Machine learning‐based classifying of risk‐takers and risk‐aversive individuals using resting‐state EEG data: A pilot feasibility study

Abstract Background Decision‐making is vital in interpersonal interactions and a country's economic and political conditions. People, especially managers, have to make decisions in different risky situations. There has been a growing interest in identifying managers’ personality traits (i.e., risk‐taking or risk‐averse) in recent years. Although there are findings of signal decision‐making and brain activity, the implementation of an intelligent brain‐based technique to predict risk‐averse and risk‐taking managers is still in doubt. Methods This study proposes an electroencephalogram (EEG)‐based intelligent system to distinguish risk‐taking managers from risk‐averse ones by recording the EEG signals from 30 managers. In particular, wavelet transform, a time‐frequency domain analysis method, was used on resting‐state EEG data to extract statistical features. Then, a two‐step statistical wrapper algorithm was used to select the appropriate features. The support vector machine classifier, a supervised learning method, was used to classify two groups of managers using chosen features. Results Intersubject predictive performance could classify two groups of managers with 74.42% accuracy, 76.16% sensitivity, 72.32% specificity, and 75% F1‐measure, indicating that machine learning (ML) models can distinguish between risk‐taking and risk‐averse managers using the features extracted from the alpha frequency band in 10 s analysis window size. Conclusions The findings of this study demonstrate the potential of using intelligent (ML‐based) systems in distinguish between risk‐taking and risk‐averse managers using biological signals.


INTRODUCTION
Decision-making is a complex cognitive process and an essential skill in everyday life (Poudel et al., 2020;Si et al., 2020;Wojcik et al., 2019).
It becomes imperative when individuals manage an organization or country, and the outcome of the decision can be decisive in economic, cultural, and political contexts. Decision-making has been part of studies of neuroscience, neuroeconomics, and related disciplines for many decades (Krain et al., 2006;Opris et al., 2020;Srivastava et al., 2020;Viacava et al., 2016;Zhang, 2018). Understanding the neural and cognitive foundations of the decision-making process is an important issue, especially in decisions made by managers. This information can play a vital role in selecting managers for an organization or even a country.
One of the reasons that different decisions are made under the same conditions is heterogeneity in brain physiology. In this regard, the functional characteristics of the brain have been evaluated by examining the brain's activity at the resting-state (Studer et al., 2013). The restingstate functional connectivity, representing intrinsic brain activity, has been used to examine individuals' impulsive economic decision-making (Li et al., 2013). Along with this idea, electroencephalography (EEG) is a powerful non-invasive tool for investigating the brain bases of human psychological processes . The method has been applied in various decision-making research domains (Ivaskevych, 2019;Lee et al., 2017;Pornpattananangkul et al., 2019;Ramsøy et al., 2018;Si et al., 2020;Wilson & Vassileva, 2018;Zheng et al., 2020). The use of intelligent systems based on learning from EEG signals has also been considered by most researchers (Al-Nafjan et al., 2017;Anjum et al., 2020;Ieracitano et al., 2020;Maitín et al., 2020;Noor & Ibrahim, 2020;Rasheed et al., 2020;Roy et al., 2019;Tzimourta et al., 2021).
Hence, developing intelligent EEG-based systems to evaluate human decision-making skills based on brain signals has become a challenging and demanding research area. For instance, Si et al. (2020) proposed an EEG-based machine learning (ML) model to predict individuals' responses. The authors could achieve higher accuracy by extracting discriminative spatial network pattern (DSNP) features from singletrial EEG data and the Linear Discriminate Analysis (LDA)-based model.
Using two EEG datasets, they obtained predictive performance in the first and the second sets with 88% and 90% accuracies, respectively, in individual response detection. In a similar attempt, Wojcik et al. The proposed ML models (Si et al., 2020;Wojcik et al., 2019) focus only on classifying individual responses, which are ML models of an intra-subject classification. Due to the importance of decision-making by managers, the automatic cross-subject separation can also divide managers into two groups, risk-taking and risk-averse, based on the features extracted from the EEG signal. This study proposes a model of ML to achieve this goal. Resting-state EEG data were recorded from managers, and the ML model was trained using information extracted from the collected signal of managers. Finally, using brain signals, this trained model can separate risk-taking and risk-averse managers.

Participants
In this study, we recruited participants from a population of 173 managers who are responsible for formulating their organization's strategies and goals, and regularly encounter unplanned decisions. As gender differences can also affect decision-making (Lin et al., 2019;Mahaldar & Aditya, 2017;Mehta, 2020), all participants were selected from the same gender in this study. The participants completed a 13-item risk-tolerance scale questionnaire to assess their willingness to engage in risky financial behavior (Grable, 1999;Grable & Lytton, 2003). Based on the questionnaire scores, managers were categorized as risk-takers (scores > 32) or risk-averse (scores < 22) in the economic decision-making process. In addition, they were asked to evaluate themselves as risk-takers or risk-averse. Among them, the thirty healthy men (N = 30; mean age = 40.80; age range = 32-55 years) were selected for EEG data acquisition in this study. After ensuring the use of no drug or alcohol by participants on the day of data collection, the experimental procedures were explained orally. All participants also approved consent forms. The experimental study was approved by the Ethics Committee of the University of Tabriz (Tabriz, Iran) and was adapted to the regulations of the Declaration of Helsinki.

EEG recording and experimental procedures
EEG data were recorded using an ANT Neuro system (DC-Amplifier, ANT Neuro, the Netherlands) with 64 scalp Ag/AgCl electrodes (waveguard cap, ANT Neuro, the Netherlands) positioned according to the 10/20 system. Cortical data were acquired at 250 Hz, and the impedance of all electrodes was maintained below 5 kΩ during data recording. The electrode AFz was also considered the ground and with the reference at the right earlobe (A2). EEG signals were collected in a room with sufficient light. Participants were asked to sit quietly in a chair while collecting data and refrain from moving their heads frequently. EEG data were collected at resting state, whereas both eyes were closed for 5 min.

EEG preprocessing and analysis strategy
All data were processed in MATLAB R2014b. The EEG data were preprocessed using an EEGLAB open-source toolbox (Delorme & Makeig, 2004), made freely available by the Schwartz Center for Computational Neuroscience. The Butterworth band-pass filter was applied to the EEG signal to filter the signal in the 0.1-85 Hz frequency range to eliminate high-frequency noise. The filter also removed the 50-Hz noise caused by the power line. A common average reference filter was applied to control the problems related to the signal-to-noise ratio F I G U R E 1 Flowchart of the electroencephalography (EEG) analysis strategy. (Ludwig et al., 2009). In addition, the independent component analysis was used to manually separate and remove artifactual components from EEG signals.
The current study used ML analysis to detect risk-taking and riskaverse managers. After the preprocessing of the brain signal, the feature extraction, feature selection, and SVM classifier are considered the consecutive steps of the ML algorithm ( Figure 1). Raw EEG signals contain noise and redundant information that is not related to risktaking behavior. Therefore, to distinguish between the two groups, it is necessary to extract meaningful features from the signals. Feature extraction helps to reduce the amount of redundant data and allows the construction of a model with less machine effort, leading to faster learning and better generalization performance. By extracting relevant features, we can identify the essential characteristics of the EEG signals that differentiate risk-taking and risk-averse individuals. These features provide valuable insights into the underlying neural mechanisms of risk-taking behavior and can be used to improve the accuracy of classification models. Frontal region EEG signals have been used in the time-frequency domain for feature extraction (Bartra et al., 2013;Gianotti et al., 2009;Krain et al., 2006;Schutter & Van Honk, 2005;Si et al., 2019;Studer et al., 2013). The ML algorithm uses a two-step process to select the appropriate feature, including statistical analysis to remove bad features and the sequential floating forward selection (SFFS) algorithm (Pudil et al., 1994) to select the feature perfectly. In the ML model, the SVM classifier is used to classify two groups of participants based on the extracted features of the EEG signal. This study uses the K-fold cross-validation algorithm to evaluate the ML model and the training and testing process. More details of each step of the ML algorithm are described below.

Feature extraction and selection
The EEG data as a complex signal contains much information about brain neural patterns. The main purpose of extracting features from the brain signal is to extract meaningful information that can be used in the ML model to classify the two groups of managers. In the feature extraction step, cleaned 5-min EEG signals in consecutive 10-s time windows with 10% overlap have been transformed into feature vectors. Time-frequency analysis was used to extract the features to overcome the nonstationary nature of the EEG signal (Bajaj, 2021;Ieracitano et al., 2020;Jacob et al., 2018). This study used the discrete wavelet transform (DWT) with the mother wavelet dB4 and level 6 as a time-frequency analysis, as shown in Figure 2. Consequently, this facilitates distinguishing between the two groups of managers based on the cortical signal. As shown in Figure 1, the filter-based and wrapper-based techniques were used consecutively to select the features. Filter feature selection methods use statistical F I G U R E 2 Extracting electroencephalography (EEG) features using discrete wavelet transform (DWT) with the mother wavelet dB4 and level 6.
techniques to eliminate features that cannot distinguish between the two groups of managers outside of the predictive models. The t-test was used to compare the extracted features between groups. Then, the features that did not show a significant difference between the two groups were removed from the feature vector.
In the wrapper method, adding and removing predictors selected in the previous step tries to achieve an optimal combination of features to maximize the model's performance. In this selection method, the algorithm is not concerned with the feature types and selects a subset of features that show the best-performing model according to the performance metric. In this study, the Student t-test was used to remove inappropriate features and the SFFS algorithm (Pudil et al., 1994) as the wrapper method that dynamically increases and decreases the number of features to find an appropriate combination of remaining features.

Classification model
One of the most widespread algorithms that researchers have used for EEG classification is the SVM classifier algorithm. The SVM is a supervised ML model that can classify high-dimensional feature space based on the hyperplane (Shanir et al., 2018;Wang et al., 2022). This classification algorithm reduces the time required for the learning phase by transforming the prediction problem into an optimization problem and has better accuracy and speed than the other algorithms. In this study, the SVM classification algorithm was used to classify participants into two groups, risk-taking, and risk-averse managers, based on the information learned from the features extracted from resting-state EEG signals.
To evaluate the classification algorithm in a range of features, a fivefold cross-validation method was adopted in the proposed study. This evaluation method has better validity due to the limited number of subjects. As shown in Figure 1, the subjects are divided into five equal portions using random selection in the fivefold validation method. The Therefore, the evaluation metrics can be calculated as follows: Accuracy represents the overall effectiveness of the classifier.
Moreover, the question of how effectively the classifier identifies risktaking and risk-averse groups is expressed by sensitivity and specificity, respectively. A single metric's combination of precision and sensitivity is shown as the F1-measure.

Statistical methods
For statistical significance of the performance of classifier between size of analysis windows, we used the t-test with Bonferroni's multiple comparison test. p < .01 was considered statistically significant.

RESULTS
In this study, 30 male managers were included in the analysis. The  Based on resting-state EEG signals, managers were classified into two risk-taking and risk-averse groups using different extracted features, the feature selection method, and the SVM classifier. Statistical features are extracted in the time-frequency domain of brain signals.

F I G U R E 3
Descriptive statistics of behavioral data and the results of the t-test as the statistical analysis. "****" Indicates p < .00001.
The two-step feature selection method was applied to the extracted features to remove bad features and select good ones. The evaluation of classifier performance, which includes testing and training, is based on the fivefold cross-validation. Figure 4 shows the classification performance for each resting-state signal frequency band. Notably, the SVM classifier with RBF kernel in the alpha band performs better than the other frequency bands, whereas the results for the theta band are also very close to the alpha band values. The classifier could separate the two groups of managers with 74.42% accuracy, 76.16% sensitivity, 72.32% specificity, and 75% F1-measure using the features extracted from the alpha band in 10 s analysis window size.
Receiver operating characteristic is shown in Figure 5. As Figure 5 illustrates Furthermore, we performed analyses for both assumptions in different frequency bands ( Figure S5). As shown, there is no significant difference in the obtained F1-measures across the frequency bands.

DISCUSSION
This feasibility study aims to assess the effectiveness of using restingstate EEG signals to identify risk-taking and risk-averse managers via ML models. To achieve this goal, statistical features were extracted from the time-frequency domain of the resting-state EEG signals. Then, suitable features are selected for classification using a two-step algorithm. In the first step, a statistical analysis is used to remove bad features. The SFFS algorithm is used to select good features appropriate to the classification in the second step. For the SVM classifier, the best results were obtained in the alpha band with an accuracy of 74.42%, a sensitivity of 76.16%, and a specificity of 72.32% (Figure 4).
These results perform better than the SVM classifier values for punishment and reward classifications (58.5% accuracy) in the beta band (Wojcik et al., 2019).
Previous studies applying classic statistical methods to the resting state have suggested a relationship between the risky decision-making processes and brain signals (Massar et al., 2014;Nakao et al., 2013;Ramsøy et al., 2018;Studer et al., 2013;Yaple et al., 2018). For instance, Studer et al. (2013) found no significant correlation between risktaking behavior and overall bilateral prefrontal cortex (PFC) power or PFC asymmetry index in the alpha frequency band. However, these inconsistent findings highlight the need for further investigation. By focusing on the power of alpha frequency in 16 channels (with a significant difference in FPz), our study suggests that this frequency band may have a potential role in predicting risky behavior, as indicated by the features calculated based on this power.
Modern automatic methods, such as ML techniques, have been used to classify a variety of diseases and psychological disorders in recent years (Craik et al., 2019;Hosseinifard et al., 2013;Müller et al., 2008;Mumtaz et al., 2017). Although attempts have been made to implement ML models in predicting risky decision-making in recent years

F I G U R E 6
The accuracy and F1-measure of the support vector machine (SVM) classification model using the features extracted from each channel in five frequency bands with 10 s analysis time window.
F I G U R E 7 Accuracy and number of features (analysis time window = 10 s) during training and feature selection phases. (Si et al., 2020;Wojcik et al., 2019), these methods have not become common. Therefore, using ML models based on EEG signals to evaluate and diagnose risky decision-making is of great importance. This classification study considers two novel aspects: (1) classification of two groups of (risk-taking and risk-averse) managers and (2) classification/prediction at the intersubject (cross-subject) level.
Although much research has been done on decision-making and cortical activity, there is still little information about resting neuronal activity in risk-taking and risk-averse managers. Despite predicting individual responses in some previous studies (Si et al., 2020;Wojcik et al., 2019), the automatic separation of managers based on resting-state brain activity has not been studied hitherto, especially at the intersubject level. Si et al. (2020) proposed an ML model for predicting individuals' responses by extracting distinctive spatial network pattern features from single-experimental brain networks. The authors presented an LDA-based model that could use DSNP consistently to achieve higher accuracy than network properties. Seven different classifiers were also used to compare their efficiency in the reward/punishment characteristic cortical activity detection and the punishment and reward classification tasks. The proposed ML models focus only on classifying an individual response, an intra-subject classification, but this study trained an ML model using selected features that could distinguish risk-taking and risk-averse individuals.
This issue becomes important when these individuals are managers, which, to our knowledge, have not been studied so far (Wojcik et al., 2019).
Our findings reveal that risk-taking and risk-averse managers can be identified using signals recorded from the frontal lobe. These results confirm the previous findings that show the role of the PFC in executive control and maintaining goals in decision-making and the task of creating fundamental impulses related to self-interest, which is a consultative process in human decision-making (Henrich et al., 2001;Huerta & Kaas, 1990;Munakata et al., 2011).
Our findings also suggest that the accuracy of the classification model decreases as the number of features is increased in all frequency bands. These findings have important implications for the practical application of our approach, as it suggests that a simpler model with fewer features may be more effective in identifying risk-taking and risk-averse managers using resting-state EEG signals.
Nevertheless, our proposed methodology suffers from some limitations. In this study, EEG signals were recorded from 30 male managers.
In order to avoid the influence of deference gender on decision-making as well as the small number of female managers, only male participants were used in this study. This has made the ML model unable to learn the unique characteristics of risk-taking and risk-aversion. Risk-taking and risk-aversion are not limited to managers and all people may be risk-taking and risk-averse. However, our study was designed to investigate the neural mechanisms underlying risk aversion in a specific population (i.e., financial decision-making professionals).
Another limitation that we faced in this study was the lack of access to a large number of managers. Only 30 managers have been used in this market. Other limitations related to this study include lack of outof-sample testing. Despite these limitations, our study demonstrates the promising potential of using intelligent systems to select managers based on resting-state biological signals which avoid potential confounding effects of external stimuli.

CONCLUSIONS
This study provides compelling evidence for the potential of ML models to identify risk-taking and risk-averse managers based on resting-state EEG signals. It also highlights the importance of exploring the relationship between intrinsic neural properties and risky behavior, as the observed changes in brain signals during this paradigm are primarily driven by internal processes rather than external stimuli. This research has significant implications for future studies in the field of managerial decision-making, particularly in the development of more sophisticated and accurate predictive models. This feasibility study highlights the possible contribution of these techniques in assigning individuals to highly demanding situations. To achive this goal, future studies' design should consider groups including moderate levels of risky behavior and an independent sample of participants to assess the generalizability of findings. Progress in the research area might also benefit from using cognitive tasks including questions with real incentives to elicit risk preferences. We hope current study may provide valuable insights into the neural mechanisms underlying risk aversion and can guide future research in this area. The approach could be helpful, especially in the case of leading managers' decisions that could reasonably impact a country's economic and political conditions.

CONFLICT OF INTEREST STATEMENT
The authors have no conflicts of interest.

FUNDING INFORMATION
This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

DATA AVAILABILITY STATEMENT
The datasets generated during and/or analyzed during the current study are available from the corresponding author upon reasonable request.