Investigation of Involvement between Specific Brain Regions and Evaluation Criteria Elements in a Two-Selections Task

It is essential to understand the neural mechanisms underlying human decision-making. Several studies using traditional analysis have attempted to explain the neural mechanisms associated with decision-making based on abstract rewards. However, brain-decoding research that utilizes the multivoxel pattern analysis (MVPA) method, especially research focusing on decision-making, remains limited. In brain analysis, decoding strategies for multivoxels are required for various decision-making evaluation criteria. This is because in daily life, the human decision-making process makes use of many evaluation criteria. In the present study, we investigated the representation of evaluation criterion categories in a decision-making process using functional magnetic resonance imaging and MVPA. Participants performed a decision-making task that involved choosing a smartphone by referring to four types of evaluation criteria. The regions of interest (ROIs) were the ventromedial prefrontal cortex (vmPFC), nucleus accumbens (NAcc), and insula. Each combination of the four evaluation criteria was analyzed based on a binary classification using MVPA. From the binary classification accuracy obtained from MVPA, the regions that reflected differences in the evaluation criteria among the ROIs were evaluated. The results of the binary classification in the vmPFC and NAcc indicated that these regions can express evaluation criteria in decision-making processes.


Introduction
Human decision-making, both simple and complex, occurs hundreds to thousands of times per day. A few examples may include decisions such as "What to eat," "Which clothes to wear," and "How to solve a problem." Among these scenarios, purchase decision-making is a familiar and frequently encountered decision. When buying a product, especially a more expensive one, people tend to compare information to aid in the decision-making process. In recent years, online shopping has given people the opportunity to purchase a greater variety of products. However, due to the large variety of product lineups, it is difcult to make the best decision that satisfes consumer requirements. Furthermore, consumers cannot check the actual products; they can only view the product information, images, and reviews displayed on the screen. Terefore, decision-making regarding the purchasing of products should be supported. An important step in this direction would be to elucidate brain representation to gain a better understanding of the human decision-making process.
Te feld of neuro-marketing, which explores the human decision-making process by objectively measuring brain reactions and taking advantage of brain science and psychophysics knowledge, is attracting increasing attention [1,2]. Te neuroscience feld uses methods such as functional magnetic resonance imaging (fMRI), functional nearinfrared spectroscopy, positron emission tomography, and electroencephalography to explore human decision-making processes [3,4]. Several studies using traditional univariate analysis have attempted to elucidate the neural mechanisms associated with decision-making based on abstract rewards.
One area that has been repeatedly shown to be activated by diverse rewarding stimuli is the ventromedial prefrontal cortex (vmPFC). In several neuroimaging studies, this brain region is active for a variety of primary and abstract rewards, including sports cars [5], cola preferences [6], pleasant odors [7], wine prices [8], facial attractiveness [9], and money [10]. Te results of these past studies support the idea that the vmPFC is involved in converting the values of diverse stimuli into a common measure for a behavioral choice. Other fndings have suggested that product preference activates the nucleus accumbens (NAcc). Knutson et al. [11] reported that preference elicits NAcc activation before a purchasing decision, whereas excessive prices can elicit insula activation and mesial prefrontal cortex (MPFC) deactivation.
Multivoxel pattern analysis (MVPA) has been attracting increasing attention, as shown in recent fMRI studies to elucidate brain activity patterns [12][13][14]. MVPA considers external stimuli, motion state, and mental content to be encoded in brain activity patterns. MVPA can save and distinguish spatial response patterns lost by averaging the responses across voxels in the region of interest (ROI), as in univariate analysis. When detecting the presence of a specifc cognitive condition in the brain, the main advantage of MVPA is its increased sensitivity. Te conventional fMRI analysis attempts to fnd voxels that exhibit statistically signifcant responses to experimental conditions. To increase the sensitivity to certain conditions, these methods spatially average the voxels that respond signifcantly to those conditions. Although this technique reduces noise, it also reduces the signal in two important ways. First, a voxel with a nonsignifcant response to a particular condition can carry some information about the presence/absence of that state. Second, spatial averaging can blur the spatial pattern that distinguishes experimental conditions. As with conventional methods, the MVPA approach also attempts to increase sensitivity by looking at the contributions of multiple voxels. However, to avoid the signal loss problem mentioned above, MVPA does not routinely involve a uniform spatial average of the voxel response. Instead, it uses a weighted average of responses that treats each voxel as a separate source of information about the participant's cognitive state. Tis technique optimizes these weights and aggregates this information among voxels and fnds ways to guide participants more accurately in terms of what they are thinking [15]. With this method, by analyzing a pattern composed of multiple voxels, it was possible to explore the brain expression of detailed information that could not be examined by conventional fMRI analysis. Te traditional univariate approach focuses on the activity changes of each voxel. By contrast, MVPA extracts information from many brain locations (voxels) at the same time to examine the spatial brain activation pattern. MVPA is often used for neural decoding. Neural decoding is a technique to estimate stimuli, behaviors, and cognitive states. Several neural decoding studies are being conducted to reconstruct visual information, cognitive judgments, and emotions [16][17][18]. Machine learning methods such as the support vector machine and neural network are important for this approach. Neural decoding is realized by learning brain activity patterns, which are multidimensional variables, using a machine learning algorithm and outputting prediction values from new brain activity patterns using the learned model. Tere are many studies on biometric data classifcation using machine learning other than neural decoding. In the medical feld, research is being conducted to detect lesions and classify benign/malignant tumors, and machine learning methods are used [19,20].
A relatively large amount of the existing research using MVPA focuses on decision-making because it is important to understand which brain regions are involved in the various metrics used during decision-making in daily life. Given this background, the present study aimed to estimate user choice in decision-making based on brain activity. To achieve this aim, we experimented to verify the brain regions involved in the evaluation criteria in decision-making processes and investigated the representation of their categories using fMRI and MVPA. Assuming a situation involving the purchase of smartphones through online shopping, participants picked one from a choice of two products with information for a single evaluation criterion. Tis task involved four types of evaluation criteria for decision-making. We focused on the voxel pattern in ROIs based on diferences in the type of evaluation criteria. To our knowledge, brain regions refecting the type of evaluation criteria used in decision-making have not been reported.
In our previous experiments, we evaluated the statistical signifcance for the whole brain and found that the brain regions in which each evaluation criterion was specifcally involved were not observed for all evaluation criteria. Moreover, no signifcant diference in brain activity was observed for each evaluation criterion in the ROI [21]. We hypothesized that the type of evaluation criteria is specifc to or commonly involved in certain brain regions. Tis analysis focused on the vmPFC, insula, and NAcc, which are considered to be involved in decision-making and were used as ROIs [22][23][24][25][26]. We hypothesized that there would be a difference in the activation patterns of the vmPFC, insula, and NAcc depending on the type of evaluation criteria. To examine this hypothesis, a decision-making task regarding each evaluation criterion was performed. Te diferences in brain activation patterns due to the diferences in evaluation criteria in these ROIs were then examined by comparing each combination of evaluation criteria. For the analysis, we used MVPA, which has been established as an efective method for identifying and classifying brain activity patterns. Based on diferent evaluation criteria, the voxel patterns from all associated regions were examined using MVPA. Each combination of the four evaluation criteria was analyzed after binary classifcation by MVPA. Finally, based on the binary classifcation accuracy obtained from MVPA, the regions among the ROIs that refected diferences in evaluation criteria were evaluated.
For many people, evaluation criteria such as price, color, and production date are important factors in making a better purchase. Te results obtained in this experiment indicate that vmPFC and NAcc respond concerning the same endpoints when making a choice. Moreover, as shown in [8,11,22,23,25], the vmPFC and NAcc are sites involved in decision-making and execution. On the contrary, there are neuromarketing methods that take advantage of fndings from brain science and apply them to marketing activities [27][28][29]. It is important to analyze the consumer psychology and behavioral principles indicated by brain science. Te traditional methods of behavior analysis were questionnaires or interviews and thus could not elicit consumers' unconscious true intentions. However, the results of this research proved that the vmPFC and NAcc measurements can be used as a model to visualize and quantitatively evaluate consumers' unconscious psychology and preferences, which are difcult to verbalize.

Materials and Methods
Te materials and methods used to in this article are described in the following sections.

Participants.
Twenty-fve participants (fve females: two left-handed, mean age 20.60 years, standard deviation 1.26 years, and age range 19-22 years) participated in the fMRI experiment. One participant who did not complete the experiment was excluded. Terefore, 24 participants were fnally included in the data analysis. Tis study was approved by the Research Ethics Review Committee of the Kochi University of Technology (approval no. 52-C3). All participants provided written informed consent before the experiment began.

Task and Stimuli.
All participants performed a decisionmaking task that involved choosing a smartphone by referring to each evaluation criterion. Figure 1 shows the experimental timeline. Tis experiment was based on a block design. It consists of several discrete periods of on-of blocks, with the "on" representing a task condition and the "of" referring to a rest state or a diferent task condition.
In this task, participants viewed a screen, an example of which is illustrated in the lower part of Figure 1. Te screen presented two identical smartphone images and the letters of each diferent combination of evaluation criteria as stimuli. Tis experimental design using these stimuli, which are illustrated in two identical images, each with diferent information, has also been adopted in other decision-making studies [30]. Te participants pushed the left or right button to select the smartphone they wanted more.
Te experimental conditions included the price, body color, and production year as the evaluation criteria. Te body color evaluation was conveyed through textual information instead of an illustration. Te reason for this was to avoid the possibility of a diference in willingness to purchase depending on the displayed color of the product [31]. Tese choices were selected as evaluation criteria because they can be evaluated easily regardless of the presence or absence of smartphone knowledge among the participants. In addition, a dummy variable (four squares as meaningless symbols) was set as the control condition. Te brain activity at the time of the main task was considered to include three main efects: decision-making, visual recognition of the stimuli, and button-pressing at the time of selection. In addition, brain activity during dummy tasks is thought to refect the efects of almost the same condition as the main task, except for decision-making. Brain activity regarding the diference between the main and dummy tasks is considered to represent only the efects of decisionmaking. Each criterion had four types of content. Table 1 shows the list of evaluation criterion labels used in this experiment. Each participant was considered to have a diferent priority for each evaluation criterion. In this experiment, it was hypothesized that a specifc brain-related region would refer to the impression for each evaluation criterion that included priority diferences. Te evaluation criteria appeared in a diferent order for each participant. Each participant performed two runs under the same conditions. Each run contained eight blocks of four separate tasks: the price choice, the color choice, the year choice, and the dummy choice. During each choice, a screen showed information for 3 seconds, followed by a rest period for 2-4 seconds. Te screen presented two images of the smartphone, shown on the left and right of the screen. Diferent labels under the images represent each evaluation criterion. In the choice tasks, the participants decided on an object in their mind and then pushed the left or right button to select the object they had chosen. In the dummy tasks, the participants were required to push either button intuitively. Te total time of one run was 306 seconds. Stimuli images and words were rear-projected onto a screen placed in the scanner bore using an LCD projector. Te screen showed two identical smartphone images and two diferent labels as an evaluation criterion, as demonstrated in Figure 1.

MRI Acquisition and Data Preprocessing.
Scanning was performed on a 3.0-tesla scanner (MAGNETOM Verio, Siemens Healthinners, Erlangen, Germany) using a 16channel head coil at the Kochi University of Technology. Functional scans were acquired with a standard gradientecho echo-planar imaging sequence to cover the whole brain (feld of view � 192 mm 2 ; repetition time � 3,000 ms; echo time � 30 ms; fip angle � 90°; slice thickness � 3.0 mm; voxel size � 3.0 mm 3 ). Each run of the functional scans obtained 102 volumes over a total duration of 306 seconds. A highresolution T1-weighted anatomical scan was acquired for each subject (1.0 mm 3 resolution).
Te frst two scans (6 seconds) in each run were discarded to account for any instability with the fMRI scanner. SPM12 software (Wellcome Centre for Human Neuroimaging, London, UK) was used to process and analyze the functional data. Functional images were corrected for differences in slice acquisition time and motion. Te data were then realigned and normalized to the Montreal Neurological Institute (MNI) standard brain model. Te brain activation degrees were analyzed on the MNI coordinates.

fMRI Analyses.
Te following four conditions were modeled: price choice, color choice, year choice, and dummy choice. Common or specifc brain regions were involved in each condition, and these regions were identifed by creating contrasts. With the frst level (single subject analysis), Computational Intelligence and Neuroscience contrasts (price vs. dummy, color vs. dummy, and year vs. dummy) were created to identify brain regions, which were commonly activated for all contrasts. Price choice vs. the two other choices, color choice vs. the two other choices, and year choice vs. the two other choices were created to identify brain regions, which were specifcally activated for each contrast. With the second level (group analysis), one sample t tests were performed to examine signifcant brain activation among the group during the contrasts mentioned. A statistical parametric map was generated using the price, color, and year vs. dummy choice contrasts. Clusters of voxels were corrected for multiple comparisons across the whole brain using family-wise error correction and a threshold of p values: p < 0.05 [32]. Te statistical parametric maps were generated using color vs. (price and year) choice contrast. Clusters were defned using a height threshold of p < 0.001 uncorrected for multiple comparisons with a cluster size threshold of k � 171 voxels. In the contrasts of price vs. (color and year) and year vs. (price and color), no suprathreshold clusters were applied.

Multivoxel Pattern Analysis.
Te results of the fMRI analysis described in the previous section showed activation in the vmPFC during decision-making about each evaluation criterion. Brain activation patterns in the decisionmaking task were investigated using MVPA based on each evaluation criterion.
MVPA was performed using a support vector machine (SVM) with a linear kernel [33], as implemented in the Pattern Recognition for Neuroimaging Toolbox [26]. Te pattern analyses were performed separately for each participant. Te β value at decision-making based on each evaluation criterion obtained from the general linear model in the previous section was taken as the input value. Tere were 48 β belonging to four evaluation criteria, including the dummy, from two runs for each participant (24/run). A binary classifcation according to the four evaluation criteria was carried out, with each beta representing a single decision-making event based on an evaluation criterion. As the evaluation criterion for each β is already known, in MVPA, these beta values were labeled "price," "color," "year," and "dummy." Te beta values, which were labeled as two types of evaluation criteria, were input to the SVM as training sets to generate a boundary of two classes. We examined into which class the sample data with either label were classifed.
Te analysis was performed using voxels in only the vmPFC, insula, or NAcc by masking using the PickAtlas toolbox. In this analysis, binary classifcation by SVM was performed to test whether every ROI could distinguish between the four evaluation criteria (e.g., "price or color" and "color or year"). WFU PickAtlas [34,35] was used to create the three ROI masks for the vmPFC, insula, and NAcc. Te mask for vmPFC selected the Brodmann areas 10, 14, 25, and 32 as defned by Finger [36]. Data were crossvalidated using the leave one block out method, with two sets of 24 data points from each participant. Only one sample data point was extracted from all the data as a test set. Te remaining data points were used as a training set. Te verifcation was repeated using all the sample data as a test set. Te number of correct answers among the 48 data points by each participant (e.g., when price data were classifed as price) was obtained from the SVM. Next, the average correct answer rate for each binary classifcation obtained from all participants was calculated. Te average correct answer rates were evaluated to reveal whether the brain activity pattern of each ROI expressed the evaluation criteria for decisionmaking.

Results
We performed a whole brain analysis to identify regions that had signifcant activation for each choice.  Table 3. In these fgures, the color bars represent the average value of all 24 participants, and the error bars represent standard errors. Te x-axis shows a combination of each binary classifcation (e.g., p-c shows the result of binary classifcation by price and color). In the precision using the vmPFC as an ROI, the highest average accuracy was 68.40% in the dummy-color binary classifcation, and the lowest was 58.51% in the price-color binary classifcation. Regarding the result of the insula as ROI, the highest average accuracy was 58.33% in the color-price binary classifcation and the lowest was 52.08% in the pricedummy binary classifcation. Regarding the result of the NAcc as an ROI, the highest average accuracy was 63.37% in the dummy-color binary classifcation and the lowest was 58.51% in the price-dummy binary classifcation.

Discussion
Te present study investigated the neural substrates associated with assessments of diferent criteria in decisionmaking. To this end, the study participants were presented with a pair of alternatives belonging to a single evaluation criterion and asked which product they wanted to choose based on the evaluation criteria. We evaluated whether decision-making based on diferent evaluation criteria could be discriminated based on the spatial activity pattern in diferent brain regions.
Considering the infuence of the context of the vmPFC, insula, and NAcc activities observed in past studies, the ROIs of those areas were used. Te vmPFC is active for a variety of primary and abstract rewards in several neuroimaging studies [22][23][24][25]. Tese fndings suggest that the vmPFC plays several roles in the representation of complex choices, which suggests that the activated region difers depending on the evaluation criteria when based on an alternative value or preference by each individual. Te insula has shown the possibility of triggering activation for the price during a purchasing decision and the NAcc of inducing activation for individual preferences [26]. Tese fndings suggest that the brain frames a preference as a potential beneft and price as a potential cost, thereby lending credence to the notion that consumer purchasing refects an anticipatory combination of preference and price considerations.
Although few studies have investigated the infuence of diferences in evaluation criteria on the brain, numerous studies have examined the infuence of preferences and pleasure/discomfort when participants evaluate two or more stimuli as a decision-making task on the brain. Tese experimental designs are roughly divided into two types. Te frst is a design that presents a single stimulus in order, a design that the subject subsequently carries out in a single evaluation. Te second presents a pair of stimuli at the same time, and the subjects select what they prefer more. In the present experiment, the latter design was adopted, in which the subjects evaluated a pair of alternatives belonging to a certain evaluation standard through comparisons. Numerous studies on decision-making have presented a pair of images or character strings and performed evaluations, while others have used an experimental design that presents the same image and character strings of diferent contents, such as the present experimental design. Some studies have also reported results regarding activity in the vmPFC, insula, and NAcc [22,23,25].
As for the average accuracy of binary classifcation by MVPA, the present results using vmPFC, insula, and NAcc as ROIs exceeded the chance level. In the insula, even though the chance level exceeded the average precision, the standard error of the binary classifcation accuracy of several combinations was lower than the chance level. Te average classifcation accuracy in the insula was lower than those in the vmPFC and NAcc for all combinations of evaluation criteria. Te average classifcation accuracy in the NAcc was about 60%, and no diference in classifcation accuracy was seen for any evaluation criterion. On the contrary, vmPFC showed a diference in the binary classifcation accuracy for each combination of evaluation criteria. In the binary classifcation of price and year, the year was 2.26% more accurate. Te color was 2.95% more accurate than the price and 6.60% more accurate than the year. Although none of the diferences in average accuracy were signifcant, classifcation in the vmPFC showed the possibility of being most classifable when color was used as the evaluation criterion. Tese results suggest that brain regions involved in the decision and the preference, such as vmPFC and NAcc, represent diferences in brain activity patterns in comparative decision-making.
Several studies applying MVPA to brain activity during decision-making have reported the following results: Jai et al. conducted an evaluation and bidding task for apparel products under three visual presentation conditions (static picture, zooming, or model rotation videos) and analyzed the brain patterns in "a buy decision" and "a not buy decision" [38]. According to the whole-brain classifcation analysis, the classifer accuracy rates were 95% in the zooming and rotation conditions, while the static condition had a 75%. By the searchlight classifcation, there were some ROIs exclusively referred by certain presentations. Tese results suggest that the brain regions to which various information corresponds exhibit characteristic activity patterns and that appropriate recognition of these patterns enables classifcation. On the contrary, Kim et al. investigated the relationship between purchase intention and Computational Intelligence and Neuroscience perceived garment ft when purchasing decision-making [39]. Tey analyzed the brain activity during the task including the phase of evaluating the ft of the model wearing clothes and the phase of purchasing consideration with the price displayed on the clothes. As a result of the MVPA for whole brain searchlight, the classifcation accuracies were more than 50% of the chance level at 11 brain regions. Among them, the superior parietal lobule exceeded 80% with the highest accuracy. One of the reasons for this accuracy higher than our results may be that their tasks Note: region labels were named based on the automated anatomical labeling template [37], which is a digital human brain Atlas with a labeled volume. Te labels indicate macroscopic brain structures. Cluster size is reported in voxels. Te T-statistic value is the total average, which was calculated for each voxel from MRI data acquired for each subject and divided by the standard deviation of all subjects.   Computational Intelligence and Neuroscience included multiple evaluation criteria such as garment ft and price. Our results showed that decision-making based on individual criteria is represented in brain patterns, but we believe that decision-making based on multiple criteria may represent more characteristic brain patterns. Tis study focused on customer preferences based on individual evaluation criteria. Although the results of the classifer did not show activity patterns in specifc brain regions that depended on the evaluation criteria, they indicated the possibility that preference-based decisionmaking could be represented in brain patterns. It is necessary to process various factors such as reward, risk, and strategy in decision-making. In particular, regarding purchasing decisions, it is said that the factor of individual preferences produces better decisions [40][41][42][43]. Te development of this research is expected to help elucidate the neural basis of decision-making and to contribute to selection, online shopping, or marketing strategies.

Conclusions
Tis study aimed to estimate user choice in decision-making based on brain activity. Te experiment focused on evaluation criteria, which motivate decision-making, and the verifcation of brain regions involved in these evaluation criteria. To achieve our purpose, we investigated the representation of evaluation criterion categories in decisionmaking using fMRI and MVPA. Price, color, and year were used as the evaluation criteria. We focused on the vmPFC, NAcc, and insula as ROIs. Each combination of the four evaluation criteria was analyzed into a binary classifcation by MVPA. From the binary classifcation accuracy obtained from MVPA, we evaluated the regions that refected differences in evaluation criteria among the ROIs. From the results of the binary classifcation by MVPA, the vmPFC and NAcc showed that these regions were capable of expressing the infuence of the evaluation criteria during decisionmaking.
In this research, we applied classifcation analysis from brain activity by adopting price, color, and year as evaluation criteria because they are often involved in purchasing decisions for various products. In actual purchase decision-making, consumers also refer to various evaluation criteria. In a future study, it will be necessary to investigate brain activity patterns expressed by evaluation criteria other than those adopted in the present study and to verify whether they can be classifed similarly. Te present analysis involved only binary classifcation by two classes. As a future challenge, the accuracy of multiclass classifcation will need to be verifed.

Data Availability
Te data used to support the fndings of the study can be obtained from the corresponding author upon request.

Conflicts of Interest
Te authors declare that they have no conficts of interest.