-Computer Interface Training of a Tetraplegic Cybathlete

Attila Korik (  a.korik@ulster.ac.uk ) Ulster University https://orcid.org/0000-0002-3757-6092 Karl McCreadie Ulster University Niall McShane Ulster University Naomi Du Bois Ulster University Massoud Khodadadzadeh Ulster University Jacqui Stow National Rehabilitation Hospital Jacinta McElligott National Rehabilitation Hospital Áine Carroll National Rehabilitation Hospital Damien Coyle Ulster University


Background
The Cybathlon championship for athletes with disabilities is a unique competition in which people with physical disabilities compete against each other to complete tasks and challenges using state-of-the-art technical assistance systems. Serving as a platform for technology developers to exchange ideas and collaborate closely with people with physical disabilities as they develop their devices -Cybathlon aims to drive research on assistance systems for everyday use, and promote public dialogue [1].
The majority of teams involved in the brain-computer interface (BCI) race at the Cybathlon championship, focus on training pilots to modulate brain rhythms through motor imagery with a BCI to control a virtual the user's psychological state in uenced their ability to concentrate, which affected their performance negatively. Hence, variability in the signal between sessions is largely considered to be due to shifts in the user's psychological state -for example, due to fatigue or loss of attention [16], [17].
The current study provides details of the BCI setup, training regime, and performance obtained during long-term BCI training of the Ulster University's Cybathlon team (Team NeuroCONCISE) for two race competitions. The ndings highlight the challenges in developing a BCI that is capable of adapting to changing user states and environment conditions that can result in temporal variations in the neural signal and create a barrier to the use of BCI systems in everyday life for individuals with a motordisability.

Participant
The study involved a single tetraplegic participant (the pilot) with normal vision and hearing, aged 49 at the time of the Cybathlon 2020. The pilot suffered a spinal injury (factures in C4-C5) in 1993 during a motorbike accident. Prior to the commencement of the training, the pilot was presented with information regarding the experimental protocol and was asked to read and sign an informed consent form to participate in the study, which was approved by the National Rehabilitation Hospital of Ireland research ethics committee. Before the beginning of the BCI training carried out in 2019 and 2020 (reported in this paper), the pilot took part in 10 basic BCI training sessions in 2009 and 12 training sessions for Cybathlon 2016.

Experimental Paradigms
Long-term BCI training involved feedback in multiple sessions using three different BCI controlled games: NeuroSensi, Triad, and BrainDriver ( Figure 1).

NeuroSensi game training for paired motor imagery tasks
The rst phase of the BCI training in both years, 2019 and 2020, involved the NeuroSensi BCI game ( Figure 1A) which is played using two motor imagery commands. The 2-class motor imagery paradigm was designed to train participants. The NeuroSensi game has a representation of a neuronal axon on both sides of the monitor. Two seconds after the beginning of the trial, a light (representing a neural spike) appears at the far end of one of the two axons to cue the participant to begin the corresponding motor imagery task. The light takes 6 seconds to travel over the 'axon' during the task period ( Figure 1A). In each NeuroSensi session, six runs were performed wherein different binary combinations of the three commands (left hand (L), feet (F), right hand (R)), and relax (X) were performed. The number of trials in each run acquired during BCI calibration varied between 30-60 (equal number/class) depending on the actual session ID (more trials in the initial session, fewer trials in later sessions). The time duration of a run, therefore, varied between 240s and 480s. The time duration of six runs involving L vs. R (LR), FR, LF, LX, FX, and XR tasks, including ve 90s inter-run pauses, varied between 20 and 30 minutes ( Figure 1B). After all pairs of runs (i.e., LR, FR, LF, LX, FX, and XR) runs was re-structured and trials involving the same class recorded from different runs were pooled (e.g., for "L" class "L" trials were pooled from LR, LF, and LX runs). Thus, in the re-structured dataset four different classes (L, F, R, and X) were separated. The number of trials per class for a single session varied between 45 and 90.
The three 'task vs. task' classi ers (LR, FR, LF) were calibrated using the corresponding trials stored in the re-structured dataset. However, in runs when the NeuroSensi game was controlled with a 'task vs. relax' (TX) task, i.e., in runs where the character was controlled with LX, FX, or XR task pairs, the same TX decoder was used. The TX decoder was calibrated using T vs. X trials from the re-structured dataset where T trials were pooled as a union of L, F, and R trials. To improve the cross-session stability of the calibrated BCI, the nal dataset for BCI calibration was prepared by pooling re-structured datasets from multiple sessions acquired prior to the calibration.
Triad game for monitoring details of the multi-class classi cation The Triad game ( Figure 1D) provides real-time continuous visual feedback from each of the four 2-class classi ers. The analogue output of the three 'task vs. task' classi ers (LR, FR, LF) are presented using a light blue ball on the three edges of a triangle. Furthermore, the linear combination of the LR, FR, and LF classi er output is presented with an additional coloured ball indicating the composite output of these three 'task vs. task' classi ers. The colour of the composite output indicator ball is assigned via the analogue output of the 'task vs. relax' (TX) classi er. The colour of the ball indicates the command is decoded as the task (green) or relaxed (dark blue) condition.
BrainDriver game to familiarize the pilot for the race in the Cybathlon BCI event After the pilot learned to control the BCI using the NeuroSensi and Triad games (2020 only), the BrainDriver BCI racing game was used to practice the control of the avatar (a virtual race vehicle) ( Figure  1C). The actual track of the BrainDriver game comprised four different zones. There are zones with left and right curves and straight zones with streetlights turned on or off. To maintain a maximal speed of the vehicle, the pilot must produce the correct race command using the 4-class BCI (e.g., left or right arm motor imagery for left or rights turns, feet imagery "headlight" and relax for "no-control"). If an incorrect command is presented the vehicle is inhibited which presents obvious negative visual feedback to the pilot to enable learning and attempt correction strategies. The pilot was instructed to relax immediately after issuing a command to allow for 'no control', or as an alternative strategy to continue to maintain the motor imagery command. Section 2.5 describes how the controller limits commands and assists in dealing with variation in control performance by the BCI and pilot.

Data Acquisition
The EEG was recorded from 32 EEG channels using a g.Nautilus Research active electrode wireless EEG system [18]. The EEG reference electrode was positioned on the left earlobe. The EEG was high-pass ltered (>0.1Hz), notch ltered (48-52Hz), and sampled (A/D resolution: 24Bits, sampling rate: 250Hz).
The ground electrode was positioned over the AFz electrode location according to the international 10/20 EEG standard ( Figure 2). The communication between the real-time BCI decoder module run in Simulink [19] (used for EEG data acquisition and online signal processing) and each of the three (NeuroSensi, BrainDriver, and Triad) games was via a 'user datagram protocol' (UDP).

Calibration of the two-class classi cation modules
The BCI framework included a lter-bank common spatial patterns (FBCSP) [20] and mutual information (MI) based feature selection [21], a well-established framework used in BCI applications that enable discrimination between imagined movements [22] performed with the left hand (L), feet (F), right hand (R), and relax (X) conditions [23]. The FBCSP-MI module, the core of the online BCI framework (Figure 2A), was calibrated o ine as described below.
EEG signal processing The pre-processed EEG dataset was downsampled from 250Hz to 125Hz and band-pass ltered in four non-overlapped standard EEG bands (8-12Hz (mu), 12-18Hz (low beta), 18-28Hz (high beta), and 28-40Hz (low gamma)) using high-pass and low-pass nite impulse response (FIR) lter modules (band-pass attenuation 0 dB, band-stop attenuation 60 dB). Trial-relevant time intervals between -2s (before) and 8s (after) the onset of the 2s pause (i.e., -4s (before) and 6s (after) the onset of task) were epoched out from the frequency ltered EEG dataset for 21 pre-selected EEG channels, and stored for spatial ltering. The epoched data using a 1s to 2s width classi cation window enabled comparison of the decoding accuracy (DA) obtained in the 0 to 2s reference baseline interval (covering the pause period) and during the 2s to 6s task interval (after the pause period).

Spatial ltering
The common spatial patterns (CSP) method was used to create spatial lters that maximize the discriminability of two classes by learning spatial lters which maximize the variance of band-pass ltered EEG signals from one class, while minimizing their variance from the other classes [24]. The linear transformation matrix de ned by CSP converts the pre-processed EEG signals into a new vector space de ned by the CSP lters.

Feature extraction
The number of selected CSP lter pairs for each 2-class classi er for each frequency band was set to three. The time-varying log-variance of the CSP ltered EEG was calculated using a 1s width sliding window, with a 40ms time lag between two windows. Thus, the offset (end-point) of the 1s sliding window was set to cover the time interval between -1s (before) and 8s (after) the onset of the pause (covering a 1s sliding width window, the 2s pause, and 6s task intervals).

Feature selection
The mutual information (MI) between features and associated target class using a quantized feature space was estimated [21] to identify a subset of features that maximize classi cation accuracy.

Two-class classi cation
A regularized linear discriminant analysis (RLDA) algorithm from the RCSP toolbox [24] was applied to classify the extracted features. Linear discriminant analysis (LDA) uses a class separator boundary in a linear hyperplane to separate data into two classes. The time-varying analogue output of the classi er, i.e., the time-varying signed distance (TSD), is the time-varying distance between the location of the classi er output and the class separation boundary in the LDA hyperplane. The class assigned to each feature vector depends on the polarity of the classi er output, determined by the relation between the location of the feature vector and the class separator boundary in the hyperplane [25]. The current TSD value is calculated using an LDA classi er as described in (1) where, is the features vector at time t in the th trial, while and are slope of the features and bias of the discriminant hyperplane. Thus, (the time-varying signed distance, obtained at the output of the classi er) is composed of a part (i.e., the feature vector dependent LDA components) and an constant bias.
The time-varying DA was calculated and compared for each of the four 2-class classi ers using the best 6, 10, 14, and 18 MI ranked features, using six-fold cross-validation. The features that provided a classi er con guration with the highest DA peak in event-related period of the task (in a 2.4 to 8s interval of the trial, covering a 0.4 to 6s interval from the onset of the task) was applied to the online BCI con guration, in the case of each 2-class classi er, separately.

Topographical analysis
To identify frequency bands and cortical areas that provide the highest contribution to the peak DA, an analysis was performed using parameters of the calibrated CSP lters and the MI weights for each of the four 2-class classi ers, separately. For the time-varying frequency analysis, the mean values of MI weights were calculated in each analyzed frequency band, and time point, separately. The obtained results were plotted in the form of subject-speci c heat maps, indicating the time-varying DA contribution of the frequency bands analyzed. The location of the source activity was plotted using the 'standardized low resolution brain electromagnetic tomography' (sLoreta) software package [26] for each 2-class classi er in each frequency band, separately, indicating cortical areas where features provided the highest contribution for calculating maximal DA.
Combining trials for different runs and sessions The objective was to nd an online BCI con guration that provides the highest DA with a high level of stability over sessions. Thus, the BCI was calibrated using different datasets that were pooled from different combinations of existing sessions. A cross-session DA analysis was performed for each BCI con guration wherein the time-varying DA plots were compared using datasets excluded from calibration data. The BCI con guration was selected for the oncoming sessions based on a manual comparison of the cross-session time-varying DA plots, frequency maps, and topographical maps obtained based on the analyzed BCI con gurations using the objectives described above (i.e., long term stability paired with a maximal level of DA).

The online BCI
The core module of the online BCI involved the same FBCSP-MI-based 2-class classi cation framework ( Figure 2A) for the NeuroSensi, Triad, and BrainDriver games which are described in Section 2.4. However, the post-processing module for these three games was different for the three BCI game applications.

NeuroSensi game
The NeuroSensi game uses only one of the four binary classi ers for controlling the character (i.e., LR, FR, LF, or TX). The baseline of the corresponding TSD signal was calibrated manually set to zero at the beginning of each run using an offset value. The amplitude of the TSD signal was corrected to a value that enabled the controlled character to move over the controllable area during the game using a scaling factor. The corresponding TSD after the baseline correction was downsampled to 25Hz and sent by UDP to the NeuroSensi game.

Triad game
As the Triad game was controlled by the TSD output of each of the four 2-class classi ers, the baseline of each of the four TSD signals was corrected, separately, as described above for the NeuroSensi game. The TSD value after baseline correction was downsampled to 25Hz and sent by UDP to the Triad game.

BrainDriver game
The online BCI framework ( Figure 2), in addition to the FBCSP-MI and TSD baseline correction modules (discussed above and presented in Figure 2A), involves a control command decoder module composed of a multi-class decoder, a stability delay timer, a dead-band (DB) control module ( Figure 2C) and game control command translator module ( Figure 2D) followed by a UDP unit for sending commands to the BrainDriver game ( Figure 2E).
The output of the multi-class decoder relies on the baseline-corrected outputs of the four binary (LR, FR, LF, and TX) classi ers. If the output polarity of two of the three 'task vs. task' classi ers (LR, FR, LF) are not con icting and the TX classi er is indicting a task condition ("T") (i.e., the pilot is not relaxed), the label of the decoded task is forwarded to the next module for a stability check. For example, in Figure 2C both LR and LF classi ers output indicate the same ("L") result and the TX classi er indicates that there is an ongoing task ("T"). Therefore, in this example, the decoded ("L") command passes through on Command Control Gate 1.
To lter out transient responses, the decoded ("L") command passes through Command Control Gate 2 only if the decoded ("L") command is maintained in the same condition for a prede ned (300ms) period. If this stability check is matched, the decoded ("L") command is translated with the game control command translator module to the game control command as shown in Figure 2D. Finally, the game control command is sent by UDP to the BrainDriver game. An example of a track section and corresponding control commands are illustrated in Figure 2E (details of the BrainDriver game in Section 2.2).
To provide the opportunity for the pilot to reach a relaxed condition before the next command is decoded and to ensure sudden changes in classi er conditions don't interrupt a correct command issued to the vehicle, a dead-band system is used. Once a command is sent to BrainDriver game, the dead-band timer module is activated, resulting in blocking new commands on the UDP for a prede ned dead-band period (e.g., in Figure 2C the dead-band period is 6s). To enable the pilot to correct an incorrect command, a dead-band-break (DBB) unit is also employed. If the pilot maintained a non-relaxed condition after a dead-band-break time is counted from the previous command (in Figure 2C the dead-band-break is 3s) the dead-band-break unit overrides the dead-band and allows the new command to be sent. To nd an optimal con guration that supports the pilot's control ability maximally, the actual value of the dead-band and dead-band-break parameters were adjusted manually over sessions and runs during the training period. The length of the dead-band was selected in a range between 20s and 8s, and the dead-bandbreak was selected in a range between 0s and 4s.

Results
Using datasets acquired in 2019 and 2020, an o ine analysis was performed to determine which modi cations of the current BCI framework and calibration methods improved the pilot's BCI control accuracy.

Calibration details
To improve the cross-session stability of the online BCI, the BCI calibration was performed using a dataset acquired from multiple sessions prior to the calibration date. To nd an optimal combination of the sessions from which the acquired data was added to the BCI calibration dataset, the BCI was calibrated and tested multiple times using data acquired from a combination of different sessions. A dataset involving two to four sessions providing the highest cross-session DA over sessions near to the calibration date were noted and used for the BCI re-calibration. Sessions used in the cross-session analysis, along with the sessions used for BCI calibration, and the sessions in which the re-calibrated BCI was used are presented in Table 1.   Figure 3B and Figure 3D, respectively. Furthermore, a comparison of the time-varying DA obtained from the same binary classi er in 2019 vs. 2020 years is presented in Figure 3E. The time-varying DA graphs show that DA during the 0-2s pause period is approximately at the chance level (50±10% (mean ± STD)). After the onset of the motor imagery task (the dotted vertical line at 2s), the online DA reached 90±10% (mean ± STD) ( Figure 3E) and was maintained by most classi ers between a period of 4s to 8s (i.e., from 2 s after the onset of the task until and of the task period). A comparison of the time-varying TSD values obtained for the two classes is presented in Figure 3F and Figure 3G using the dataset recorded in online sessions in 2019 and 2020, respectively. The graphs show that during the pause period (i.e., before the onset of the task) the TSD varied in the same range for both classes around the zero baselines and the TSD separability for the two classes is a maximal around 4s (i.e., 2s after the onset of the task).

Frequency analysis and topographical results
A comparison of frequency bands and cortical areas providing the highest DA contribution in the four binary classi ers (LR, FR, LF, TX) involved in the BCI con guration applied to the Cybathlon challenge in 2019 and 2020 is presented in Figure 4.
The results of the frequency analysis using CSP and MI weights of the BCI that was applied 2019 indicate an increased level of motor imagery task-related brain activity for each of the three 'task vs. task' classi ers (LR, FR, LF) in the 18-28 Hz (high beta) band ( Figure 4B). The highest contribution for the 'task vs. relax' (TX) classi cation, similar to the LF, FR, and LF classi cation, was obtained in the high beta band. However, regarding the separation of the task and relaxation conditions, in addition to the 18-28Hz (high beta), the 12-18Hz (low beta) and 28-40Hz (low gamma) bands also contribute to high accuracy. 3.4 BrainDriver scores, baseline correction, and dead-band con guration Each year, after practice sessions using the NeuroSensi game, our pilot's training focused on improving his BCI control ability using the BrainDriver game. The game completion time achieved by our pilot in 2019 and 2020 is presented in Figure 5.
In both years, during the training period, the race completion time decreased over sessions. However, around four days before the competition, each year, the required time for nishing the BrainDriver race increased (highlighted with the dotted oval in Figure 5A and Figure 5B). Furthermore, on the day(s) of the competition, the required time to nish the BrainDriver race increased, resulting in substandard performance in the nal competition race compared to earlier training performance (e.g., up to session 16 in 2019 and up to session 21 in 2020).
The baseline shift on the TSD of the 2-class classi ers ( Figure 5C and Figure 5D) is correlated with the increase in race times ( Figure 5A and Figure 5B) as highlighted with a dotted and solid oval.
To nd a BCI con guration that supports maximally the pilot's control ability, the values of the dead-band and dead-band-brake parameters were set over sessions and runs in a range of possible options. The applied values of the dead-band and dead-band-brake parameters are presented in Figure 5E and Figure   5F. Based on manual observation of the game completion times and feedback information reported by the pilot about which con guration that best supported control ability, the BCI was con gured for the nal challenge in both years, 2019 and 2020, using a 6s dead-band and a 3 s dead-band-break parameter value.
3.5 BrainDriver race performance and EEG power spectral density (PSD) An analysis was performed to evaluate a possible connection between the change in BrainDriver race times and power spectral density (PSD) of EEG over sessions and runs during BrainDriver game control. The logarithmic magnitude of PSD values was obtained in a frequency  range applied to the BCI in 2 Hz steps over each BrainDriver race, separately. As the EEG dataset on race day in 2020 was not archived, this analysis was performed only for the dataset acquired in 2019. 3.6 BrainDriver race times and frequency bands providing the highest DA contribution Some sessions before the competition races, a decrease occurred in the pilot's BrainDriver game control ability ( Figure 5A and Figure 5B). Based on the NeuroSensi dataset acquired in 2020, an analysis was performed to investigate a possible connection between the decreases in the BCI control ability and the change in frequency bands providing the highest contribution to 2-class classi cation, comparing results from sessions 19-21 and sessions 22-24.
The results of the analysis (Figure 7) for each 'task vs. task' classi ers (LR, FR, LF) shows that the taskrelated DA contribution of the 18-28Hz (high beta) and 28-40Hz (low gamma) bands show a relatively high value (red-coloured area) for sessions 19-21 ( Figure 7A) closer to the onset of the task (dotted vertical line) compared to results obtained for sessions 22-24 ( Figure 7B). As indicated by the light colour areas in Figure 7C, the highest difference in DA contribution for 'task vs. task' 2-class classi cation between the two session groups was obtained during the task period around 4-5s (i.e., 2-3s after the onset of the motor imagery task).
In the case of the 'task vs. relax' (TX) classi cation, the DA contribution map obtained from the twosession groups showed more diverse patterns, especially in the 8-12Hz (mu) band. However, areas with the highest contribution for sessions 19-21 and 22-24 are consistent.

The reliability of the features in terms of the ability to separate classes
We investigated the feature separability encoded by the four binary classi ers using the dataset from 2020 when the pilot trained with the NeuroSensi game. The session varying values of the LDA components for the four 2-class classi ers are presented for class1 and class 2 in Figure 8A and Figure 8B, respectively. Figure 8C shows the difference of the LDA components obtained for class1 and class2. Figure 8C indicates clearly that the largest difference in LDA components calculated for the two classes is obtained (for each 2-class classi er) at the highest mutual information ranked feature (indicated with rank1 at the vertical axes in Figure 8C). The gure also demonstrates that some features (e.g., for LR classi er: features 4, 5, 8) provide more negative than the positive contribution to the classi cation.
The change in each LDA component using a baseline corrected LDA components (i.e., session 10) is presented for class 1 and class 2 in Figure 8D and Figure 8E, respectively. Figure 8D and Figure 8E indicate, that for the three 'task vs. task' (LR, FR, LF) classi ers, the LDA components (e.g., for LR classi er: features 1, 2, 7, 8) are signi cantly (p < 0.05) different in sessions 19-21 compared to earlier sessions. The session numbers where the change occurred match the sessions when the pilot's BrainDriver control ability was at its best (see BrainDriver race completion times 2020 in Figure 5B between sessions 19 and 21). This change in the level of the features over sessions in the case of the TX classi cation is lower compared to that obtained for the 'task vs. task' classi ers.

Discussion
This paper provides an overview of a long-term pilot training and BCI strategy implemented in preparation for the Cybathlon BCI race event in 2019 and 2020. In both years, the initial phase of the project focused on calibrating the 2-class classi ers using the NeuroSensi game. Our results show that decoding accuracy increased during the online sessions following the calibration period. This observation is in line with similar studies that have demonstrated improvements in the Cybathlon pilot's BCI control ability due to long-term practice periods -although this has not always translated into high performance on the day of the event, as was the case with the current study and for both [14] and [15]. Possible reasons for this drop in race-day performance, and suggested strategies to address the issue are discussed.
The TSD results from the current study indicate that although the BCI was re-calibrated between 2019 and 2020, the timing of the pilot's accuracy control strategy was consistent between 2019 and 2020. The TSD output of the 'left vs. right' (LR) classi er between 0s and 3s (reference baseline-related interval) in 2020, show less uctuation in positive or negative directions compared to that obtained in 2019. The more balanced TSD track deviation, prior to the effect of the class-speci c motor imagery task in 2020, may be because the pilot was asked to maintain a relaxed mental state whilst keeping the controlled character in the center position (i.e., at the zero baseline) during the pause period of the 'task vs. task' runs in 2020, but not in 2019. Furthermore, the track of the TSD values for the 'task vs. relax' (TX) classi er within a 2s width interval around the onset of the task (i.e., between 1s and 3s in the trial) shows that the controlled character moves in the negative direction. The above-described effect is more clearly observable in the 2020 results ( Figure 3G, 'task vs. relax') compared to the results from 2019 ( Figure 3F, 'task vs. relax'), indicating that the pilot followed the instructions and tried to relax during pause periods of the TX runs, especially in 2020.
The comparison of the frequency and topographic analysis results obtained in 2019 and 2020 using CSP and MI parameters of the calibrated BCI, indicate a similar pattern each year ( Figure 4). However, the frequency and topographical patterns obtained based on the 2020 BCI con guration involve some speci c features that are not evident in the 2019 BCI. The performance in 2019, for each binary classi er, relied on a single area in the frequency and topographical maps (indicated in A and B panels of Figure 4).
However, in 2020, for most binary classi ers, two separate areas provide similarly high contributions for the 2-class classi cation ( Figure 4C and Figure 4D). For example, the 'feet vs. right' (FR) classi cation in 2019 relied mostly on 18-28Hz (high beta) oscillations in the central area of the primary motor and somatosensory cortex -a cortical area commonly activated when able-bodied participants perform the task associated with a kinaesthetically imagined feet movement. This observation is in line with the ndings reported by Müller-Putz et al. [27], which reveals a post-movement beta rebound within a mean range of 17. 3-29.7 Hz. For the FR classi cation in 2020, in addition to the high beta activity in the central area of the primary motor and somatosensory cortex, a similarly high contribution was obtained from low gamma oscillations in the left hemisphere of the somatosensory association and occipito-temporal cortex. Thus, the above discussed CSP-MI patterns obtained in 2020 indicate task-speci c cortical activity for both compared tasks ('feet vs. right hand'), as opposed to highlighting only one task (i.e., 'feet'), as was the case in 2019. The appearance of more detailed task-speci c CSP-MI patterns in 2020 indicates a possible improvement in the pilot's 2-class classi cation control strategy compared to the BCI control strategy in 2019.
Each year, after the early sessions, which served to help the pilot achieve control con dence in the 2-class paradigm applied to the NeuroSensi game, the focus of the sessions turned to the BrainDriver game. The game completion time of the BrainDriver during the training period improved signi cantly. The game completion time in 2019 ranged from 274s to 156s over 17 days including 10 sessions wherein the mean±std statistic of game completion times in the rst BrainDriver session (session 9, Figure 5A) was 255±24s and the pilot reached 191±14s in the 10th BrainDriver sessions (sessions 16, Figure 5A). The game completion time in 2020 ranged from 230s to 168s over 18 days including 13 sessions wherein the mean±std statistic of game completion times in the rst BrainDriver session (session 9, Figure 5B) was 214±14s and the pilot reached 181±4s in the 13nd BrainDriver session (session 21, Figure 5B). However, on both competition occasions, towards the race date, the pilot's performance for controlling the BrainDriver game is decreased signi cantly. For reference the winning race times were 183s and 172s in 2019 and 2020, respectively. The results con rm not only that the multi-session online training using the BrainDriver game had a positive impact on the pilot's performance, manifested in an increased BCI control ability that enabled the pilot to achieve competitive race completion times, but also highlights that experiences in 2019 were transferred to 2020 resulting in race completion times achieved after one year were in the range of the best race completion times achieved at the end of the training performed in 2019.
In terms of other competitors, the NITRO 1 team, Benaroch et al. reported that the game completion time of their pilot uctuated between 250 and 340s during seven training sessions before the Cybathlon BCI series (2019) event [14]. However, their pilot could not nish the track within the 4 minutes limit. between 310s and 214s [29]. The best three ranked teams completed the track in the nal challenge of the Cybathlon BCI series (2019) in the following order. Rank 1st : WHI team (500m within 183s), Rank 2nd : Mirage 91 team (500m within 229s), Rank 3rd : NeuroCONCISE team (386m within the 240s limit) [30], Page 17/33 [15]. The best three ranked teams in the Cybathlon Global Edition (2020) completed the track in the following order. Rank 1st : WHI team (500m within 172s), Rank 2nd : MAHIDOL BCILAB BCI team (500m within 176s), Rank 3rd : Neurorobotics team (500m within the 213s). Our team, the NeuroCONCISE team), completed the Cybathlon Global Edition (2020) in Rank 6th (452m within the 240s limit) [31]. For completion, we applied similar but less developed strategy in Cybathlon 2016 with the same pilot.
Cybathlon 2016 is not comparable due differences in race track and total race time however our pilot achieved the 3rd best time of all competitors in the competition but had a poor qualifying lap which meant it was 6th place overall [32].
Factors that impact the performance Each year, up to a few days before the competition, the race completion time achieved by the pilot from session to session decreased, indicating an improvement in BCI control ability and/or a more re ned parameters selection for the BCI. However, a few days before the competition, each year, a signi cant increase was detected in the race completion time (dotted oval area in Figure 5A and Figure 5B).
Furthermore, on the day of the competition, the race completion time increased once again (solid oval area in Figure 5A and Figure 5B). During the last six sessions before the competition in 2019 and 2020, when this negative effect was detected, the BCI con guration had not changed, nor was there a change in the game control strategy reported by the pilot. There may be a number of factors associated with this change in performance, including increased arousal and stress levels, fatigue, and/or changes in living and dietary patterns (e.g., the pilot was living away from home for extended periods during the lead up to race day). Benaroch et al. demonstrated that their Cybathlon pilot's strategy of adapting their brain patterns to match the training data distribution, helped to improve BCI control [14]. However, they also report that they were not able to translate this to event day performances, which is consistent with our ndings. Additionally, Turi et al. reported a similar outcome for competition day results, citing multivariate factors that in uence and potentially disrupt pilot performance, including training time, change in routine, differences in the training method with the nal game, and differences in training and event environment [15]. A longitudinal study involving another Cybathlon tetraplegic pilot [29] investigated factors affecting the long-term use of their system by analyzing several performance indicators including activations maps, completion time, classi cation, and the personal experience of the pilot, by measuring their subjective experience of both their physical and mental readiness on a scale of 1 to 5. The ndings support the use of a closed-loop calibration system with real-time feedback, due to better online median classi cation performance, compared to open-loop calibration paradigms, and improved pilot engagement. Although only a single subject study, the team recommends striving to keep the training paradigm closely matched with the nal event by including closed-loop real-time feedback, which helps boost classi cation performance whilst increasing brain activations due to the increased engagement felt by the pilot. Promisingly, not all entrants reported a decrease in performance on the event day. Hehenberger et al. described a correlation between external in uencing factors and performance on their nal race [28]. As was the case for the current study, the authors describe the pilot's increased performance over time, but did not evidence a decrease in performance during the nal races. Moreover, the pilot achieved a personal best in his performance at the 2019 Cybathlon, leading the authors to speculate that their pilot performs better in front of an audience.
In an attempt to understand how to mitigate all negative in uences on pilot performance, it has been argued that the peak performance of Paralympians is driven not only by psychological factors, but through their convergence with a su cient support network, lifestyle, and attuned methods of performance [33]. Hence, recent research has focused on counteracting the effects of these variables.
One approach has been to use neurofeedback to improve BCI task performance through training selfregulation of arousal states via attention mechanisms [34]. Cognitive control underlies executive function within the brain and interacts critically with arousal systems to activate approach-avoidance behavior [34], [35]. This interaction is the principal behind the Yerkes-Dodson law -a psychological concept that posits an optimal (moderate) level of arousal is necessary for improved task performance -while a low level of arousal reduces motivation and a high level of arousal negatively impacts cognitive information processing, thus impairing task performance. Therefore, this relationship between the state of arousal and performance on a task is best described as an inverted-U, known as the Yerkes-Dodson curve [34], [36]. Arousal levels have been found to in uence sensory-motor cognition -spontaneous high-frequency oscillations known as "pilot-induced oscillation" (PIO) are generated when performing a highconsequence task. These unstable oscillations have deleterious effects on performance when ampli ed by the pilot's over-correction of small errors in control [34], [37]. Faller and colleagues (2019) used audio feedback in a closed-loop neurofeedback BCI, comprising a synthetic slow heartbeat (60bpm), which became louder with increased arousal, to train users to self-regulate their arousal levels while performing a virtual reality (VR)-based boundary avoidance task (BAT). Task performance improved signi cantly for the users who received veridical feedback compared to those in the sham or no-feedback control conditions. This result was corroborated by heartrate variability (HRV) data and measures of pupil dilation, which indicated a learned ability to shift arousal state and increase task performance through neurofeedback training.
However, it remains uncertain whether the advantage afforded by feedback would be maintained in the absence of that feedback during the competition itself. Mindfulness training could provide a more sustainable approach to the self-regulation of arousal, via underlying cognitive mechanisms.
Mindfulness training has been shown not only to have a positive impact on stress reduction [38] but is also gaining momentum in the BCI research community with some studies showing an improvement in performance for those using mindfulness over a control group of almost 20% [39]. Mindfulness is a metacognitive process as it requires self-regulation of attention to control cognitive processes -while simultaneously monitoring conscious experience [40]. Individuals who are experienced in meditation skills have been found to demonstrate higher resting SMR power, a more stable resting mu rhythm, and greater BCI control, compared to those who are not practiced in meditation techniques [41]. Furthermore, mindfulness-based stress reduction (MBSR) training has been found to improve BCI learning and performance on BCI tasks that communicate the user's intent via motor imagery commands and volitional rest [42]. The mechanism for this improvement is brought about by the user's ability to volitionally increase alpha-band neural activity as a consequence of the MBSR training. Stieger et al.
evidenced an increase in alpha activity recorded during the user's volitional resting-state, across MBSR sessions, which was correlated with mindfulness practice and predicted BCI performance [42]. Strategies to reduce challenges and stress when preparing for the competition include anticipation and preparation through detailed planning, including contingency planning, and expectation management, i.e., focusing on the process rather than the outcome [43]. Training in MBSR would t with these approaches to stress management, and enhance BCI task performance under the pressure of competition.
Factors that impact SMR-BCI performance that are less dependent on the BCI user and more dependent on external elements include distractors, time spent training, various types of feedback, and features of the EEG system and the data preprocessing algorithms [36]. For example, González-Franco et al., whose research studied the in uences of positive and negative visual feedback on motor imagery task performance using EEG and electrocardiography (ECG) found that over-biased negative feedback caused mental stress that is detected in the form of signi cantly higher heart rate variability (HRV), compared to sessions where over-biased positive feedback was presented, and accuracies correlated with the polarity (-/+) of the biased feedback [44]. Adaptive feedback, or BCI setup, that limits the negative feedback may be an alternative strategy where the onus is on the BCI to deal with the changes in the pilot's affective state. Performance may be improved by replacing the BCI framework proposed here with an adaptive BCI method that could update the BCI in real-time and adapt to the pilot's actual mental state [45]. In the lead up to the competition, the average TSD was offset from zero (indicating classi er bias to one of the classes). Even manual correction (offsetting) before each session was not enough to counteract the drift.
This baseline drift is a well-known issue in BCI and is associated with changes in the distribution of the features -covariate shift [46], [47]. As shown in Figure 6 -Figure 8 we observed changes in the temporal evolution of the frequency response and feature importance changes over time, in addition to the inclusion of some features that negatively impact performance. Although more di cult to manage, an adaptive classi er approach [48] or feature adaptation [49] or data space adaption approach [50], [51], may combat this issue and maintain performance regardless of the pilot's affective state.

Conclusion
We described the long-term training of a tetraplegic pilot in preparation to compete at the Cybathlon Competition in 2019 and 2020. Training was undertaken in the pilot's home, in hotel rooms, and at the race venues. All home based training in 2020 was supported remotely by the team (due the COVID 19 pandemic a non-expert at the pilot's home assisted with cap preparation, etc.). Our results demonstrate signi cant improvement in performance as a result of our user training strategy, BCI approach and optimisation of the BCI parameters. The results demonstrate that applying multiple binary classi ers along with additional post-processing modules and training with multiple neurogaming technologies is effective at improving the capacity to control a virtual avatar with movement, directly via brain activity, and to maximise that control ability to continue to reduce race times and achieve state-of-the art performances for the challenge. Our pilot has developed into a BCI expert, even though he has been tetraplegic for 37 years, as demonstrated by consistently achieving accuracies above 90% and competitive race times outside the competition days. We did however observe that performance was signi cantly impacted by changes in cognitive state, possibly due to heightened arousal arising from competition day pressure on the pilot. We conclude that helping the pilot to maintain consistent cognitive states is of critical importance to ensure race day performances are consistent with best training day performance. This should be supplemented by adaptive BCI strategies that can autonomously adapt to cognitive state changes to maintain performance. However, maintenance of cognitive state stability is likely to be the most important criteria for success at the Cybathlon championship for athletes with disabilities. We will focus on this and compete again at the CYBATHLON Edition in 2024. The research study presented here, received ethical approval from the Ulster University research ethics committee (UREC), and was carried out in accordance with the Declaration of Helsinki. Prior to the commencement of the training, the participant was presented with information about the experimental protocol and allowed to discuss details with the researchers and ask questions, following which consent was obtained.

Consent for publication
Not applicable

Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. using the results of the three 'task vs. task' 3-class classi ers (LF, LR, FR). The colour of the ball within the triangle is controlled by the 'task vs. relax' (TX) classi er (blue: task (T), green: relax(R)). The training timeline is shown for 2019 and 2020. In both years, the pilot was trained in his house followed by preparation at the competition location and race venue where the competition was held (Cybathalon BCI series, Technical University of Graz, 2019 and Cybathlon Global Edition, Ulster University, Derry, 2020).      for Class1 and Class2, respectively, using a baseline corrected to zero at session 10. The optimal number of mutual-information based ranked LDA features were selected for each of the four 2-class classi ers, separately. Therefore, the number of features presented for different classi ers are not the same. The highest-ranked feature indicated feature number 1. The interval between the two vertical dotted lines indicates sessions (19-21) that the pilot demonstrated the greatest improvement in BrainDriver.