Contents lists available at ScienceDirect

Behavioural Brain Research

journal homepage: www.elsevier.com/locate/bbr

Research report

Classification of somatosensory cortex activities using fNIRS

Keum-Shik Honga,b,⁎, M. Raheel Bhuttab, Xiaolong Liua, Yong-Il Shinc

a School of Mechanical Engineering, Pusan National University, Busan 46241, South Koreab Department of Cogno-Mechatronics Engineering, Pusan National University, Busan 46241, South Koreac Department of Rehabilitation Medicine, School of Medicine, Pusan National University Yangsan Hospital, Yangsan 50612, South Korea

A R T I C L E I N F O

Keywords:fNIRSMulti-sensory decodingSomatosensory cortexTactile stimulationLinear discriminant analysis

A B S T R A C T

The ability of the somatosensory cortex in differentiating various tactile sensations is very important for a personto perceive the surrounding environment. In this study, we utilize a lab-made multi-channel functional near-infrared spectroscopy (fNIRS) to discriminate the hemodynamic responses (HRs) of four different tactile sti-mulations (handshake, ball grasp, poking, and cold temperature) applied to the right hand of eight healthy malesubjects. The activated brain areas per stimulation are identified with the t-values between the measured dataand the desired hemodynamic response function. Linear discriminant analysis is utilized to classify the acquireddata into four classes based on three features (mean, peak value, and skewness) of the associated oxy-he-moglobin (HbO) signals. The HRs evoked by the handshake and poking stimulations showed higher peak valuesin HbO than the ball grasp and cold temperature stimulations. For comparison purposes, additional two-classclassifications of poking vs. temperature and handshake vs. ball grasp were performed. The attained classifi-cation accuracies were higher than the corresponding chance levels. Our results indicate that fNIRS can be usedas an objective measure discriminating different tactile stimulations from the somatosensory cortex of humanbrain.

1. Introduction

The aim of the present study is to decode the hemodynamic responses(HRs) evoked by multiple pain and touch stimuli in the somatosensorycortex using functional near-infrared spectroscopy (fNIRS). Researchershave examined different neurophysiological signals to understand theworking of the somatosensory cortex for different pain and touch sti-mulations [1]. Among those signals are event-related potentials (ERPs),which are acquired by electroencephalography (EEG) from the scalp[2,3]. Recently the authors have examined the relationship betweenempathy and social touch as reflected in attenuation of EEG oscillationsin the mu band [3]. EEG, with its high temporal resolution, detects brainsignals very quickly but the main disadvantage is its poor spatial re-solution [4,5]. The other technique widely used to detect brain areas thatare activated in the course of tactile stimulation is functional magneticresonance imaging (fMRI). fMRI, with its high spatial resolution relativeto that of EEG [6–9], can easily localize changes in the regional cerebralblood flow (rCBF). A comprehensive review of fMRI has found the cur-rent technique to be limited due to the high cost of its scanners, bulkysize, and high sensitivity to motion artifacts [10].

Researchers therefore have begun to explore another brain-imagingtechnique, fNIRS, which measures brain activity through hemodynamic

responses associated with neuron behavior [11–14]. fNIRS is a portable,noninvasive, inexpensive method of monitoring cerebral hemodynamicactivity at moderate depths (surface cortices), which makes it suitablefor the study of various stimulations in remote and challenging en-vironments [15–18]. fNIRS can detect changes in both oxy-hemoglobin(HbO) and deoxy-hemoglobin (HbR), which are strong absorbers ofnear-infrared light (650–1000 nm). In this spectrum of light, bones, skinand tissue are mostly transparent [19]. fNIRS provides a better tem-poral/spatial resolution tradeoff as compared with EEG and fMRI. Inthe recent past, researchers have used fNIRS in a variety of fields suchas neuroscience [20,21], sports medicine [22], behavioral studies[23–26], clinical medicine [27], and brain-computer interface (BCI)[28–31].

To date, only limited research has been undertaken to decode differentstimulations using fNIRS [32–37]. In all of these studies, the authors haveidentified only different brain regions or changes in HRs related to dif-ferent noxious or innocuous stimulations. The authors, meanwhile, haveapplied only specific types of noxious stimulation including thermal[32,33], painful pressure [35], electrical stimulation [36], and dental pain[34,37], and have compared the corresponding HR to innocuous stimu-lation. There is as yet no study that has used classification methods todistinguish between even noxious and innocuous stimulation.

http://dx.doi.org/10.1016/j.bbr.2017.06.034Received 25 January 2017; Received in revised form 10 June 2017; Accepted 20 June 2017

⁎ Corresponding author at: School of Mechanical Engineering, Pusan National University, 2 Busandaehak−ro, Guemjeong−gu, Busan 46241, South Korea.E-mail address: kshong@pusan.ac.kr (K.-S. Hong).

Behavioural Brain Research 333 (2017) 225–234

Available online 28 June 20170166-4328/ © 2017 Elsevier B.V. All rights reserved.

MARK

In the present study, we decoded different pain and touch stimu-lations (poking, cold temperature, handshake and ball grasp) usingfNIRS. Linear discriminant analysis (LDA) was used to distinguish be-tween the four HRs based on the mean, peak value and skewness ofHbO. Pre-processing techniques of noise removal and statistical analysiswere used to enhance classification accuracy. The results obtained inthis study demonstrate that fNIRS-determined features can be used todistinguish and classify the HRs of the four different tactile stimula-tions.

2. Materials and methods

2.1. Subjects

Eight healthy male subjects (average age: 30.75 ± 2.81 years)participated in the experiment. All of the subjects had normal sensationin their right hand and no history of any neurological disorder. All ofthem were informed about the nature of the experiment as well as itspurpose, before providing written informed consent. During each ex-periment, the subject was asked to sit down on a comfortable chair, torelax, close his eyes, and avoid any major body movement. Their righthand was placed palm-up on a table in a comfortable position. Thecomplete experimental procedure was conducted in accordance withthe Declaration of Helsinki.

2.2. Sensory stimuli

Four different sensory experiments were performed with the sub-jects. Two were passive stimuli (poking and cold temperature) and twowere active stimuli (handshake and ball grasp). In the application ofpassive stimuli, the subject does not have to move any of his body toproduce the sensation, whereas in the application of active stimuli thesubject has to move his hand and fingers to feel the sensation. Eachexperiment consisted of 10 trials. Each trial included a task duration of10 s followed by a resting-time duration of 30 s. Before the start of eachexperiment, 60 s baseline data were collected in the resting state. Eachexperiment was 460 s in total (60 + (10 + 30) × 10). For each ex-periment, prompts for the start and end of the task period were shownto the investigator using a Microsoft power point presentation withpredefined time intervals.

Fig. 1 shows the experimental paradigm for a task, in which fourtasks (poking, temperature, hands shake, and ball grasp) are seriallyperformed. For the poking experiment, the subject’s right-hand middlefinger was poked by a nail during the task period with a frequency of1.5–2.0 Hz. Before the start of the experiment, every subject was pokedtwice in order to determine the level of pain they could feel. Care wastaken to poke the same finger with the same frequency and pressure.

For the temperature sensation experiment, a computer-controlledthermode (3 × 3 cm) was used (Q-Sense, Medoc Advanced MedicalSystems, Israel). The thermode was placed on the right-hand index andmiddle fingers of the subject. The temperature was set to 20 °C. Thethermode was set at the correct temperature for the experiment, appliedto the fingers of the subject upon the start of the task period, and re-moved at the end of each task period. Care was taken to apply thethermode to the same site with the same pressure.

For the handshake experiment, the investigator placed his righthand on the subject’s right hand during each task period. Because thesubject was sitting with closed eyes, two different sounds were used to

signify the start and end of the task period so as to indicate the time toclose and open his hand, respectively.

For the ball grasping experiment, the investigator placed a billiardball in the right hand of the subject in such a way that it was nottouching the palm, the subject grasping the ball using only his thumband fingers. Two different sounds were played to signify the start andend of the task period to indicate the time to close and open his fingers,respectively.

2.3. fNIRS signal acquisition

For this study, we upgraded our own developed multi-wavelengthfNIRS system to 32 channels [38,39]. A special cap, made of nylon, wasdesigned so that it can be worn on the subject’s head tightly and thelight emitting diodes (LEDs) and photodiodes can be attached to thesubject’s scalp through the holes. While attaching the emitters (LEDs)and detectors (photodiodes) on the subject’s head, hair was moved tothe sides so that the optodes could be attached properly to the subject’sscalp. This system consists of eight tri-wavelength LEDs (L735/805/850-40B32, Epitex Inc., Japan) and fifteen photodiodes (S1223-01, SiPhotodiode, Hamamatsu Photonics, Japan). Each LED can emit light atthree wavelengths: 735 nm, 805 nm, and 850 nm. The LED has a totalradiant power of 9 mW at each wavelength. We used LEDs rather thanlasers as the light sources because LEDs are inexpensive, easy to use andthere is no need for fiber cables (LEDs being directly attachable to thehuman body). Each photodiode, used in our system, can detect light atwavelengths ranging from 320 nm to 1100 nm, with an active area of3.6 mm x 3.6 mm. A uniform intensity of LEDs and the gain of thephotodiodes were controlled by the drive circuitry. The head cap wasdesigned in such a way as to enable later accommodation of sensorsfrom an EEG system as well.

Fig. 2(a) shows the placement of the LEDs and photodiodes over theleft somatosensory cortex of the subject. The red circles indicate thedetectors and blue squares indicate the LEDs. The numbers written inbetween the LEDs and photodiodes indicate the channel numbers. Ac-cordingly, following the International 10–20 system, channel 18 waspositioned at the Cz location. It should be noted that the complete so-matosensory cortex area was covered by this configuration in the lefthemisphere. Fig. 2(b) shows the picture of a subject wearing the cap.Each LED was turned on and off sequentially, and the light diffusedthrough the cortical region was detected at the nearest detectors. Atotal of 64 light-intensity signals (8 LEDs × 2 wavelengths × 4 neigh-boring detectors) were acquired at a sampling rate of 5.3 Hz. Finally,during the experiment, all the lights in the room were switched off tominimize signal contamination from the ambient light sources.

2.4. Data processing

The signals from the fNIRS system were imported and further ana-lyzed and classified offline using MATLAB™ 7.9.0 (MathWorks, USA).The data were stored in a host-computer text file in the form of digitizedraw-intensity values. From those values, the changes in optical densityΔA, could be calculated at each discrete time k as follows.

= − = ⋅ ⋅ΔA k λ I λI λ

l D λ Δμ k λ( , ) ln ( )( )

( ) ( , ),aout

in (1)

where λ is the wavelength of light, Iin is the intensity of incident light,Iout is the intensity of detected light, l is the distance between the

Fig. 1. Experimental paradigm: An experiment includes a total of 10 trialspreceded by 60 s rest.

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

226

emitter and the detector, D is the differential pathlength factor (DPF),and Δμa is the change of the absorption coefficient of the tissue. Thechanges of oxy-hemoglobin (ΔcHbO) and deoxy-hemoglobin (ΔcHbR)were measured using the modified Beer-Lambert law [40,41]

⎡⎣⎢

⎤⎦⎥

= ⎡⎣⎢

⎤⎦⎥

× ⎡⎣⎢

⎤⎦⎥

−Δc kΔc k

l D λ α λ l D λ α λl D λ α λ l D λ α λ

ΔA k λΔA k λ

( )( )

( ) ( ) ( ) ( )( ) ( ) ( ) ( )

( , )( , )

,HbO

HbR

1 HbO 1 1 HbR 1

2 HbO 2 2 HbR 2

11 1

2 2

(2)

with λ1 = 735 nm, λ2 = 850 nm, D(λ1) = 6.3125, and D(λ2) = 5.235[42], according to the values for the wavelength-dependent absorptioncoefficients αHbO, αHbR taken from the UCL Department of MedicalPhysics and Bioengineering website [43]. Since HbO signals are moredirect to the given stimuli than HbR signals (i.e., the signal-to-noise

ratio of HbO is higher than that of HbR) [44], the mean, peak value,and skewness values only of the HbO concentration changes were usedin the subsequent analysis.

fNIRS, in detecting the hemodynamic response, picks up the phy-siological noises of respiration, pulse and low-frequency Mayer waves.To remove these noises and to minimize signal variations, fNIRS dataneed to be pre-processed. Since each of our trials included a 10 s taskperiod followed by a 30 s rest period, the frequency in each trial wasapproximately 0.025 Hz (i.e., 1/40). First, we applied a second-orderlow-pass filter with a cutoff frequency of 0.15 Hz to remove the phy-siological noises of respiration (about 0.3 Hz) and cardiac activity(about 1 Hz). Second, the detrending technique was used to remove thetrend of the signal (e.g., a low-frequency drift) from the time-series data[45]. For this purpose, we utilized Matlab™’s detrend function, which

Fig. 2. Channel configuration: a) Locations of emit-ters (LEDs) and detectors (photodiodes) on the leftsomatosensory cortex; the red circles indicate thedetectors and the blue squares represent the emitters,the numbers denote the channels, and b) a picture ofthe cap on a participant. (For interpretation of thereferences to colour in this figure legend, the readeris referred to the web version of this article.)

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

227

removes the best fit straight-line (i.e., x= at+ b) from the given data x(t). In the final step of pre-processing, due to the wide variation in thefiltered data, the rescaling method was applied [46].

2.5. Brain activation map

In this study, we adopted the t-value comparison between themeasured data and the desired hemodynamic response function (dhrf)to select the active trials for further analysis [47]. The dhrf, which is theexpected HR for a given stimulus, was computed by convolving thestimulus pattern (i.e. 10 s task and 30 s rest) with the canonical HRfunction. In this study, we used the canonical HR function with thefollowing parameters: dispersion of response: 1 s, dispersion of under-shoot: 1 s, delay of response from onset: 6 s, delay of undershoot: 16 s,ratio of peak to undershoot: 6, length of kernel: 21 s (see [41] for fur-ther information on the canonical HR function).

The t-value is the ratio of the weighting coefficient to the modeledHR (in the process of fitting the measured HR to the dhrf), and itsstandard error. In a t-test, we normally try to find the evidence of asignificant difference between the measured HR and the dhrf. The t-value measures the size of the difference relative to the variation in oursample data. A positive high t-value indicates that the measured signalis highly correlated with the dhrf, whereas a near to zero or a negative t-value indicates that the measured HR is significantly different from thedhrf. For each trial, the t-value was computed for all 32 channels, andthen the averaged t-value for each channel was computed by taking themean of all 10 trials. To calculate the t-value, we used the robustfitfunction available in Matlab™. The robustfit function, [b, stat] = ro-bustfit (dhrf, t1_S(:,i)), uses robust regression to fit the measured HR of aspecific trial t1_S of specific channel i as a function of dhrf, and returnsthe vector b of coefficient estimates and a stat structure with the sta-tistical data including the t-value, p-value, standard error, etc. Thechannels whose t-value was greater than the critical t-value (tcrt) was

selected for further analysis. In this study, the value of tcrt used was1.652. This value is computed from the degree of freedom (the no. ofdata points in one trial) and the statistical significance level. In thisstudy, the total number of data points in one trial was 213, and thesignificance level was set to 0.05 for one tailed test [47].

All t-values were normalized within the range of 0–1 using thefollowing equation

′ = −−

a a aa a

min( )max( ) min( )

,(3)

where a denotes the original t-value, a′ the normalized t-value, and min(a) and max(a) the minimum and maximum t-values in the respectivetrial. All t-values were normalized to keep equal standard for all plots.

2.6. Feature selection and classification

The active channels for each stimulation were averaged together tomake a mean signal, and then the trials were separated from the meansignal. In this study, the signal mean (SM), peak value (SP) and skew-ness (SK) of HbO were used as the classification features. The SM valuewas computed by averaging the data points within a 0.5–15 s windowin order to differentiate the occurrence of activation from the restingstate. The SP was calculated by taking the maximum value of the signalwithin 0.5–15 s window. The SK was measured for the same 0.5–15 swindow to represent the asymmetry of a signal in terms of the prob-ability distribution around its mean relative to a normal distributionand, thus, differentiate the shapes of HbO signals.

In this study, LDA, due to its simplicity and execution speed, wasused for the offline classification of four different stimuli. LDA is alinear classifier that discriminates between different classes of databased on hyper-planes [19]. The separating hyper-plane is designed insuch a way as to minimize the interclass variance and maximize thedistance between the class means. For classification purposes, we used

Fig. 3. Activation map (handshake).

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

228

40 signals (10 trials × 4 experiments) for each subject. The classifica-tion accuracies were calculated using 10 runs of 10-fold cross-valida-tion; that is, the data were randomly distributed into 10 sets (i.e., 40/4 = 10), trained on 9 datasets and tested on 1. This process was re-peated 10 times and after which the mean accuracy was obtained.

3. Results

Figs. 3–6 plot the t-values for each trial and each experiment for alleight subjects. These plots were made to check the consistency of theactive channels in each experiment across all ten trials. The numbers

Fig. 4. Activation map (ball grasp).

Fig. 5. Activation map (poking).

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

229

shown in the figure represent the channel numbers in the covered brainregion, and the color in each pixel represents the signal intensity. Fig. 7plots the t-values calculated by comparing the averaged HR for eachactivity with the dhrf. The averaging was done over all ten trials and alleight subjects. Because the data used were averaged over all eightsubjects, the plot demonstrates the overall activation trends. Fig. 8 isthe representation of data presented in Fig. 7 in a bar graph format.

Fig. 9 compares the average HbO signals and its standard deviation

for all four stimulations applied to the subjects. The averaging wasperformed over all 10 trials and eight subjects for all of the stimula-tions. The shaded areas around the mean signals represent the standarderrors.

Fig. 10 represents Subjects 2′s three-dimensional (3D) plot for theSK, SM, and SP values of HbO (x-, y-, and z-axes, respectively). Thegraph displays the data for all 40 features (10 trials of each experi-ment), which are marked by crosses (× in red color) for handshake,

Fig. 6. Activation map (cold temperature).

Fig. 7. Comparison of the averaged activation maps (over all 8 subjects) in the left somatosensory cortex upon four different stimulations.

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

230

circles (o in blue color) for ball grasp, steric (* in black color) forpoking, and plus (+ in magenta color) for temperature. It can be seenthat the feature values were scaled between 0 and 1 using Eq. (3).

Cross-subject analysis was performed to determine the consistencyof the classification accuracies across the subjects. Table 1 provides theclassification accuracies with their standard deviations for individualsubjects across all four stimulations. The average classification accuracywas 50.31 ± 2.81%. After investigating the classification accuraciesfor the four stimulations, we also investigated the classification ac-curacies within the active and passive stimuli. Table 1 additionallyprovides the classification accuracies for both of these investigations.The average classification accuracy for poking vs. cold-temperaturestimuli was 78.12 ± 2.87% and, for handshake vs. ball-grasp stimuli,it was 75.94 ± 2.04%.

4. Discussion

In this study, the authors used the fNIRS technique to decode theresponses of four different tactile stimulations from the left somato-sensory cortex of the brain. fNIRS, because of its noninvasiveness, lowcost, and ease of operation, is widely used in brain imaging studies.Previous studies have reported that, along with the high activation ofprimary somatosensory cortex, different brain areas including the pre-frontal cortices, anterior insular cortex and thalamus were also acti-vated during tactile stimulation as applied to subjects [1]. In this study,four different tactile stimulations (handshake, ball grasp, poking andcold temperature) were investigated.

To compare the different brain activities, first the spatial and tem-poral information of the fNIRS signals were analyzed. The activation

Fig. 8. Bar graphs of averaged t-values.

Fig. 9. Averaged HbO responses and their standard deviations of four stimulations (data were averaged over all ten trials and eight subjects).

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

231

maps shown in Figs. 3–6 illustrate the spatial distribution of activationamong the 32 channels. The plots were shown to illustrate the con-sistency of activation area in each trial for the specific experiment. Thedata shown in these figures show that, for a specific task, the activatedbrain region is consistent within most of the trials for individual sub-jects. Small inconsistencies were found in some of the trials but thesecan be attributed as the trial-to-trial variation present in the fNIRSsignals [48,49]. Figs. 3–6 also demonstrate the variations, in activatedbrain region, present within the subjects for the same task. The subjectwise variations in the somatosensory cortex activities could be seen inthe previous fNIRS literature as well [50].

Fig. 7 illustrates the spatial distribution of activation among the 32channels within the area under observation. The activation map wasgenerated by computing the correlation level between the measured HRand the dhrf for a specific stimulus. The activation maps for the fourdifferent stimulations indicate that for different stimulations differentbrain regions are activated. The activation map can only provide in-formation about the active location in the brain; it cannot reveal anyinformation about the signal strength in a given location. Therefore,along with the activation map, the temporal magnitude of HR (shown inFig. 9) was examined to determine the intensity of the brain activity.Fig. 9 is the averaged HbO response (averaged over all 10 trials and alleight subjects) for the individual stimulations. Fig. 9 shows clear andsignificant differences between the HbO signals corresponding to thefour stimulations. It serves to demonstrate, thereby, that the responses

for each stimulation can be separated from the others based on therespective relevant HRs from the somatosensory cortex.

The classification was performed using LDA to distinguish HbOsignals as fNIRS data features. Fig. 10 shows a 3D scatter plot for thefour different classes (hand shake, grasp, poking, and temperature)after classification of the fNIRS data based on the selected features(mean, skewness and peak value) for Subject 2. The separability of theHRs arising during the four stimulations is clear. The average classifi-cation accuracy achieved in this study was 50.31 ± 2.81%. Regardingthe HR-based classification using fNIRS, the previous study by Poweret al. has performed a 3-class classification (mental singing, mentalarithmetic, and rest) and achieved an average classification accuracy of56.2% [51]. Hong and Santosa [47] have obtained an average classi-fication accuracy of 46% while classifying four different sounds in theauditory cortex of brain. In another study by Naseer and Hong [13], theauthors have reported an average four-class classification accuracy of73.3% for the prefrontal and motor cortex signals. But in [13] the au-thors have used two different brain regions and distinctive tasks toacquire four different brain signals whereas in this study, we used onlythe left somatosensory cortex to acquire four different brain signals andthe average classification accuracy of 50.31%. In a four-class classifi-cation study, the chance level is 25% (i.e., 100/4), and our obtainedclassification accuracies for all subjects were higher than the chancelevel; even the lowest accuracy from Sub. 1 was 37.25 ± 4.72%. In thetwo-class classification cases, average classification accuracies of78.12 ± 2.87% and 75.94 ± 2.04% were obtained for the passivetasks (poking vs. cold temperature) and active tasks (handshake vs. ballgrasp), respectively. It is noted that the classification accuracies variedamong the subjects and even among the different tasks. These varia-tions can be attributed to both individual-subject differences and trial-to-trial variations in the signals caused by background activity or as-yet-unknown sources [48,49].

In this study, the experiments were performed only with malesubjects aged 27–36, though differences in the HRs of females and olderindividuals relative to those of young males have already been reported[52–54]. One additional factor that should be noted is that all of oursubjects were healthy subjects, whereas HRs can differ for patients withlocked-in syndrome (LIS), amyotrophic lateral sclerosis (ALS) or anyother neurological disease [55]. One final factor is that, in this ex-perimental paradigm, the experiments were performed separately toincrease the number of trials per experiment. If the same stimulationswere provided in a random order during the same session [47,56], the

Fig. 10. 3D scatter plot of the mean, peak, and skewness values of the HbO of 40 trials (Sub. 2).

Table 1Four- and two-class classification accuracies for all eight subjects.

Subject No. Four-classclassificationaccuracies [%]

Poking vs. coldtemperatureclassificationaccuracies [%]

Handshake vs. ballgrasp classificationaccuracies [%]

1 37.25 ± 4.72 64.5 ± 4.38 68 ± 2.582 55.25 ± 2.65 76.5 ± 1.6 72 ± 2.583 54.25 ± 1.69 61.5 ± 5.48 85 ± 1.24 51 ± 2.43 74.5 ± 2.83 71.5 ± 3.375 46 ± 3.16 90 ± 1.14 70.5 ± 1.66 53.25 ± 2.06 91.5 ± 2.41 80 ± 2.117 54.25 ± 3.07 80 ± 1.74 75 ± 1.278 51.25 ± 2.7 83.5 ± 3.37 85.5 ± 1.58Average 50.31 ± 2.81 78.12 ± 2.87 75.94 ± 2.04

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

232

HRs and classification accuracies would have differed and the paradigmwould be more suitable for BCI purposes.

5. Conclusions

This study investigated the feasibility of functional near-infraredspectroscopy (fNIRS) for classification of hemodynamic responsesevoked by four different tactile stimulations (handshake, ball grasp,poking and cold temperature). The data were collected from the leftsomatosensory cortex of the brain utilizing our own developed multi-channel fNIRS system. The active channels for individual stimulationswere mapped by t-value comparison between the measured signal andthe desired hemodynamic response function. The classification of thefour different stimulations was performed using linear discriminantanalysis as a classifier. The higher classification accuracies obtained inthis study show that the hemodynamic responses for the different tac-tile stimulations are distinguishable. The overall results of this studyindicate that fNIRS offers a great potential for application to the de-coding of different tactile stimulations in the somatosensory cortex.

Acknowledgments

This work was supported by the National Research Foundation ofKorea under the auspices of the Ministry of Science, ICT and FuturePlanning, Korea (grant no. NRF-2014-R1A2A1A10049727)

References

[1] A.V. Apkarian, M.C. Bushnell, R.D. Treede, J.K. Zubieta, Human brain mechanismsof pain perception and regulation in health and disease, Eur. J. Pain 9 (2005)463–484.

[2] E. Niechwiej-Szwedo, J. Chin, P.J. Wolfe, C. Popovich, W.R. Staines, Abnormalvisual experience during development alters the early stages of visual-tactile in-tegration, Behav. Brain Res. 304 (2016) 111–119.

[3] L. Peled-Avron, E. Levy-Gigi, G. Richter-Levin, N. Korem, S.G. Shamay-Tsoory, Therole of empathy in the neural responses to observed human social touch, Cogn.Affect. Behav. Neurosci. 16 (2016) 802–813.

[4] A. Turnip, K.-S. Hong, M.Y. Jeong, Real-time feature extraction of P300 componentusing adaptive nonlinear principal component analysis, Biomed. Eng. Online 10(2011) e83.

[5] C. Popovich, W.R. Staines, Acute aerobic exercise enhances attentional modulationof somatosensory event-related potentials during a tactile discrimination task,Behav. Brain Res. 281 (2015) 267–275.

[6] Z. Yuan, J. Ye, Fusion of fNIRS and fMRI data: identifying when and where he-modynamic signals are changing in human brains, Front. Hum. Neurosci. 7 (2013)e676.

[7] M. Ghio, P. Schulze, B. Suchan, C. Bellebaum, Neural representations of novel ob-jects associated with olfactory experience, Behav. Brain Res. 308 (2016) 143–151.

[8] M.M. Gillis, S. Garcia, B.M. Hampstead, Working memory contributes to the en-coding of object location associations: support for a 3-part model of object locationmemory, Behav. Brain Res. 311 (2016) 192–200.

[9] J. Kim, Y.G. Chung, S.C. Chung, H.H. Bülthoff, S.P. Kim, Decoding pressure sti-mulation locations on the fingers from human neural activation patterns,Neuroreport 27 (2016) 1232–1236.

[10] M. Verriotis, P. Chang, M. Fitzgerald, L. Fabrizi, The development of the nociceptivebrain, Neuroscience 338 (2016) 207–219.

[11] G.D. Tempest, R.G. Eston, G. Parfitt, Prefrontal cortex haemodynamics and affectiveresponses during exercise: a multi-channel near infrared spectroscopy study, PLoSOne 9 (2014) e95924.

[12] U. León-Domínguez, J.F. Martín-Rodríguez, J. León-Carrión, Executive n-back tasksfor the neuropsychological assessment of working memory, Behav. Brain Res. 292(2015) 167–173.

[13] N. Naseer, K.-S. Hong, Decoding answers to four-choice questions using functionalnear infrared spectroscopy, J. Near Infrared Spectrosc. 23 (2015) 23–31.

[14] H.D. Nguyen, K.-S. Hong, Y.I. Shin, Bundled-optode method in functional near-in-frared spectroscopy, PLoS One 11 (2016) e0165146.

[15] C.-K. Kim, S. Lee, D. Koh, B.-M. Kim, Development of wireless NIRS system withdynamic removal of motion artifacts, Biomed. Eng. Lett. 1 (2011) 254–259.

[16] M.J. Khan, M.J. Hong, K.-S. Hong, Decoding of four movement directions usinghybrid NIRS-EEG brain-computer interface, Front. Hum. Neurosci. 8 (2014) e244.

[17] S.K. Piper, A. Krueger, S.P. Koch, J. Mehnert, C. Habermehl, J. Steinbrink, H. Obrig,C.H. Schmitz, A wearable multi-channel fNIRS system for brain imaging in freelymoving subjects, Neuroimage 85 (2014) 64–71.

[18] M.R. Bhutta, M.J. Hong, Y.-H. Kim, K.-S. Hong, Single-trial lie detection using acombined fNIRS-polygraph system, Front. Psychol. 6 (2015) e709.

[19] N. Naseer, K.-S. Hong, Classification of functional near-infrared spectroscopy

signals corresponding to the right-and left-wrist motor imagery for development ofa brain–computer interface, Neurosci. Lett. 553 (2013) 84–89.

[20] H.D. Nguyen, K.-S. Hong, Bundled-optode implementation for 3D imaging infunctional near-infrared spectroscopy, Biomed. Opt. Express 7 (2016) 3491–3507.

[21] A. Zafar, K.-S. Hong, Detection and classification of three-class initial dips fromprefrontal cortex, Biomed. Opt. Express 8 (2016) 367–383.

[22] V. Quaresima, R. Lepanto, M. Ferrari, The use of near infrared spectroscopy insports medicine, J. Sports Med. Phys. Fit. 43 (2003) 1–13.

[23] A. Roos, F. Robertson, C. Lochner, B. Vythilingum, D.J. Stein, Altered prefrontalcortical function during processing of fear-relevant stimuli in pregnancy, Behav.Brain Res. 222 (2011) 200–205.

[24] H. Santosa, M.J. Hong, K.-S. Hong, Lateralization of music processing with noises inthe auditory cortex: an fNIRS study, Front. Behav. Neurosci. 8 (2014) e418.

[25] M.J. Khan, K.-S. Hong, Passive BCI based on drowsiness detection: an fNIRS study,Biomed. Opt. Express 6 (2015) 4063–4078.

[26] K.-S. Hong, N. Naseer, Reduction of delay in detecting initial dips from functionalnear-infrared spectroscopy signals using vector-based phase analysis, Int. J. NeuralSyst. 26 (2016) e1650012.

[27] M.C. Taillefer, A.Y. Denault, Cerebral near-infrared spectroscopy in adult heartsurgery: systematic review of its clinical efficacy, Can. J. Anesth. 52 (2005) 79–87.

[28] S. Fazli, J. Mehnert, J. Steinbrink, G. Curio, A. Villringer, K.R. Müller, B. Blankertz,Enhanced performance by a hybrid NIRS–EEG brain computer interface,Neuroimage 59 (2012) 519–529.

[29] N. Naseer, M.J. Hong, K.-S. Hong, Online binary decision decoding using functionalnear-infrared spectroscopy for the development of brain–computer interface, Exp.Brain Res. 232 (2014) 555–564.

[30] L.C. Schudlo, T. Chau, Single-trial classification of near-infrared spectroscopy sig-nals arising from multiple cortical regions, Behav. Brain Res. 290 (2015) 131–142.

[31] N. Naseer, K.-S. Hong, fNIRS-based brain-computer interfaces: a review, Front.Hum. Neurosci. 9 (2015) e3.

[32] L. Becerra, W. Harris, M. Grant, E. George, D. Boas, D. Borsook, Diffuse opticaltomography activation in the somatosensory cortex: specific activation by painfulvs. non-painful thermal stimuli, PLoS One 4 (2009) e8016.

[33] L. Becerra, W. Harris, D. Joseph, T. Huppert, D.A. Boas, D. Borsook, Diffuse opticaltomography of pain and tactile stimulation: activation in cortical sensory andemotional systems, Neuroimage 41 (2008) 252–259.

[34] A.J. Racek, X. Hu, T.D. Nascimento, M.C. Bender, L. Khatib, D. Chiego,G.R. Holland, P. Bauer, N. McDonald, R.P. Ellwood, A.F. DaSilva, Different brainresponses to pain and its expectation in the dental chair, J. Dent. Res. 94 (2015)998–1003.

[35] N. Üçeyler, J. Zeller, S. Kewenig, S. Kittel-Schneider, A.J. Fallgatter, C. Sommer,Increased cortical activation upon painful stimulation in fibromyalgia syndrome,BMC Neurol. 15 (2015) e210.

[36] M.A. Yücel, C.M. Aasted, M.P. Petkov, D. Borsook, D.A. Boas, L. Becerra, Specificityof hemodynamic brain responses to painful stimuli: a functional near-infraredspectroscopy study, Sci. Rep. 5 (2015) e9469.

[37] N. Higaki, T. Goto, T. Ichikawa, Periodontal tactile input activates the prefrontalcortex, Sci. Rep. 6 (2016) e36893.

[38] M.R. Bhutta, K.-S. Hong, B.-M. Kim, M.J. Hong, Y.-H. Kim, S.-H. Lee, Note: threewavelengths near-infrared spectroscopy system for compensating the light absor-bance by water, Rev. Sci. Instrum. 85 (2014) e026111.

[39] M.R. Bhutta, S.-W. Woo, M.J. Khan, K.-S. Hong, Effect of anodal tDCS on humanprefrontal cortex observed by fNIRS, 6th IEEE Int. Conf. on Biomed. RoboticsBiomechat. (BioRob) (2016) 957–961.

[40] L. Kocsis, P. Herman, A. Eke, The modified Beer–Lambert law revisited, Phys. Med.Biol. 51 (2006) N91–N98.

[41] K.-S. Hong, H.D. Nguyen, State-space models of impulse hemodynamic responsesover motor somatosensory, and visual cortices, Biomed. Opt. Express 5 (2014)1778–1798.

[42] F. Scholkmann, M. Wolf, General equation for the differential pathlength factor ofthe frontal human head depending on wavelength and age, J. Biomed. Opt. 18(2013) e105004.

[43] UCL Department of Medical Physics and Biomedical engineering, Tissue Spectra,(2005) (Accessed 20.12.2016), http://www.ucl.ac.uk/medphys/research/borl/intro/spectra.

[44] L. Holper, M. Biallas, M. Wolf, Task complexity relates to activation of corticalmotor areas during uni-and bimanual performance: a functional NIRS study,Neuroimage 46 (2009) 1105–1113.

[45] H. Santosa, M.J. Hong, S.-P. Kim, K.-S. Hong, Noise reduction in functional near-infrared spectroscopy signals by independent component analysis, Rev. Sci.Instrum. 84 (2013) e073106.

[46] K.-S. Hong, N. Naseer, Y.-H. Kim, Classification of prefrontal and motor cortexsignals for three-class fNIRS-BCI, Neurosci. Lett. 587 (2015) 87–92.

[47] K.-S. Hong, H. Santosa, Decoding four different sound-categories in the auditorycortex using functional near-infrared spectroscopy, Hear. Res. 333 (2016) 157–166.

[48] X.S. Hu, K.-S. Hong, S.S. Ge, Reduction of trial-to-trial variability in functional near-infrared spectroscopy signals by accounting for resting-state functional con-nectivity, J. Biomed. Opt. 18 (2013) e017003.

[49] L. Holper, N. Kobashi, D. Kiper, F. Scholkmann, M. Wolf, K. Eng, Trial-to-trialvariability differentiates motor imagery during observation between low versushigh responders: a functional near-infrared spectroscopy study, Behav. Brain Res.229 (2012) 29–40.

[50] C. Habermehl, S. Holtze, J. Steinbrink, S.P. Koch, H. Obrig, J. Mehnert,C.H. Schmitz, Somatosensory activation of two fingers can be discriminated withultrahigh-density diffuse optical tomography, Neuroimage 59 (2012) 3201–3211.

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

233

[51] S.D. Power, A. Kushki, T. Chau, Automatic single-trial discrimination of mentalarithmetic, mental singing and the no-control state from prefrontal activity: towarda three-state NIRS-BCI, BMC Res. Notes 5 (2012) e141.

[52] G. Wang, N. Erpelding, K.D. Davis, Sex differences in connectivity of the subgenualanterior cingulate cortex, Pain 155 (2014) 755–763.

[53] S. Brodoehl, C. Klingner, O.W. Witte, Age-dependent modulation of the somato-sensory network upon eye closure, Behav. Brain Res. 298 (2016) 52–56.

[54] B. Wang, M. Zhang, L. Bu, L. Xu, W. Wang, Z. Li, Posture-related changes in brain

functional connectivity as assessed by wavelet phase coherence of NIRS signals inelderly subjects, Behav. Brain Res. 312 (2016) 238–245.

[55] M. Naito, Y. Michioka, K. Ozawa, Y. Ito, M. Kiguchi, T. Kanazawa, A communica-tion means for totally locked-in ALS patients based on changes in cerebral bloodvolume measured with near-infrared light, IEICE Trans. Inf. Syst. 90 (2007)1028–1037.

[56] X. Liu, K.-S. Hong, Detection of primary RGB colors projected on a screen usingfNIRS, J. Innov. Opt. Health Sci. 10 (2017) e1750006.

K.-S. Hong et al. Behavioural Brain Research 333 (2017) 225–234

234