Talking with hands and feet: Selective somatosensory attention and fMRI enable robust and convenient brain-based communication

In brain-based communication, voluntarily modulated brain signals (instead of motor output) are utilized to interact with the outside world. The possibility to circumvent the motor system constitutes an important alternative option for severely paralyzed. Most communication brain-computer interface (BCI) paradigms require intact visual capabilities and impose a high cognitive load, but for some patients, these requirements are not given. In these situations, a better-suited, less cognitively demanding information-encoding approach may exploit auditorily-cued selective somatosensory attention to vibrotactile stimulation. Here, we propose, validate and optimize a novel communication-BCI paradigm using differential fMRI activation patterns evoked by selective somatosensory attention to tactile stimulation of the right hand or left foot. Using cytoarchitectonic probability maps and multi-voxel patterns analysis (MVPA), we show that the locus of selective somatosensory attention can be decoded from internally generated fMRI signal patterns from fMRI-signal patterns in (especially primary) somatosensory cortex with high accuracy and reliability, with the highest classification accuracy (85.93%) achieved when using Brodmann area 2 (SI-BA2) at a probability level of 0.2. Based on this outcome, we developed and validated a novel somatosensory attention-based yes/no communication procedure and demonstrated its high effectiveness even when using only a limited amount of (MVPA) training data. For the BCI user, the paradigm is straightforward, eye-independent, and requires only limited cognitive functioning. In addition, it is BCI-operator friendly given its objective and expertise-independent procedure. For these reasons, our novel communication paradigm has high potential for clinical applications.


Introduction
Social behavior and communication are essential to human nature. When interacting with others, humans not only use spoken language but also nonverbal behavior to reveal their intentions. For healthy people, these actions are easy to perform and self-evident but they rely heavily on the functional integrity of the neuromuscular system.
The neuromuscular system can be crucially compromised in several clinical conditions ( e.g. , motor neuron disease, multiple sclerosis, stroke), resulting in severe motor paralysis. In an extreme case, the socalled complete 'locked-in' syndrome (CLIS; Guger et al., 2017 ), patients suffer from a total loss of motor function while remaining fully conscious and maintaining their sensory and cognitive abilities as well as their emotional experience. In the most devastating case, natural communication is impossible ( Laureys et al., 2004 ).
One way to restore communication with severely motor-impaired patients is using a brain-computer interface (BCI). A BCI is a system that uses voluntarily modulated brain signals (instead of motor output) has many obvious advantages ( e.g. , high temporal resolution, portability, relatively low costs), some disadvantages can still be noted. EEG has a relatively low spatial resolution and depth pervasion, meaning that only superficial brain activity can be measured and exploited for BCI implementations ( Naci et al., 2012 ). The main limitation, however, is that not all (even healthy) individuals can elicit robust EEG responses exploitable for BCI purposes -a phenomenon called EEG-BCI illiteracy ( Blankertz et al., 2008 ). It is of crucial importance to increase the spectrum of available BCI options by considering other brain imaging techniques. This will increase the chance that at least one BCI approach might work in a patient in need of an alternative communication channel.
As an alternative to electrophysiological approaches, hemodynamic brain-imaging techniques, such as functional magnetic resonance imaging (fMRI; Boly et al., 2007 ;Monti et al., 2010 ;Owen et al., 2006Owen et al., , 2007Yoo et al., 2004 ) and functional near-infrared spectroscopy (fNIRS; Abdalmalak et al., 2017Abdalmalak et al., , 2020Benitez-Andonegui et al., 2020 ;Coyle et al., 2007 ;Gallegos-Ayala et al., 2014 ;Nagels-Coune et al., 2020 ;Naito et al., 2007 ) have been suggested and explored for brain-based communication and control in recent years. Due to the high single-trial reliability of the fMRI/fNIRS responses, hemodynamic BCIs enable effective communication and control often right from the very first session. In addition, hemodynamic BCIs require less setup time and practice for the participant ( Nagels-Coune et al., 2020 ;Sorger et al., 2012 ). While fNIRS has the advantage of being portable and relatively easy to use, fMRI provides superior spatial resolution and full brain coverage ( Naci et al., 2012 ). Moreover, it requires only a limited setup time and MRI scanners are part of standard hospital equipment. Therefore, an fMRI-based BCI setup can be implemented relatively quickly at many clinical sites. Recent advancements in real-time fMRI data analysis ( e.g. , Goebel, 2021 ;LaConte et al., 2007 ;Mathiak and Posse, 2001 ) have put forward further methodological development and application of fMRIbased BCI systems ( Bardin et al., 2012 ;Kaas et al., 2019 ;Sorger et al., 2009Sorger et al., , 2012. The majority of fMRI-BCI studies used mental imagery (auditory, emotion, motor, tactile, and visual imagery) to hemodynamically encode BCI commands Kaas et al., 2019 ;LaConte, 2011 ;Monti et al., 2010 ;Owen et al., 2006Owen et al., , 2007Senden et al., 2019 ;Sorger et al., 2009Sorger et al., , 2012Yoo et al., 2004Yoo et al., , 2003Yoo et al., , 2001. However, mental imagery imposes high cognitive load on participants ( e.g. , maintaining information in working memory) and for some individuals, mental imagery is very difficult or even impossible to perform. Therefore, attention-based BCI paradigms have been suggested -which might require alternative mental abilities and are potentially less cognitively demanding for the BCI user. For example, Naci et al. (2013) employed selective auditory attention to let individuals (including patients) encode yes/no answers. The latter BCI paradigm is especially favorable, as it does not require intact visual capabilities. Generally, fMRI-BCI paradigms rely on the visual sensory modality ( LaConte et al., 2007 ;Sitaram et al., 2007 ;Sorger et al., 2009Sorger et al., , 2012Yoo et al., 2004 ). While a few visual-BCI approaches exist that do not require voluntary eyemovement control ( e.g., Reichert et al., 2020 ), most visual BCI strategies are suboptimal or not suited at all for the target patient populations who often suffer from severe visual impairments. An alternative, very promising sensory BCI-input modality might be the tactile domain. This approach may be particularly suited for unresponsive patients ( e.g. , LIS and CLIS), given that the tactile modality is often still preserved at different stages of the disease ( Guger et al., 2017 ;Kaufmann et al., 2013 ;Murguialday et al., 2011 ;Patterson and Grabois, 1986 ). The tactile modality has already been exploited in some EEG-BCI studies (for an overview see Huang et al., 2022 ). For example, Guger et al. (2017) successfully applied an EEG-based tactile BCI to established basic communication in nine out of 12 (C)LIS patients.
To our knowledge, the only study that employed the tactile domain for hemodynamic brain-computer interfacing is a recent study from our lab in which topographic somatosensory imagery was used as an information-encoding strategy ( Kaas et al., 2019 ). Here, we propose, optimize and validate a novel selective somatosensory attention paradigm for fMRI-based communication of yes/no answers by letting healthy participants attend to vibrotactile stimulation of either the right hand or left foot. The novel fMRI-BCI paradigm is inspired by the studies of Goltz et al. (2013Goltz et al. ( , 2015 which demonstrated attentional modulatory effects within the primary and secondary somatosensory cortex (SI and SII) when participants alternatively attended to vibrotactile stimulation of the right or left hand. For BCI purposes, unilateral handmotor/sensory tasks may be suboptimal as they often produce bilateral brain activation -with the result of overlapping activity in both BCI conditions and inferior classification accuracy. From a theoretical viewpoint, the risk of overlapping activity can be reduced by implementing different body sites that have spatially distinct neural representations ( e.g., hand and foot). Note that while hand-and foot-related brain activation is closer in (brain) space than activation evoked for the right/left hand, fMRI offers the spatial resolution required to still reliably differentiate adjacent brain activation. Moreover, from the BCI user's perspective, it may be more convenient/easier to focus on body parts that are as different as possible. When applying vibrotactile stimulation to the right hand/left foot in our novel paradigm, the attentional foci are spatially more distinct (thus might feel more different) then when using vibrotactile stimulation to the right/left hand.
As a first aim, the current study investigated whether the locus of selective somatosensory attention to the right hand or left foot (cued by auditory instructions) can be reliably decoded from fMRI activation patterns across the whole brain or within specific brain regions (SI and SII). To increase future clinical applicability, we explored several data-analysis strategies for developing an effective and efficient decoding procedure. For example, to maximize sensitivity ( i.e. , classification accuracy), we used multi-voxel pattern analysis (MVPA). In contrast to univariate analysis, MVPA is considered more sensitive to latent multidimensional representations and processes ( Davis et al., 2014 ) which are involved in higher cognitive functions, such as attention. A priori-defined regions of interest were selected for the MVPA based on published cytoarchitectonic probability maps ( Eickhoff, Amunts, et al., 2006Geyer et al., 1999Geyer et al., , 2000Grefkes et al., 2001 ). We exploit the inter-subject consistency of these maps for developing a more efficient, objective, and time-saving methodology for the BCI operator. To determine the optimal combination of location and probability level of the used cytoarchitectonic maps, we systematically investigated the effect of these factors on classification accuracy.
As a second aim, the current study investigated whether selective somatosensory attention can be used for brain-based communication of yes/no answers. In BCI applications, it is highly desirable to keep the duration of the procedure as short as possible. This is especially important in clinical contexts where time and costs constitute crucial factors. Additionally, and most pressing, the patient's cognitive abilities might be considerably limited. To optimize BCI performance and usability for future clinical applications, we systematically investigated the effect of varying a) the amount of training data ( i.e. , the number of Classifiertraining trials ) and b) the amount of testing data ( i.e. , the number of trial-repetitions considered for classification; see below) on the classification accuracy. By using an appropriate amount of training and testing trials, not only time and financial resources can be saved, but also the burden for the investigated patient will be limited to what is necessary.
For hemodynamic BCIs, the user has to be able to apply an explicit task or strategy to control their brain signal and thereby the BCI output. For convenient communication, this task should be intuitive, almost instant to control, and not cognitively overwhelming for the BCI user. In addition, the BCI strategy should rely on modalities that are still preserved in patients to achieve successful communication. Since the somatosensory modality is one of the few modalities still preserved in (C)LIS patients ( Murguialday et al., 2011 ), it constitutes an alternative control strategy for BCI communication. Therefore, we tested whether the locus of selective somatosensory attention can be reliably decoded To summarize, we investigated the following research questions: 1 Can the locus of selective somatosensory attention to stimulation of the right hand and the left foot be reliably decoded from fMRI brain activation? 2 Using probabilistic atlas-based cytoarchitectonic somatosensory regions, what is the optimal combination of considered location (SI or SII) and size (determined by the across-subject overlap probability level) resulting in the highest classification accuracy? 3 Can selective somatosensory attention to either right-hand or the left-foot stimulation be used to conveniently and successfully communicate yes/no answers? Note that for communication purposes, classification accuracies should reach at least the threshold of 70% in a 2-class BCI ( Kübler et al., 2001 ).

Participants and ethics
Nine healthy participants (five men; age: M = 26.4 years, SD = 4.2 years; one left-handed) were recruited among students and staff members from the Faculty of Psychology and Neuroscience at Maastricht University (Maastricht, The Netherlands; see Table 1 for an overview). Before participation, they gave written informed consent and were screened for MRI compatibility. The study was approved by the Ethics Review Committee of the Faculty of Psychology and Neuroscience (ERCPN) at Maastricht University  and was in accordance with the Declaration of Helsinki.

Materials
2.2.1. Selective somatosensory-attention paradigm 2.2.1.1. General concept. Selective attention BCI paradigms are based on the idea that two different attentional states, associated with two distinct brain-activation patterns, can be reliably decoded. It has been shown that different attentional (brain) states can be decoded successfully from hemodynamic signals in the auditory ( e.g. , Riecke et al., 2017 ) and visual domains ( e.g. , Monti et al., 2013 ). In the present study, we investigated whether the tactile modality can be exploited. As a first step, we investigated whether the locus of selective somatosensory attention to the right index finger or two left toes (instructed by auditory stimulation) can be reliably decoded from fMRI brain activation. As a second step, we tested whether the novel somatosensory-attention paradigm can be successfully used for basic brain-based yes-no communication.

Stimulation.
Vibrotactile stimulation in the flutter range (25 Hz) was generated by a piezostimulator (QuaeroSys, St. Johann, Germany). Transducers were attached with hook-and-loop tape to the right index finger and the great and second left toes. Each transducer had ten plastic pins arranged in a 2 × 5 matrix, with a pin diameter of 1.5 mm and an inter-pin distance of 2.5 mm. The vibrotactile stimulation was simultaneously and continuously applied to the right index finger and the two left toes with randomly occurring interruptions of fixed 150ms duration. Interruption patterns were different for the hand and foot stimulation to facilitate the attentional process. The number of interruptions per run ranged between 84 and 102 ( M = 92.47, SD = 5.07). The auditory speech stimuli used for instructions ( "hand ", "foot ", "start ", and "rest ") were computer-generated (neutral female voice).

Instructions to the participants.
In the first part of the study ( i.e. investigating whether the locus of selective somatosensory attention can be reliably decoded from fMRI brain activation), participants were auditorily guided to focus their attention to either the right hand ( "hand ") or left foot ( "foot ") stimulation and count the number of interruptions of the vibrotactile stimulation that occurred on the to-be-attended body part (tactile gap-counting task). After each functional run, participants were asked to report the number of counted interruptions. In the second part of the study ( i.e. testing the feasibility of yes-no communication based on selective somatosensory attention), participants freely chose the locus of attention depending on their selected answer ( right hand for "yes " and left foot and "no ") by performing the tactile gap-counting task. This task was used to increase the participants' engagement and to monitor their attention. The outcome of the participants 'reports ( Supplemental Figure 1 and Supplemental Table 1 ) suggests that participants followed the task instructions and focused their attention on the to-beattended body part.

Experimental design
The attention-trial design of the two study parts was identical: 2 s after the onset of the auditory instruction for focusing somatosensory attention ( "hand ", "foot ", "start "), the vibrotactile stimulation started. The stimulation lasted for 16 s during which participants performed the gapcounting task. The trials were alternated with 18-22 s jittered resting periods (indicated to participants by "rest ") where no stimulation was applied and participants had no specific task to perform. Each of five to six Classifier-training runs (first study part) consisted of nine right-hand and nine left-foot somatosensory attention trials that appeared in pseudorandom order. Each of the four Communication runs (second study part) consisted of five trials encoding the same answer for the same question allowing to investigate single as well as multi-trial response accuracies. During Communication runs , participants encoded two "yes " and two "no " answers to avoid bias in the answer selection. However, they were free to choose the order of answers so that the researchers were blind to the answer given in a particular run. Classifier-training runs lasted 11:20 min and Communication runs took 3:12 min. To make the experiment as varied as possible for the participants, we performed the Communication runs after the first three Classifier-training runs . See Fig. 1 for a visualization of the run order, stimulation, and task protocol.

General procedure
The experiment was performed in two sessions, one training session and one (f)MRI scanning session .
Training session. Two to five days before fMRI scanning, participants attended a 1-h training session. Participants were introduced to the study and performed two Classifier-training and two Communication runs in a mock scanner . The purpose was to familiarize the participants with the fMRI environment and to practice the selective somatosensory attention paradigm. The typical fMRI noise was not simulated but all the steps as performed in a real fMRI session (providing earplugs, headphones, placing coil over their heads and performing coil alignment) were included.
(f)MRI scanning session. Participants' heads were fixated with foamed cushions to minimize head movement. Participants were further provided with headphones for enabling auditory instruction during scanning and were asked to lie still and to keep their eyes closed during all functional runs. The latter was based on the participants' reports after the training session that closing the eyes was beneficial for focusing attention. Once the participants felt comfortable in the scanner, the fMRI session started with the shimming procedure and an anatomical scan. For organizational reasons, participants S01 and S02 completed six Classifier-training runs only. The remaining participants (S03-S09) completed five Classifier-training runs and four Communication runs . After the scanning session, participants rated the subjectively experienced easiness level and vibrotactile stimulation strength of the attention task using a Likert scale (1-10). For the majority of participants, stimulation to the right hand was rated to be stronger as compared to the left foot (see Supplemental Table 1 ). In addition, participants' reports indicated that the attention task was easy to perform (see Supplemental Figure 2 ). Finally, participants reported the order of the encoded answers to an independent researcher.

(f)MRI data acquisition
(f)MRI data were acquired at the Maastricht Brain Imaging Center (Maastricht University, The Netherlands) using a 7-T MRI scanner (Siemens Healthcare, Erlangen, Germany) equipped with a 32 receiverchannel (Nova) head coil. Anatomical images were obtained using a three-dimensional T1-weighted Magnetization Prepared 2 Rapid Acquisition Gradient Echoes sequence (MP2RAGE; Marques et al., 2010 ). The following scanning parameters were used: 240 slices, no gap, 0.7 mm isovoxel resolution, matrix size: 256 ×256 mm, field of view: 224 ×224 mm, TR/TE = 5000/2.47 ms, flip angle = 5°The total scanning time was 8:02 min. Functional images were acquired using echoplanar imaging (EPI). Scanning parameters were: 78 slices, 1.5 mm isovoxel resolution, matrix size: 96 ×96, field of view: 198 ×198 mm, TR/TE = 2000/21.6 ms, flip angle = 60°The number of volumes was 340 ( Classifier-training runs ) and 96 ( Communication runs ) resulting in a scanning time of 11:20 min and 3:12 min, respectively. Note that for each participant, five functional volumes were acquired in opposite phaseencoding direction to be used for EPI-distortion correction during data analysis.

(f)MRI data analyses
fMRI data obtained in the first study part ( i.e., investigating whether the locus of selective somatosensory attention can be reliably decoded) was analyzed using BrainVoyager 22.2 (Brain Innovation, Maastricht, the Netherlands). Brain data for the second part of the study ( i.e., testing the feasibility of yes-no communication based on selective somatosensory attention) was analyzed in simulated real-time using Turbo-BrainVoyager 4.2 (Brain Innovation B.V., Maastricht, the Netherlands).

fMRI data analysis of first study part (decoding locus of selective somatosensory attention)
Pre-processing. Anatomical data were corrected for intensity inhomogeneities and transformed to Talairach space. Pre-processing of functional data started with head-motion detection and correction using trilinear/sinc interpolation. Per participant, the first volume of each run was aligned to the first volume of the first functional run. Further preprocessing included slice-scan time correction, high-pass filtering (cutoff: 6 cycles/run), EPI-distortion correction ( Cope v0.5 BrainVoyager plugin , Brain Innovation, Maastricht, the Netherlands), co-registration with anatomical data, and Talairach transformation.
Definition of regions of interest (ROIs) . Analysis of the functional data that was submitted to MVPA was restricted to multiple regions of interest including the whole functional volume measured as well as several predefined Brodmann (BA) and Operculum (OP) regions derived from the cytoarchitectonic probabilistic maps of the Jülich Center ( Eickhoff, Amunts, et al., 2006Geyer et al., 1999Geyer et al., , 2000Grefkes et al., 2001 ). Besides the whole functional volume, the more restricted regions were: SI and SII merged, SI and SII separately and in total eight sub-regions within SI (SI-BA1, SI-BA2, SI-BA3a, and SI-BA3b) as well as SII (SII-OP1, SII-OP2, SII-OP3, and SII-OP4). The whole-functional volume ROI was defined individually for each participant and included all functional voxels covering brain tissue. For the more spatially restricted ROIs, different overlap probability levels of the cytoarchitectonic maps were used ranging from 0.1 (10% overlap) to 1.0 (100% overlap). To give an example, an anatomical mask with a probability level of 0.6 (60%) contains voxels that were classified as belonging to that particular cytoarchitectonically-defined region in at least 60% of the postmortem human brains in the database used to construct the probability map. Note that the higher the probability level, the smaller the mask size (see Fig. 2 C ). For the region SII-OP1 at the probability level of 1.0, no ROI could be defined (no single voxel shared across postmortem brains). Therefore, MVPA was performed within a total of 110 ROIs.
Decoding the locus of selective somatosensory attention using MVPA. Decoding accuracies were obtained using linear classifiers (support vector machines, SVMs) following a (run-level) cross-validation procedure. All voxels within a particular ROI (see above) were used to define the features for the MVPA. Parameter values (t-values) were estimated in a temporal (peri -onset) window ( − 2 s to 18 s) for each trial, attention condition, and voxel applying a general-linear model (GLM) fitting pro-cedure using a standard 2-gamma hemodynamic response function. For each run-level split, data of four runs were combined to train the SVM and the remaining runs were used to test classifier performance. In total, 110 classification accuracies (for the 110 different ROIs) were obtained per participant. For every single trial, the classifier's prediction was compared to the condition supposed to be performed by the participant. Individual-and group-mean accuracies of the five run-level splits were calculated for each of the 110 ROIs.
To check for a potential bias of the SVM classifier, individual-and group-mean classification accuracies were obtained for the two attention conditions separately (focusing on the right hand and left foot). This was done only for the ROI-probability level combination that showed the highest decoding accuracy on the group level.
Significance testing. To assess whether classification accuracies reached significance ( p < .05), we ran permutation tests ( n = 1000) for each case to obtain a distribution of accuracy values under the null hypothesis. This was performed for the optimal ROI-probability level combination and separately for all five run-level splits per participant. An average accuracy over all splits was calculated as well. Moreover, a Wilcoxon Signed-Ranked test was used to investigate whether there was a significant difference between hand and foot trials (not hypothesized).

fMRI data analysis of second study part (testing novel communication paradigm)
Pre-processing. Functional data were analyzed based on slice data in native resolution (1.5mm 3 ) in simulated real-time using Turbo-BrainVoyager v4.2 (Brain Innovation B.V., Maastricht, the Netherlands). Pre-processing of functional data included head-motion detection and correction (using the standard incremental Turbo-BrainVoyager procedure) and spatially realigning each functional volume to the first recorded volume of the first functional run of each participant. Functional data were also high-pass filtered ( Classifier-training runs : 6 cycles/run; Communication runs : 2 cycles/run), spatially smoothed with a 2 mm full-width-at-half-maximum (FWHM) kernel, and transformed into Talairach space. Anatomical and functional data were co-registered.
ROI selection. Simulated real-time data analysis was done using the 'optimal' ROI-probability level combination that had resulted in the highest group-mean decoding accuracy in the first study part.
Obtaining communication accuracies. In the simulated real-time data analysis, the same MVPA approach (including MVPA-feature extraction) as in the offline data analysis was followed. This was motivated by previous real-time MVPA studies demonstrating the feasibility and potential of such an approach for BCI purposes ( Beauchamp et al., 2009 ;LaConte et al., 2007 ). SVM 'training' was based on the data of one to five Classifier-training runs allowing for the investigation of the effect of the amount of training data. The Communication runs' data were used for testing the SVM-classifier performance. Single-trial and multi-trial classification accuracies were calculated for each SVM classifier, separately for all participants and the group. Single-trial accuracies were calculated as mean accuracies across all individual trials. For obtaining multi-trial accuracies, the means of the six single-trial classifier values per Communication run were calculated and considered for the decision. Fig. 2 A shows the classification accuracies (group means) for the regions SI, SII, and SI + SII merged at probability levels 0.1-1.0 as well as for the whole functional volume measured. The highest accuracies were obtained for SI (84.94%; see Supplementary Table 2 for all mean accuracies). Fig. 2 B shows the group means for all BAs separately (see Supplementary Table 3 for all mean accuracies) revealing that area SI-BA2 at a probability level of 0.2 (20%) resulted in the highest classification accuracy of 85.93%. The individual classification accuracies for this area can be inferred from Fig. 3 (yellow bars). Note that for all individual run-level splits ( n = 45 for five splits per participant) investigated for SI-BA2, the obtained classification accuracy was significantly above the empirical chance level as assessed by permutation testing (see Supplemental Table 4 ). Fig. 2 C shows the different mask sizes for each probability level and somatosensory region (SI and SII). Note that higher probability levels correspond to smaller mask sizes.

Classification accuracies separately for left-foot and right-hand attention trials (for SI-BA2)
Further results are described for the ROI-probability level combination that yielded the best classification accuracy, i.e., SI-BA2 at probability level 0.2. Fig. 3 shows participants' classification accuracies separately for left-foot and right-hand attention trials and all trials together. These individual classification accuracies were generally high ranging from 80% − 95.56% for left-foot trials, from 73.33% − 93.33% for righthand trials, and 76.67% − 93.33% when considering all trials. The groupmean accuracy for attention trials to the left foot, the right hand, and all trials was 84.89%, 82.67%, and 85.93%, respectively (see Supplemental  Table 5 for individual accuracies). A Wilcoxon Signed-Rank test indicated that there was no significant difference between right-hand and left-foot classification accuracies ( Z = − 0.70, p = .48). Fig. 4 A shows the individual single-and multi-trial communication accuracies for five Classifier-training runs . The mean individual singletrial communication accuracy was high (83.57%) and ranged from 75% to 90% on the individual level. As expected, higher accuracies were obtained with multi-trial analyses, resulting in a mean multi-trial communication accuracy of 92.86% and an individual range from 75% to 100%. Fig. 4 B shows the effect of the amount of training data (range: one to five Classifier-training runs ) on mean single-and multi-trial communication accuracies. Communication accuracies ranged from 66.43% to 83.57% and 75% to 92.86% for single-and multi-trial mean accuracies, respectively (see Supplementary Table 6 for all classification accuracies). The minimum accuracy threshold suggested for a 2-class BCI (70%; Kübler et al., 2001 ) was reached using two Classifier-training runs for single-trial accuracies (73.57%) and using only one Classifiertraining run for multi-trial accuracies (75%). In addition, a clear trend can be seen for the single-trial accuracies. The classification accuracies increased up to using four Classifier-training runs and seemed to level off when adding more training data . In contrast, the multi-trial accuracies further increased when including up to five Classifier-training runs .

Discussion
The aim of the current study was to propose, validate and optimize a novel selective somatosensory attention paradigm for convenient and effective fMRI-based yes/no communication. We first investigated whether the locus of selective somatosensory attention to either the right index finger or two left toes can be reliably decoded from fMRI brain activation patterns within the whole functional volume measured or within pre-defined task-relevant (somatosensory) brain regions. More particularly, we systematically investigated which region at which across-subject overlap probability level (when using cytoarchitectonic maps) resulted in the highest classification accuracy. We finally tested the novel fMRI-based yes/no communication procedure and investigated the effect of varying a) the amount of training data ( i.e. , the number of trials considered for classifier training) and b) the amount of testing data ( i.e. , the number of trial-repetitions considered for classification) on the classification accuracy.

Fig. 2. Classification accuracies for different probability levels and somatosensory areas.
Panel A shows the classification accuracies (group means) for each probability level ×ROI (SI, SII, and SI + SII merged) combination and for the whole functional volume. The highest accuracies were obtained for SI with a classification accuracy of 84.94%. Panel B shows the group means for all BAs separately. It demonstrates that area SI-BA2 at a probability level of 0.2 (20%) resulted in the highest classification accuracy (85.93%). Panel C shows the different mask sizes for each probability level for SI and SII, with the number of corresponding voxels shown between brackets. Note that the higher the probability level, the smaller the mask size. Abbreviations: SI, primary somatosensory cortex; SII, secondary somatosensory cortex; BA, Brodmann area; OP, Operculum; ROI, region of interest.

The locus of selective somatosensory attention can be decoded with high accuracy, especially within the primary somatosensory cortex
High classification accuracies were obtained within the whole functional volume and the primary (SI) and secondary (SII) somatosensory cortex for almost all tested overlap probability levels of the cytoarchitectonic atlas ( Fig. 2 A ). A general clear trend could be observed, namely that considering lower overlap probability levels ( i.e. , larger ROI size) increased classification accuracies. The highest classification accuracy was obtained for SI (84.94% at probability level 0.3) followed by the whole functional volume (73.83%) and SII (73.09% at probability level 0.1; Fig. 2 A ). Results obtained for the combined SI + SII ROI did not exceed the classification accuracies obtained for SI alone. These findings are similar to the selective somatosensory imagery 2-class BCI study of Kaas et al. (2019) which demonstrated that higher classification accuracies were obtained for SI (up to 82%), with lower probability levels resulting in higher classification accuracies. The lower SII classification accuracies can be explained by the fact that the somatotopy in SII is considerably less fine-grained than in SI -with SII neurons having larger receptive fields resulting in partially overlapping fMRI-activation patterns during hand and foot stimulation ( Ruben et al., 2001 ).
When considering the different subregions of SI (BA1-3b) and SII (OP1-4), significant classification accuracies were obtained for almost all probability levels (see Fig. 2 B ). The highest classification accuracy was obtained for SI-BA2 at a probability level of 0.2 (85.93% ) -again with a clear trend that a lower probability level ( i.e. , larger ROI size) resulted in higher classification accuracies ( Fig. 2 B ). Note that for all 45 individual run-level splits for SI-BA2, the obtained classification accuracy was significantly above the empirical chance level (see Supplemental  Table 4 ). The area-specific attentional modulation within SI is similar to the study of Goltz et al. (2013) where it was demonstrated that attention to vibrotactile stimulation to the right hand induced activity in BA2 of the primary somatosensory cortex.
The participants experienced the gap-counting task as generally straightforward while reporting slightly higher easiness ratings for the attention-to-hand condition (see Supplemental Figure 2 ). However, classification results for left-foot and right-hand attention trials did not show significant differences indicating that classification was not biased to one attention condition (group-mean accuracies for SI-BA2 at probability level 0.2 being 84.89% and 82.67%, respectively).
While SI-BA2 at probability level 0.2 yielded the highest classification accuracy, several combinations of SI-BAs and (especially low) probability levels resulted in similarly high classification accuracies (see Fig. 2 B ). This indicates that selective somatosensory attention is represented rather widely across SI subregions but probably not so much in SII and beyond. Therefore, several primary somatosensory subregions and right-hand (red bars) attention trials as well as for all trials together (yellow bars) for area SI-BA2 at probability level 0.2. Classification accuracies ranged from 80% − 95.56% for left-foot trials, from 73.33% − 93.33% for righthand trials, and 76.67% − 93.33% for all trials together. The mean accuracy for attention to the left foot, right hand, and all trials together was 84.89%, 82.67%, and 85.93%, respectively. Abbreviations: S01-S09, participant number.
might be well-suited as BCI-input regions. Note also that using the entire SI mask resulted in only slightly lower classification accuracies ( ca. − 1% on average). Using SI at probability level 0.3 might also be a viable option for future clinical applications as the larger SI-ROI size might be less vulnerable to head motion (which is more likely in patients) and less affected by brain damage and following brain plasticity.

Convenient and effective yes/no communication based on selective somatosensory attention
As the ultimate aim of the study, we tested whether selective somatosensory attention can be utilized for a simple yes/no brain-based communication approach. Across participants, we observed a very high run-based (multi-trial) communication accuracy of 92.9%, with five out of the seven tested participants reaching 100% ( Fig. 4 A ). Every participant could successfully use our communication paradigm as shown by all individual communication accuracies surpassing the 70-% criterion, which is considered sufficient for effective communication in a 2-class BCI ( Kübler et al., 2001 ).
The obtained across-participant communication accuracy was similarly high in a recent 2-class fMRI-BCI study that utilized selective somatosensory imagery and observed an average run-based communication accuracy of 92.3% ( Kaas et al., 2019 ). However, in the current study, a considerably lower number of training trials ( − 20%) was needed while implementing an identical trial design. As a potential reason, our novel attention -based communication paradigm might be less cognitively demanding for the BCI user ( e.g. , concerning cognitive load/working memory). In mental imagery, individuals have to generate the internal stimulus themselves, while in an attention paradigm, stimulation is provided and therewith facilitates the encoding process.
The mean individual single-trial communication accuracy was also high (83.57%) and ranged from 75% to 90% on the individual level ( Fig. 4 A ). Therewith, our novel paradigm could also be implemented as a single-trial BCI paradigm. However, trial averaging ( i.e. , following a multi-trial approach) might still be necessary for clinical applications (due to lower data quality) or could allow for considerably reducing the amount of training data (see Fig. 4 B ).
These findings demonstrate that selective somatosensory attention can be considered an almost instant control strategy for fMRI-based communication. As attention potentially requires alternative mental abilities, it constitutes a viable additional option to the mental imagery-based encoding approach.
In BCI applications, it is highly desired to keep the duration of the procedure as short as possible. When looking at the communication accuracies obtained with different amounts of training data ( Fig. 4 B ), it can be concluded that, for surpassing the generally accepted 70-% threshold for a 2-class BCI ( Kübler et al., 2001 ), only one and two Classifier-training runs would be needed in the multi-and single-trial approach, respectively. By tailoring the appropriate amount of training data to individuals and specific situations, not only time and financial resources can be optimally used, but also the burden for the BCI user can be limited. This is especially important in clinical contexts in which all these factors are crucial. Fig. 4. Single-and multi-trial communication accuracies. Panel A shows the individual singleand multi-trial communication accuracies for five Classifier-training runs for area SI-BA2 at probability level 0.2. The single-trial communication accuracies ranged from 75% to 90% (mean: 83.57%). The multi-trial communication accuracies ranged from 75% to 100% (mean: 92.86%), with five out of the seven tested participants reaching 100%. Panel B shows the effect of the amount of Classifiertraining data for single-and multi-trial mean classification accuracies for SI-BA2 at probability level 0.2, which ranged from 66.43% to 83.57% and 75% to 92.86%, respectively. Abbreviations: S03-S09, participant number.
Overall, we consider the novel communication paradigm to be a viable option in clinical situations where MRI facilities are available; it is straightforward as it is easy-to-apply and requires relatively low cognitive and mental capabilities. It is also robust and efficient, which is demonstrated by its high communication accuracy and immediate applicability, respectively. Clinical populations that might benefit from the developed procedure are especially severely paralyzed patients but also other unresponsive patients ( e.g. , patients with disorders of consciousness). In the latter, our somatosensory-attention approach could also be used to detect remaining consciousness via wilful modulation of brain activity ( e.g. , Bardin et al. 2012, Owen et al., 2006.

BCI-user and -operator friendliness beneficial for future clinical applications
Generally, to achieve successful brain-based communication, BCI control strategies should rely on sensory modalities that are still preserved in unresponsive patients. Most developed fMRI-BCI paradigms relied on the visual modality ( LaConte et al., 2007 ;Sitaram et al., 2007 ;Sorger et al., 2009Sorger et al., , 2012Yoo et al., 2004 ). While a few visual-BCI approaches exist that do not require voluntary eye-movement control ( e.g. , Reichert et al., 2020 ), most visual BCI strategies are suboptimal or not suited at all for the target patient populations who often suffer from severe visual impairments. The proposed tactile BCI paradigm takes potential visual impairments of patients into account. It is fully independent of visual input as it exclusively relies on auditory instructions and vibrotactile stimulation. We think implementing the auditory and tactile domain for BCI applications is especially promising in (C)LIS patients as these sensory modalities are often well preserved through different stages of the disease ( Guger et al., 2017 ;Kaufmann et al., 2013 ;Murguialday et al., 2011 ;Patterson and Grabois, 1986 ) and can therefore still be experienced on a daily basis which constitutes a crucial advantage. A previous study relying on neuroelectric brain signals already exploited the tactile and auditory domain and enabled basic communication even in (C)LIS patients, with similar communication accuracies ( Guger et al., 2017 ). Comparable to the publication by Kaas et al. (2019) , Fig. 5. Classification accuracies for different probability levels and somatosensory areas (3-T fMRI data). Panel A shows the classification accuracies (group means) for each probability level ×ROI (SI, SII, and SI + SII merged) combination and for the whole functional volume. The highest accuracies were obtained for SI with a classification accuracy of 81.11%. Panel B shows the group means for all BAs separately. It demonstrates that area SI-BA2 at a probability level of 0.4 (40%) resulted in the highest classification accuracy (82.78%). Abbreviations: SI, primary somatosensory cortex; SII, secondary somatosensory cortex; BA, Brodmann area; OP, Operculum; ROI, region of interest.
the current study demonstrates that also hemodynamic brain responses within the primary somatosensory cortex can be used for effective communication of binary responses in healthy participants.
In recent hemodynamic BCI applications, mostly mental imagery (auditory, emotion, motor, tactile, and visual imagery) was used to encode BCI commands Kaas et al., 2019 ;LaConte, 2011 ;Monti et al., 2010 ;Owen et al., 2006Owen et al., , 2007Senden et al., 2019 ;Sorger et al., 2009Sorger et al., , 2012Yoo et al., 2004Yoo et al., , 2003Yoo et al., , 2001. However, for some (even healthy) individuals, mental imagery is very difficult or even impossible to perform ( Naci et al., 2012 ) due to its generally high cognitive load. For example, mental imagery requires maintaining a considerable amount of information in working memory and has to be fully self-initiated. In contrast, in our novel paradigm, participants receive vibrotactile stimulation and just have to pay attention to the provided stimulation which requires only little practice and effort. This should considerably limit the cognitive burden. The convenience of somatosensory attention is supported by the high easiness ratings of the study participants (see Supplemental Figure 2) . For all these reasons, we think that our novel BCI paradigm might be a well-fitting approach for patients lacking certain cognitive capacities.
Next to the advantages for the BCI user, the proposed procedure is easy to implement for the BCI operator as well. By using an optimal a priori -defined brain region from given cytoarchitectonic maps, the decoding procedure is fully objective. Therewith, no specific expert knowledge on functional neuroanatomy or the individual definition of regions of interest is required and considerable time and effort can be saved. This expertise independence improves the BCI usability for future clinical applications.

Limitations and future work
This study only included a limited number of mainly young and healthy participants. More research is needed to test whether the current paradigm can be applied with similar success in larger and more diverse samples. Given the obtained high decoding accuracies (92.86% in the multi-and 83.57% in the single-trial approach), it is quite likely that this BCI application might also work in populations for which lower signal quality is expected. For example, in clinical populations, signalquality loss might happen as a result of brain damage and/or atrophy as well as due to cognitive impairments ( e.g. , attention deficits).
In our study, brain-based communication was realized by using simulated real-time data analysis which means that answers were decoded only after the session. Therewith, no back-and-forth communication ( e.g. , Sorger et al., 2009Sorger et al., , 2012 was possible. Note, however, that even offline procedures can be very useful, especially in clinical situations ( e.g. , Bardin et al., 2012 ;Monti et al., 2010 ). Finally, as the real-time analysis procedure would be identical to the performed simulated realtime procedure, we do not expect a significant drop in accuracy in a real-time analysis situation. Real-time feedback on the decoding outcome could, however, positively affect the BCI performance, e.g., due to higher motivation of participants and the possibility of feedback learning.
The communication paradigm was tested for the best atlas-based ROI-probability level combination on the group level ( e.g. , SI-BA2 at probability level 0.2). The resulting ROI might not optimally match with individual functional neuroanatomy. In principle, we could have taken the most optimal ROI-probability level combination on the individual level (obtained from the first study part) or anatomically/functionally define the ROI. This might have resulted in even higher classification accuracies. However, we did not follow this more individualized approach because it is more time-consuming while more immediate communication may be desired, especially in clinical settings.
Another time-saving methodological strategy might be the "betweensubject " approach. Training data of (as many as possible) participants would be combined and the resulting classifier would be used to classify communication data of another (new) participant who would not have to perform C lassifier-Training runs. From a theoretical point of view, there is a considerable chance that the effects that we exploited in our BCI paradigm are spatially robust enough to go for a "between-subject " approach.
The current study was performed on a 7-T fMRI scanner. While increasing, MRI scanners with this (high) field strength are very rare as compared to 3-T scanners. The measurements were performed at 7-T fMRI only for logistical reasons and the paradigm should also work with more common clinical MRI scanners as indicated by the results of our 3-T fMRI pilot study (see Fig. 5

and Supplementary Material 3-T Pilot Study ).
A disadvantage of an fMRI-based communication procedure like ours is still that the availability of MRI scanners is limited and associated costs are high. However, it might be possible to transfer our novel communication paradigm to fNIRS -a method that is based on the same hemodynamic principles. fNIRS is gaining more and more interest in the field of BCI research especially due to its lower costs, ease of application, and mobility ( Cui et al., 2011 ). Given that somatosensory experiences to hand and foot stimulation can be measured with fNIRS ( Eto et al., 2014 ;Hong et al., 2017 ), it constitutes a suitable alternative candidate for a somatosensory-attention BCI. If successful, an fNIRS-based procedure would allow for regular communication in daily life which is the ultimate goal of BCI development. Note that we deemed the righthand/left-foot approach to be suitable for hemodynamic BCIs. It may still be an optimal strategy to use the right/left-hand approach for EEGbased BCIs considering the lower spatial resolution of EEG ( Guger et al., 2017 ).

Conclusion
This is the first study that uses hemodynamic brain activation evoked by different somatosensory attentional states for BCI communication. The locus of selective somatosensory attention to either the right hand or the left foot can be reliably decoded from fMRI response patterns in the primary somatosensory cortex resulting in high yes/no communication accuracies. In addition, only limited training data is needed to achieve effective communication. The novel hemodynamic informationencoding paradigm is convenient for BCI users by being straightforward, eye-independent, and requiring only limited cognitive capabilities. It is also BCI-operator friendly as it is objective and expertise-independent. Therewith, the paradigm has a high potential to be applied in various clinical populations for brain-based communication and control as well as the detection of remaining awareness in disorders-of-consciousness patients. Its potential could be further enhanced if a move to more costeffective and mobile fNIRS were successful.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability
The datasets generated and analyzed during the current study are not available due to ethical and privacy restrictions. Participants did not explicitly consent to their data being made public. The regulations regarding data sharing adopted by the authors comply with the requirements of the Faculty of Psychology and Neuroscience, Maastricht University and the EU legislation on the general data protection regulation (GDPR).

Ethics approval and consent to participate
This study was carried out with approval from the Ethics Review Committee of the Faculty of Psychology and Neuroscience (ERCPN) at Maastricht University . According to the Declaration of Helsinki, all participants gave written informed consent before participation.