All protocols were approved by the Hallym University Animal Care and Use Committee, and were performed according to the Guide for the Care and Use of Laboratory Animals of the U.S. National Institutes of Health as well as the guidelines of the Animal Welfare Act. Adult Sprague-Dawley (SD) rats (350~400 g, male, n=5) were used in this study. The rats where obtained from the experimental animal center of Orient Bio Co. (linked to the Charles River Lab.). The breeding room environment was maintained at a temperature of 23±2℃ and relative humidity of 55±10%. Artificial lighting was maintained for 12 hours per day. Two animals were housed in each cage with unlimited food and water. Subjects were anesthetized with i.p. injections of urethane (20%, 1.25 g/kg body weight) and they were transferred to a stereotaxic apparatus and fixed in a prone position. Prior to the surgery and optical recording, we checked their motor responses. After a one-inch middle incision and removal of the skin and soft tissue, the skull was exposed.

We defined the coordinates of the main olfactory (MOB) regions (anterior to posterior: +4.0 mm~+12.0 mm, later to medial: 0~±4.0 mm) from bregma point on the bilateral hemisphere [8].

A commercial sixteen source-channel frequency-domain NIRS system (Imagent, ISS, IL., USA) was used to measure the cerebral hemodynamic responses to olfactory stimulation. This system uses two wavelengths (690 nm, 830 nm), and each channel contains two optical fibers with 400 µm core diameter (FT400EMT, Thorlabs, NJ, USA) for 690 nm and 830 nm. The optical fibers were arranged as shown in Fig. 1A and a picture of the actual experiment is shown in Fig. 1B. The NIRS channels, source and detector, were separated by 7 mm and the corresponding penetration depth is approximately 2 mm, which is sufficient to cover the cortex area [10]. They encompass both left and right olfactory bulbs. After the optical fibers were placed at appropriate locations, dental cement (KetacCem, 3M, USA) was applied to the rat skull to maintain a constant boundary condition between optical fibers and skull [11, 12]. The sampling rate of the NIRS system was 28.4 Hz.

Odorants were diluted with mineral oil to the expected dilution rate (approximately 350 PPM) based on individual odorant vapor pressure. Following the study of Wilson (2008), mixtures were produced by odorant supplements to mineral oil in amounts that provided identical component dilution rates within the mixture [13]. Fig. 2 shows the odor stimulatior sytem and the NIRS sytem where odorants were delivered by delivered by a motorized odor stimulator and presented 5 mm from the rat's nose for 10 seconds with plain air. The approximate dilution rate was 1 : 1 with plain air. Fig. 3 explains each stimulus was delivered at an interval of about 170 seconds and repeated 20 times for each experiment.

We changed silicone tubes and the bottles containing the odorant each time we used another odorant to prevent unexpected dilution rate. Odorants are: (i) plain air as a reference (Blank), (ii) 2-Heptanone (HEP), (iii) Isopropylbenzene (IB).

We used an analysis program of Boas and Huppert's [14] to calculate cerebral hemodynamic responses. This program provides calculation of ΔHbO₂ and ΔHbr concentration, block average process and image reconstruction. For the recorded intensities from each channel, a modified Beer-Lambert law used to calculate the hemoglobin concentration.

where OD means optical density, I_Final is the measured light intensity, and I_initial is the initial light intensity. ε is the extinction coefficient of oxy-hemoglobin and deoxy-hemoglobin, and B indicates differential path length factor. λ is the wavelength of light (690 nm and 830 nm). Using this ΔOD^λ, we can obtain ΔHbO₂ (oxy-hemoglobin concentration) and ΔHbr (deoxy-hemoglobin concentration) with the following equations:

and

To extract clear patterns about hemodynamic responses, signals were low pass-filtered with 0.5 Hz cut-off frequency. Then signals were segmented into two groups: the response state and the relax sate. The relax state indicates 20 seconds of signals before chemical stimuli, and the response state indicates of 20 seconds of signals after stimuli. The recorded signals are presented on Fig. 5A. To remove the differences of sensitivity on each channel, the signals were normalized by fixing their variances to 1. Black vertical lines indicate the temporal location of stimuli. When the stimuli is given on time point t, each of the three lines are located on t-20, t and t+20. So, the time periods between the first and second line belong to the relax state, and periods between the second and third lines belong to the response state. As shown in Fig. 5A, hemodynamic responses for chemical stimuli were observed as a gradual increase of HbO₂ levels and a decreasing of Hbr levels.

Using these responses, we tried to detect an exact time point of the odor stimuli (t). To increase sensitivity for the detection, we applied Principal Component Analysis (PCA). PCA is a widely used mathematical procedure to find orthogonal basis and it is useful for explaining general shape of data distribution. The method can also be used for extracting linear filters to maximize sensitivity for specific responses. We concatenated all signals in the relax and response states, then found the most principal component, as follows,

when X_relax, X_res are NIRS signals for relax and response states. A useful characteristic of this component is that it can make distinctive alterations from the stimuli. The filtered signal obtained from w^TX is shown in Fig. 5B. By this process, the signals of 16 channels were transformed into a single channel signal. We could see that the obtained signal shows drastic peaks near the stimuli. However, gradual alteration of the signal's baseline acts as an obstacle for extracting the exact time point of stimuli. To remove this drift, we applied an additional low-pass filter to the signal as shown in Fig. 5C. The time ranges of response states, which we tried to detect, are presented as the red box above the plot. Then by using mean values of relaxes and response states, we calculated the threshold for classification. The threshold is drawn as horizontal dotted line in Fig. 5C. In conclusion, we could successfully detect time periods for response states as shown in Fig. 5D.

From the previous analysis, we could extract temporal locations of the stimuli. Our next object was classifying the kinds of odor stimuli. For the classification, we analyzed the spatial patterns of hemodynamic responses. From each trial (a single odor stimuli), we calculated each channel's mean values for relaxes and response states, and measured their differences. So, from a single trial, we obtained a 16-channel data vector. Odor stimuli were repeated 20 times for each kind of odor, and three kinds of odors (HEP, IB and Blank) were continuously given to the rat, as the sequence of Blank, HEP and IB. Then 5 sessions of the above experiment were performed with 5 individual rats. As a result, we collected 300 trials of the difference data vector (20 trials * 3 classes * 5 sessions) and made the difference data matrix X∈ℝ^16×300. To fulfill our object, the classification of odorants, we re-segmented X into X_blank, X_HEP and X_IB, and compared each of them.

Fig. 6 shows the analysis results for the experiment of 4th session of the experiment. From X, we extracted only the data vectors belonged to session 4, and put that data as matrix X₍₄₎∈ℝ^16×60. Then X₍₄₎ segmented to X_blank(4), X_HEP(4) and X_IB(4).

Fig. 6A shows a comparison results between X_HEP(4) and X_blank(4). At first, we calculated mean values X_HEP(4) and X_blank(4), then showed their differences on the table on the right side. As already explained, hemodynamic responses for olfactory stimuli are observed as increased HbO₂ levels with decreased of Hbr levels. Results of the table fit well with the phenomena. The results also presented as topoplots on the left side. The plots show that the differences of channel 1A and 6F are more significant than other channels.

To investigate the possibility of the classification, we performed a t-test between X_HEP(4) and X_blank(4), then drew topoplots of the bottom row based on its results. The p-values obtained from two-sample t-test are finer measures than simple difference to prove discriminativeness between groups. It considers not only differences of means, but also variances, which are closely related to overlapping between groups. On the t-test, smaller p-values mean that the difference is relatively larger. So, to intensify the regions with smaller p-values, we applied a -log function on p-values and drew the topoplots by the values. Unlike upper-row mean value-based plots, 2B, 7G of HbO₂ levels and 4D of Hbr levels showed more significant differences on p-value based plots.

The same kind of analysis performed between X_IB(4) and X_blank(4) as shown in Fig. 6B. Trends of HbO₂ and Hbr levels are observed quite often with Fig. 6A, as shown on the table and upper plots. Furthermore, HEP and IB shares similar spatial patterns. Mean value-based plots also indicates 1A and 6F as discriminative channels like Fig. 6A and p-value based plots point to 2B, 7G (in HbO₂) and 4D (in Hbr) as informative channels. Although 1A and 6F more significantly emphasized on Hbr levels of p-value analysis, general patterns showed similar shapes. However, magnitudes of alteration showed clear difference between HEP and IB. The differences caused by IB were smaller than the results of HEP. The difference caused by HEP stimuli caused HbO₂ levels to vary from 0.445 to 0.619 (mean value=0.538) and -0.713 to -0.401 (mean value=-0.576) on Hbr levels, but IB showed value range of 0.162~0.265 (mean value=0.211) and -0.306~-0.182 (mean value=-0.247). Because of these small differences, p-value also showed a higher value range of 10^-7~10^-3.5.

From the above two analyses, we observed that the hemodynamic alteration caused by HEP and IB share similar patterns but their magnitude is different. In the third step of analysis, we compared X_HEP(4) and X_IB(4) as shown in Fig. 6C. This analysis is more important in practical terms than the previous two analyses, because it is directly related to the classification of odors.

The results in the table were calculated by mean (X_HEP(4))-mean (X_IB(4)). Because of X_HEP(4) has larger value ranges than X_IB(4), values for HbO₂ levels have a positive range and values for Hbr have a negative range. p-value based plots extract 1A, 4D and 7G (on HbO₂) and 4D and 8H (on Hbr) as the discriminative channels. These results are slightly different from the previous two comparisons. This difference implies that even though the direction of hemodynamic alteration by HEP and IB is similar, there is a subtle difference between them, and it could be beneficial for following classifications.

We extended the above analysis to the entire data set X from 5 sessions as shown in Fig. 7. The results were not much different to those in Fig. 6. The stimuli increased HbO₂ levels and decreased Hbr levels. HEP caused more significant response than IB. Mean value-based topoplots showed similar results to Fig. 6, but p-value based topoplots showed slightly different results because data from different rats were mixed.

By the above analysis, we confirmed that the classification of each odor stimuli could be possible. As the final step of the analysis, we tried classification of signals. At first, we attempt the classification between relax signals (X_blank) and response signals (X_HEP, X_IB). For the classification, we used the linear-kernel based SVM (Support Vector Machine) method. To evaluate classification performances, we performed 5-fold cross validation test. We calculated the test for each session, and the results are presented in Table 1. The performance fluctuated sharply between sessions from 64.8% (session 5) to 88.9% (session 4).

Secondly, we performed classification between odorant responses for two different stimuli. As already explained in the previous subsection, the difference between X_HEP and X_IB is smaller than the difference between relax and response signals. To overcome this lesser discriminability, we applied RBF (radial basis function) kernel-based SVM for the classification.

For the classification of each session, we tried both linear kernel and RBF kernel-based classification approaches, and we show best results in Table 2. In the first session and the third session, RBF kernel was used and in other sessions, linear kernel was used. The mean accuracy of the classifying stimuli was 77.9%, while the mean accuracy of odor classifications was 75.0%. 77.9% and 75.0% may not be satisfactory scores from the viewpoint of interface research. However, we confirmed that the magnitude of hemodynamic response from two different odor stimuli is significantly different, and sometimes it enables plausible accuracy of the classification as shown in the result of the session 4. We also observed subtle differences of spatial patterns between two responses on the t-test. These aspects of spatial patterns differences imply that improved measuring device/techniques and sophisticated classification algorithms can enhance the performance of the BMI system.