Next Article in Journal
Research on Stochastic Evolutionary Game and Simulation of Carbon Emission Reduction Among Participants in Prefabricated Building Supply Chains
Previous Article in Journal
Innovative Stackable Multijet-Printed Templates for Precise Veneer Preparation: A Dental Technique
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Applied Sciences
Volume 15
Issue 9
10.3390/app15094980
applsci-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Using the β/α Ratio to Enhance Odor-Induced EEG Emotion Recognition

by
Jiayi Fang
1,
Genfa Yu
1,
Shengliang Liao
2,3,4,
Songxing Zhang
1,
Guangyong Zhu
1,* and
Fengping Yi
1,*
1
School of Perfume and Aroma Technology, Shanghai Institute of Technology, Shanghai 201418, China
2
East China Woody Fragrance and Flavor Engineering Research Center of National Forestry and Grassland Administration, Jiangxi Agricultural University, Nanchang 330045, China
3
Camphor Engineering Research Center of National Forestry and Grassland Administration, Jiangxi Agricultural University, Nanchang 330045, China
4
College of Forestry, Jiangxi Agricultural University, Nanchang 330045, China
*
Authors to whom correspondence should be addressed.
Appl. Sci. 2025, 15(9), 4980; https://doi.org/10.3390/app15094980
Submission received: 18 February 2025 / Revised: 25 April 2025 / Accepted: 29 April 2025 / Published: 30 April 2025
(This article belongs to the Section Biomedical Engineering)
Browse Figures
Versions Notes

Abstract

:
Emotion recognition using an odor-induced electroencephalogram (EEG) has broad applications in human-computer interaction. However, existing studies often rely on subjective self-reporting to label emotion, lacking objective verification. While the β/α ratio has been identified as a potential objective indicator of arousal in EEG spectral analysis, its value in emotion recognition remains underexplored. This study ensured the authenticity of emotions through self-reporting and EEG spectral analysis of 50 adults after inhaling sandalwood essential oil (SEO) or bergamot essential oil (BEO). Classification models were built using discriminant analysis (DA), support vector machine (SVM), and random forest (RF) algorithms to identify low or high arousal emotions. Notably, this study introduced the β/α ratio as a novel frequency domain feature to enhance model performance for the first time. Both self-reporting and EEG spectral analysis indicated that SEO promotes relaxation, whereas BEO enhances attentiveness. In model testing, incorporating the β/α ratio enhanced the performance of all models, with the accuracy of DA, SVM, and RF increasing from 70%, 75%, and 85% to 75%, 80%, and 95%, respectively. This study validated the authenticity of emotions by employing a combination of subjective and objective methods and highlighted the importance of β/α in emotion recognition along the arousal dimension.

1. Introduction

Emotion recognition has garnered increasing attention from researchers across various interdisciplinary fields. A primary scientific problem in affective computing lies in enabling computers to accurately process, identify, and interpret emotional information conveyed by humans [1]. Although facial expressions and vocal cues offer valuable insights, they are highly susceptible to individual and cultural differences, potentially reducing the accuracy of emotion recognition [2]. Therefore, it is essential to employ more direct and objective methods for assessing emotional states. The electroencephalography (EEG) method, which records the electrical activity of pyramidal neurons in the cerebral cortex as a physiological indicator [3], is characterized by being non-invasive, rapid, portable, possessing high temporal resolution, providing objective measurement of emotions, and being sensitive to emotional changes [4]. It has attracted significant attention from both academia and industry [5,6,7].
In emotion recognition research using EEG signals, features are typically extracted from the δ, θ, α, and β frequency bands. However, subdivided frequency bands (α1, α2, β1, and β2) contain more detailed emotional information [8,9,10], which can enhance model performance [11,12]. Furthermore, the β/α ratio is often regarded as an indicator of arousal, as a higher ratio is associated with greater attentional focus [13,14,15]. This ratio effectively represents emotions in the arousal dimension. Consequently, incorporating the β/α ratio into the feature construction of emotion recognition models is a potential means to improve their performance.
With the growing popularity of olfactory aromatherapy, individuals are increasingly using essential oils to regulate their mood, as this method is convenient, fast-acting, and safe [16]. Due to their distinct chemical compositions, different essential oils can elicit specific emotional responses. For instance, Satou et al. [17] investigated the relationship between emotional behavior in mice and brain concentrations of (+)-α-santalol, the primary volatile component of sandalwood essential oil (SEO). They observed its transfer to the brain, indicating that this compound exerts a sedative effect through pharmacological mechanisms. Similarly, Chang et al. [18] examined the effects of inhaling bergamot essential oil (BEO) on depression-like behaviors and hippocampal neural plasticity in rats subjected to chronic unpredictable mild stress (CUMS). Their findings suggested that BEO alleviates depressive symptoms by preserving hippocampal neuronal plasticity, thereby enhancing mood. This approach of modulating emotions via olfactory stimulation has attracted significant attention from researchers in the field of emotion recognition [19,20].
A comprehensive understanding of emotions is beneficial to emotional recognition research. In psychology and computer vision, emotions are classified as categorical or dimensional models [21,22,23]. In the categorical model, Ekman et al. [23] defined the basic human emotions as happiness, anger, disgust, fear, sadness, and surprise. Sauceda et al. [24] utilized three neural network algorithms (ShallowFBCSPNet, Deep4Net, and EEGNetv4) to recognize emotions, including happiness, sadness, disgust, neutrality, and fear, in the SEED-V dataset. In contrast, the dimensional emotional model quantifies emotions along multiple dimensions, including valence, arousal, and dominance, thereby capturing more complex emotional states [22]. Kroupi et al. [25] established a model based on pleasant, neutral, and unpleasant odors, finding that the model was particularly sensitive to unpleasant smells and exhibited good classification performance. However, the accuracy of the model decreased markedly when distinguishing between pleasant and neutral odors. This may indicate that the use of olfactory stimuli to regulate emotions is effective, as exogenous stimuli can enhance emotional self-monitoring [26], thereby facilitating the understanding of emotions. Despite these advancements, existing research in emotion recognition has primarily focused on the valence dimension. In reality, each emotional state is a linear combination of various dimensions. Accurate emotion recognition requires the simultaneous consideration of multiple emotional dimensions. The widely used VA (valence–arousal) dimensional emotional model links valence to the degree of pleasure and arousal to the degree of excitement, with higher dimension values corresponding to greater emotional intensity. This model treats emotional experiences as a continuum of related but often ambiguous states, effectively capturing the nuances of emotional expression [27].
This study investigates the emotional effects of inhaling SEO, which is believed to induce relaxation in olfactory aromatherapy, and BEO, which is thought to enhance concentration. By combining subjective self-report scores with objective EEG spectral analysis, we assessed and validated the emotional effects induced by inhaling SEO or BEO. Using EEG data collected after the inhalation of SEO or BEO, we introduced a novel feature construction method, namely extracting features from six frequency bands (δ, θ, α1, α2, β1, and β2) across all electrodes in five brain regions, and incorporating the β/α ratio as an initial feature. This approach led to the development of a high-performance model for classifying low-arousal (SEO) and high-arousal (BEO) emotions. The model enables rapid and accurate differentiation between various emotional states, facilitating timely monitoring of emotional changes. It provides an effective tool for screening essential oils that can evoke distinct emotional responses in various applications, such as aromatherapy, human–computer interaction, and multimedia.

2. Materials and Methods

2.1. Participants

Fifty young healthy adults (twenty-eight females, aged 22.71 ± 1.98 years; twenty-two males, aged 22.73 ± 1.24 years, mean ± SD) were recruited from Shanghai Institute of Technology. All participants provided written informed consent and were compensated monetarily.

2.2. Olfactory Stimulation

According to previous studies, sandalwood essential oil and bergamot essential oil are commonly used in olfactory aromatherapy and are capable of eliciting significant emotional responses [16]. Specifically, sandalwood essential oil has been associated with emotional relaxation [17], while bergamot essential oil has been linked to mood enhancement [18]. These contrasting emotional effects are advantageous for developing machine learning models, as they provide distinct emotional states for classification and analysis. Therefore, in this study, undiluted sandalwood (Santalum album L.) essential oil (SEO) and bergamot (Citrus medica L. var. sarcodactylis Swingle) essential oil (BEO), sourced from Quintis Trading Co., Ltd. (Xiamen, China) and Zhejiang Golden Hand Biotechnology Co., Ltd. (Zhejiang, China), respectively, were employed to produce olfactory stimuli. For the convenience of the participants, unscented aroma diffuser woods were procured as carriers for the essential oils. The EEG data obtained during the inhalation of SEO was designated as 0 (negative sample), and the data corresponding to BEO inhalation was marked as 1 (positive sample).

2.3. Experimental Design

The experiment comprised three distinct sessions (Figure 1A). Prior to each session, participants engaged in a brief period of physical relaxation (up to one minute) in preparation for the EEG data acquisition. During the EEG recording of each session, participants held a diffuser wood (no fragrance) approximately 3 cm from their nose to inhale the essential oil for a duration of two minutes, while being instructed to limit bodily movements and maintain normal eye openness. The diffuser wood was subjected to the following treatments across the sessions: (1) no treatment; (2) addition of 0.1 g SEO; (3) addition of 0.1 g BEO. Subsequent to the EEG data collection in the second and third sessions, participants completed a 30 s questionnaire to assess their emotional responses to the essential oils, based on the valence–arousal (VA) emotion model [27]. They rated the oils on a scale of valence and arousal from 1 to 9, with 1 indicating low pleasantness and excitement, 9 indicating high, and 5 being neutral. To prevent olfactory cross-contamination, the experimental area was ventilated for one minute following each inhalation of essential oil.

2.4. EEG Recording

EEG data were captured using an EEG acquisition device (eggoTM mylab, ANT Neuro, Hengelo, The Netherlands), equipped with 32 AgCl electrodes arranged according to an expanded 10–20 system [28] (refer to the “Channel Locations” module in Figure 1B). The impedance of the electrodes was maintained below 5 kΩ. A band-pass filter was applied, featuring a low-pass cutoff frequency of 1 Hz and a high-pass cutoff frequency of 30 Hz. Concurrently, a notch filter was utilized to attenuate the 50 Hz power line noise, with cutoff frequencies set at 49 Hz and 51 Hz. The sampling rate of EEG device was established at 512 Hz.

2.5. EEG Data Preprocessing

The EEG data were re-referenced to the average values of the M1 and M2 electrodes at the bilateral mastoids, after which the reference electrode was excluded, leaving 30 electrodes for analysis. The data were processed using EEGLAB (version 2024.0) [29] and MATLAB (R2024a) functions. Contaminants, such as eye blinks and motion artifacts, were mitigated using the second-order blind identification (SOBI) algorithm [30], with all parameters remaining default. Baseline drift was minimized by applying a detrending function to the continuous linear trend present in the data from each electrode using MATLAB, with all parameters remaining default. Ultimately, one minute of continuous EEG data were preserved for further analysis.

2.6. Spectral Decomposition

The power spectral density (PSD) was estimated using the Welch method, employing a Hanning window of 2 s for segmentation and an overlap of 256 samples (50% overlap rate). The number of discrete Fourier transform (DFT) points was 1024.
Since high-frequency signals are often doped with electromyographic (EMG) artifacts, these are difficult to remove cleanly in long-term monitoring [31], which, in turn, affects the experimental results. The 1–30 Hz range is a common EEG study frequency band range which can provide most of the emotional information [32]. Therefore, this study focused on the PSD within the 1–30 Hz frequency range, which constitutes the band of interest. Given the distinct biological significance of EEG signal frequency bands [8,9,10,33,34], the 1–30 Hz range was segmented into six sub-bands, namely delta (δ, 1–4 Hz), theta (θ, 4–8 Hz), alpha1 (α1, 8–10 Hz), alpha2 (α2, 10–13 Hz), beta1 (β1, 13–20 Hz), and beta2 (β2, 20–30 Hz). The PSD of all frequency points within each frequency band at each electrode was averaged to obtain the mean PSD value for that band. Additionally, we introduced the β/α ratio as a new metric, since it reflects the level of brain arousal [13,14,15]. Subsequently, we averaged the PSD of the β/α ratio across all electrodes. It should be noted that in the β/α ratio, β represents the entire β frequency band (13–30 Hz), while α represents the entire α frequency band (8–13 Hz).
The brain was divided into five regions of interest (ROIs), namely F-ROI, T-ROI, C-ROI, P-ROI, and O-ROI, with electrodes in each ROI distinguished by unique colors (refer to the “Channel Locations” module in Figure 1B). Given that this study focuses exclusively on the arousal dimension of emotions, the α frequency band, which is associated with internal attention [35], and the β/α ratio, which serves as a potential indicator of arousal [14], are sufficient to support the electroencephalographic analysis conducted in this research. In addition, the α frequency band is considered to have the highest retest reliability and is a characteristic of the internal stability of the individual [36], so no other separate EEG bands (1–30 Hz) were visualized and further analyzed. To facilitate an intuitive comparison of the PSD across different brain ROIs before and after inhaling SEO or BEO, we averaged the PSD of three metrics (α1, α2, and β/α) at each electrode in both the resting state (RS) group and the inhalation of SEO or BEO groups, and then computed the intragroup averages to generate topographic maps. A paired t-test (one-tailed) was conducted on the average PSD of the three metrics at each electrode between groups, with a significance level of α = 0.05. Additionally, we plotted the spectrograms of the intersubject average PSD across different ROIs for all frequency points within the 8–30 Hz range (with a frequency resolution of 0.5 Hz). The PSD for each ROI was calculated as the average of all electrodes in that region. Paired t-tests (one-tailed) were also performed on the intersubject average PSD of each metric between the RS and SEO or BEO groups across different ROIs, with a significance level of α = 0.05.

2.7. Division of the Training Set and Test Set

The division of the dataset into training and test sets followed the conventional 8:2 ratio. To ensure a balanced representation of positive and negative samples, 40 positive and 40 negative samples were randomly selected to form the training set, while 10 positive and 10 negative samples constituted the test set.

2.8. Model Selection and Feature Construction

We employed three classic models for emotion recognition, namely discriminant analysis (DA), support vector machine (SVM) [37], and random forest (RF). In the label assignment, EEG data from SEO inhalation were labeled as 0 (negative class), and EEG data from BEO inhalation were labeled as 1 (positive class). Given the potential of the β/α ratio as an arousal indicator, we introduced it during feature construction. The mean PSD of traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2) and the β/α ratio at each electrode in all ROIs were used as initial features. Each ROI thus contributed 7 initial features, resulting in a total of 35 features (see the “Dataset” module in Figure 1B). To assess the impact of incorporating the β/α ratio on model performance, we constructed two sets of models. The first set (DA-1, SVM-1, RF-1) used traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2) for feature construction, while the second set (DA-2, SVM-2, and RF-2) incorporated the β/α ratio, with all other conditions kept constant.
To enhance numerical stability and expedite model convergence, we applied standardization and normalization to the dataset, which consists of a 100 by 35 array. We performed both processes column-wise. Standardization adjusted the mean of each feature to zero and the standard deviation to one. Normalization scaled the feature data to the [−1, 1] interval, thereby mitigating the influence of outliers on the model. Given that the DA and RF models are insensitive to the range of feature data, we normalized the datasets only for these models. Conversely, the SVM model, which is sensitive to feature data range, underwent both standardization and normalization. To eliminate redundant information and improve model performance, we applied the partial least squares (PLS) regression algorithm for dimensionality reduction and selected the first m components, where the cumulative explained variance ratio exceeded 95%. The final dataset used for model input had a size of 100 by m (samples: 100, features: m).

2.9. Model Optimization

We constructed the models based on the training set to find the optimal hyperparameters. For the DA model, due to its simplicity, hyperparameter optimization was not performed, and the discriminant type was set to “diaglinear”. For the SVM and RF models, 5-fold cross-validation was conducted during each grid search (Supplementary Figure S1) to identify the best hyperparameters. Notably, the kernel function used in the SVM was the radial basis function (RBF). The objective for DA and SVM was accuracy, while for RF, the objective was accuracy plus AUC. In the SVM, the hyperparameters selected were cost (c) and gamma (g), both of which were subject to grid search with exponential growth (base 2). The grid search range for c and g was from 2−8 to 28. For RF, the hyperparameters selected were the number of decision trees and the minimum number of leaves, with step sizes of 10 and 1, respectively. The grid search range for these parameters was from 10 to 300 and from 1 to 20.

2.10. Model Testing

To effectively and quantitatively assess the performance of each binary classification model, we utilized the following metrics on the test set: accuracy, precision, sensitivity, specificity, and F1 score. Accuracy is defined as the ratio of the number of samples correctly predicted by the model to the total number of samples. Precision denotes the fraction of positively predicted samples that are indeed positive. Sensitivity is the proportion of actual positive samples that are correctly identified by the model. Specificity refers to the proportion of actual negative samples that the model accurately classifies as negative. The F1 score, the harmonic mean of precision and sensitivity, considers both precision and recall. These metrics are calculated according to the following “Equations (1)–(5)”:
A c c u r a c y = T P + T N / ( T P + T N + F P + F N )
P r e c i s i o n = T P / ( T P + F P )
S e n s i t i v i t y = T P / ( T P + F N )
S p e c i f i c i t y = T N / ( T N + F P )
F 1   s c o r e = 2 × P r e c i s i o n × S e n s i t i v i t y / ( P r e c i s i o n + S e n s i t i v i t y )
F P R = F P / ( F P + T N )
where TP and TN represent the counts of positive and negative samples that the model correctly identified, respectively. Conversely, FP and FN denote the counts of positive and negative samples that the model incorrectly identified, respectively.
We constructed the receiver operating characteristic (ROC) curve to assess the overall performance of model. The ROC curve is derived from a series of varying classification thresholds for the model. At each threshold, we calculate the true positive rate (TPR) and false positive rate (FPR) of the model, which serve as the coordinates (FPR, TPR) for the curve [38]. The TPR, synonymous with sensitivity, and the FPR, defined by “Equation (6)”, represent the proportion of negative samples incorrectly classified as positive by the model. The metrics in “Equations (1)–(6)” range from 0 to 1, with values closer to 1 indicating superior model performance. Notably, an AUC of 0.5 suggests that the classification capability of model is no better than random chance.

2.11. Statistical Analysis

Unless otherwise stated, all statistical tests were conducted using the built-in functions in MATLAB (R2024a). All statistical tests reported in this study were paired t-tests (one-tailed), the confidence level was 95%, and p-values < 0.05 were considered significant. When performing paired t-tests, the RS group was used as the first input sample data group. This study conducted multiple paired t-tests between electrodes or regions, but did not perform statistical corrections (such as FDR correction), which may lead to an increase in Type I errors, reducing the reliability of validating hypotheses to some extent. However, this approach also reduces the generation of Type II errors, contributing to the screening of potential candidate electrodes or brain regions.

3. Results

3.1. Demographics and Experimental Setup

We collected both self-reported (subjective level) and EEG (objective level) data on emotional evaluation following the inhalation of SEO or BEO from fifty participants (twenty-eight females, 22.71 ± 1.98 years; twenty-two males, 22.73 ± 1.24 years, mean ± SD). The experimental procedure is summarized in Figure 1A.

3.2. Subjective Evaluation

We employed self-report measures to quantify the emotional responses of participants to inhaling SEO or BEO along the dimensions of arousal and valence.

3.2.1. Arousal Dimension

In the arousal dimension (Figure 2A), participants who inhaled SEO exhibited a lower average arousal level (3.86 ± 1.73, mean ± SD), in contrast to those who inhaled BEO, who displayed a higher average arousal (6.68 ± 1.71). It is noteworthy that, despite considerable variation among the participants, only seven individuals (three females) categorized SEO in the high-arousal zone (>5), while four participants (two females) categorized BEO in the low-arousal zone (<5). Furthermore, the normal distribution curve clearly distinguishes between SEO and BEO (Figure 2A). Based on these subjective assessments, we conclude that inhaling SEO induces low-arousal emotions, whereas inhaling BEO elicits emotions characterized by high arousal, which are the opposite of those induced by SEO. The emotional ratings for female participants inhaling SEO were notably clustered (1 to 6), with a higher mean value (4.04 ± 1.40), whereas male participants exhibited a broader range of responses (1 to 8) with a lower mean value (3.64 ± 2.08). For both female and male participants inhaling BEO, the ratings followed a similar normal distribution. Females displayed a wider distribution with a lower mean (6.32 ± 1.68), compared to males, whose ratings were more narrowly distributed with a slightly higher mean (7.14 ± 1.67).
In summary, from a subjective standpoint, inhaling SEO and BEO distinctly affects emotions along the arousal dimension: SEO induces low-arousal emotions, whereas BEO induces high-arousal emotions. These findings lay the groundwork for label assignment in subsequent machine learning models.

3.2.2. Valence Dimension

In the valence dimension (Figure 2B), participants who inhaled SEO or BEO demonstrated higher average valence scores (5.9 ± 1.45 and 6.22 ± 2.11, respectively). Similar to the arousal dimension, there is considerable variation among participants; however, only nine individuals (six females) categorized SEO in the low valence zone (<5), and ten participants (three females) categorized BEO in the same low valence zone (<5). Additionally, a paired t-test was conducted on the valence ratings following the inhalation of sandalwood essential oil (SEO) and bergamot essential oil (BEO) across all participants (α = 0.05). The results showed no significant difference between the two conditions (p = 0.3924; see Supplementary Table S1 for detailed rating results). Therefore, based on these subjective assessments, we conclude that inhaling either SEO or BEO elicits similar valence emotions. The distribution of SEO ratings was relatively narrow for both genders, with females reporting a slightly lower mean (5.64 ± 1.45) compared to males (6.23 ± 1.41). Conversely, BEO ratings by females were more narrowly distributed with a higher mean (6.64 ± 1.77), while males exhibited a broader distribution with a lower mean (5.68 ± 2.42).
In conclusion, both essential oils lead to similar emotional experiences in the valence dimension, with no significant difference, so the valence dimension will not be further analyzed in this study.

3.3. EEG Spectral Analysis

To reveal the activation of pyramidal neurons across different brain ROIs upon inhalation of SEO or BEO, we processed EEG data from fifty participants, resulting in the creation of topographic maps (Figure 3A–G) and spectrograms (Figure 3H,I).

3.3.1. Analysis of Topological Maps

Compared to the RS, inhalation of either SEO or BEO leads to a varying increase in the intersubject average PSD across the α1 and α2 frequency bands in all ROIs. Notably, inhalation of SEO reduces the intersubject average PSD of the β/α ratio across all ROIs, whereas inhalation of BEO increases it (Figure 3A–G).
For the SEO group (Figure 3D,F), compared to the RS group, the PSD of the α1 rhythm significantly increased, primarily in the F-ROI, the right T-ROI (“T8”), the right and central C-ROI, and several electrodes in the P-ROI (“P7”, “Pz”, “P8”), as well as in the left O-ROI (“O1”). The PSD of the α2 rhythm significantly increased mainly in the F-ROI, a single electrode in the central C-ROI (“Cz”), the right and central P-ROI (“P8”, “Poz”), and the O-ROI. The PSD of the β/α ratio decreased significantly across all ROIs, except for “P7” (specific p-values and t-values are provided in Supplementary Table S2).
For the BEO group (Figure 3E,G), compared to the RS group, the PSD of the α1 rhythm significantly increased, primarily in several electrodes in the anterior F-ROI (“Fp1”, “Fpz”, “Fp2”, “F7”, “F3”, “FC1”), the left T-ROI (“T7”), a few electrodes in the central C-ROI (“CP1”, “CP2”), and a few electrodes in the central P-ROI (“Pz”). The PSD of the α2 rhythm significantly increased mainly in the F-ROI, the right and central P-ROI (“Pz”, “P4”, “P8”), and the left O-ROI (“O1”). The PSD of the β/α ratio decreased significantly only at the “T8” electrode. Specific p-values and t-values are provided in Supplementary Table S3. The exact electrode locations are shown in the “Channel Location” module of Figure 1B.

3.3.2. Analysis of Spectrograms

For the α1 and α2 frequency bands, different ROIs exhibited similar patterns (Figure 3H,I), which supports the presence of a spatial ambiguity effect [39] in EEG signals. This effect occurs because voltage fluctuations measured at any electrode on the scalp result from the combined activity of multiple sources across different locations.
The spectrogram results show that, compared to the RS group, inhalation of either SEO or BEO led to varying degrees of increase in the PSD of α1 and α2 across all ROIs. (Figure 3H,I). Specifically, for the alpha1 rhythm (Figure 3H), inhalation of SEO resulted in a significant PSD rise in all ROIs (F-ROI: t = −2.192, p = 0.017; T-ROI: t = −1.916, p = 0.031; C-ROI: t = −2.436, p = 0.009; P-ROI: t = −1.820, p = 0.038; O-ROI: t = −1.766, p = 0.042), whereas inhalation of BEO led to a significant increase solely in the F-ROI (F-ROI: t = −1.700, p = 0.048; T-ROI: t = −1.482, p = 0.072; C-ROI: t = −1.593, p = 0.059; P-ROI: t = −1.385, p = 0.086; O-ROI: t = −0.502, p = 0.309). In the case of the alpha2 rhythm (Figure 3I), all ROIs except the T-ROI and C-ROI regions showed a significant enhancement due to SEO (F-ROI: t = −2.213, p = 0.016; T-ROI: t = −0.898, p = 0.187; C-ROI: t = −1.087, p = 0.141; P-ROI: t = −1.908, p = 0.031; O-ROI: t = −2.604, p = 0.006), with BEO affecting only the F-ROI (F-ROI: t = −2.151, p = 0.018; T-ROI: t = −0.732, p = 0.234; C-ROI: t = −1.075, p = 0.144; P-ROI: t = −1.671, p = 0.051; O-ROI: t = −1.646, p = 0.053).
In conclusion, compared to RS, inhalation of either SEO or BEO resulted in varying degrees of increase in the PSD of α1 and α2 across all ROIs. The distinction lies in the alpha1 frequency band, where inhalation of SEO significantly increases PSD in most ROIs, while inhalation of BEO does so only at a few electrodes. Notably, the PSD of the β/α ratio decreased across all ROIs in the SEO group, while it increased across all ROIs in the BEO group. In conjunction with the self-reported data, which indicated that inhalation of SEO and BEO induced low and high arousal emotions, respectively, these results suggest that the β/α ratio has the potential to serve as an arousal marker [13,14,15]. These findings also provide objective evidence that inhaling SEO induces low-arousal emotions, while inhaling BEO induces high-arousal emotions.

3.4. Model Evaluation

Based on the subjective (self-reported) and objective (EEG) results, inhalation of either SEO or BEO clearly differentiates emotions along the arousal dimension (inhaling SEO induces low-arousal emotions, while inhaling BEO induces high-arousal emotions). However, in the valence dimension, no significant difference is observed between the two (See Supplementary Table S1), as both SEO and BEO induce high-valence emotions. Therefore, we focus solely on the arousal dimension of emotions, with SEO labeled as low arousal (label: 0) and BEO labeled as high arousal (label: 1).
We employed six models (DA-1, SVM-1, RF-1, DA-2, and SVM-2, RF-2) for binary classification. The hyperparameter optimization diagram for SVM-1 (Figure 4A) illustrates that when c is fixed, the validation set accuracy generally increases and stabilizes as g decreases. Notably, the highest accuracy (66.25%) occurs when c = 1 and g = 0.5. The hyperparameter optimization diagram for SVM-2 (Figure 4B) indicates a similar trend to SVM-1, with the difference being that validation set accuracy rises to a plateau as c increases within the range of 16–256. The highest accuracy (70%) is achieved when c = 16 and g = 0.5.
The hyperparameter optimization diagrams for RF-1 and RF-2 (Figure 5) show that when the number of decision trees is fixed, the validation set accuracy and AUC generally increase and stabilize as the minimum number of leaves increases, with changes in AUC being particularly noticeable. For RF-1, the optimal values (accuracy = 0.7, AUC = 0.7406) are achieved with 20 decision trees and a minimum of 13 leaves. For RF-2, the optimal values (accuracy = 0.7625, AUC = 0.7594) are achieved with 30 decision trees and a minimum of 14 leaves.
In the final evaluation of the test set (sample size = 20), we used the above hyperparameters for model prediction (Figure 6). The results from six metrics (accuracy, precision, sensitivity, specificity, ROC AUC, and F1 score) demonstrate that the performance of DA-1, SVM-1, and RF-1 models is lower than that of DA-2, SVM-2, and RF-2, indicating that incorporating the β/α ratio into feature construction effectively enhances model performance. Furthermore, comparing the performance across different model types (DA, SVM, and RF), the results from the six metrics show that DA and SVM models perform similarly and poorly, while RF models excel across all six metrics, significantly outperforming DA and SVM. In summary, for feature construction, incorporating the β/α ratio combined with the RF algorithm is an effective method for odor-induced EEG emotion recognition, compared to traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2).

4. Discussion

This study is the first to use the inhalation of SEO or BEO as an olfactory stimulus for emotion recognition based on the EEG signals of participants. By subdividing the α and β frequency bands, introducing new variables (β/α), and incorporating different ROIs, new EEG spectral features were constructed, leading to the development of a high-performance emotion recognition model. In the performance test of the classification models, the introduction of the β/α value markedly improved the performance of all models (DA, SVM, and RF). Among them, the RF model demonstrated the best performance in handling high-dimensional frequency–spatial domain EEG signal characteristics, indicating its effectiveness as a tool for analyzing odor-induced EEG emotion recognition.
The subjective self-reported results showed that inhaling SEO generally induced low-arousal, high-valence emotions, while inhaling BEO generally induced high-arousal, high-valence emotions. These results align with previous studies showing that SEO has relaxing and calming properties [40]. This may be due to its main volatile component (+)-α-Santalol [17]. BEO has mood-boosting and attention-focusing properties [41], and limonene may play an important role as the main volatile component of BEO [42]. Notably, the self-reported results also conveyed interesting information: in the arousal dimension, male participants were more sensitive to SEO than female participants, perceiving SEO as inducing lower arousal emotions; in the valence dimension, the ratings for BEO were highly dispersed, especially among male participants.
Current EEG spectral analysis research generally recognizes that different frequency bands have different physiological meanings [33,43]. For example, alpha rhythm synchronization is associated with enhanced calmness [44], while beta rhythms usually occur when we are alert, focused, and engaged in problem solving [45]. Furthermore, studies have shown that a higher β/α ratio is associated with increased brain focus [13,14,15]. In the topographic analysis, we found that compared to RS, inhaling SEO significantly increased the PSD of the alpha1 frequency band in the frontal lobe, right temporal lobe, central area, and parietal lobe, and significantly increased the PSD of the alpha2 rhythm mainly in the frontal lobe, right and central parietal lobes, and occipital lobe. Inhaling BEO significantly increased the PSD of the alpha1 rhythm in the prefrontal lobe, left temporal lobe, and a few electrodes in the central area, and significantly increased the PSD of the alpha2 rhythm mainly in the frontal lobe, right and central parietal lobes, and left occipital lobe. It is noteworthy that inhalation of SEO or BEO significantly alters the alpha1 and alpha2 rhythms in the frontal lobe, possibly due to the involvement of the orbitofrontal cortex in odor processing [46]. More interestingly, inhalation of SEO significantly reduced the PSD of the β/α ratio across all ROIs, while inhalation of BEO significantly increased the PSD of the β/α ratio in the right temporal lobe. These findings suggest that inhaling SEO promotes brain relaxation, while inhaling BEO enhances focus, which is consistent with the self-reported results in this study. Additionally, they further support the potential of the β/α ratio as an arousal marker. Our spectral analysis results show that, compared to the RS, inhalation of SEO significantly increases the PSD of the alpha1 rhythm in all ROIs, while the alpha2 rhythm shows a significant increase only in the frontal lobe. In contrast, the inhalation of BEO results in a significant increase in the PSD of both alpha1 and alpha2 rhythms exclusively in the frontal lobe. The differences between the alpha1 and alpha2 results may be due to the former reflecting general task demands, such as attention processes [9,34], and the latter possibly reflecting specific task demands, such as semantic memory processes [10].
In the final model test, the comprehensive performance of the classifiers was ranked as follows: RF > SVM > DA. This suggests that more complex classifiers are better at identifying emotions induced by odors and reflected by EEG. RF (random forest) likely outperforms others [47] because it can more accurately capture nonlinear relationships between features by integrating multiple decision trees [48]. SVM (support vector machine), while capable of addressing nonlinear problems through kernel functions that map data to high-dimensional space, requires careful parameter tuning [49] and is more sensitive to feature engineering than RF [50]. Consequently, RF surpasses both the linear model (DA) and the kernel-driven SVM. This finding also supports the notion that “the mapping relationship between EEG signals and emotional states is inherently nonlinear”. Additionally, when comparing models with the new β/α features (DA-2, SVM-2, and RF-2), the performance indices were higher than those of models without these features (DA-1, SVM-1, and RF-1). This indicates that the β/α ratio is directly linked to emotional arousal. Furthermore, the improvement in cross-model performance with the β/α features suggests that the β/α ratio is a universal, robust, and interpretable biomarker. By integrating this emotion marker with advancements in wearable device miniaturization technology [51,52] and edge computing deployment frameworks [53], this framework enables the construction of an olfactory stimulation-driven EEG emotion recognition model. The model supports rapid and precise identification of diverse emotional states, enables real-time monitoring, and advances the application of olfactory-driven emotion recognition in such domains as aromatherapy, human–computer interaction, and multimedia.
The limitations of this study include the following: (a) the absence of additional demographic data (e.g., educational level, ethnicity, and handedness) that may influence EEG signals, potentially limiting the generalizability of findings to broader populations; (b) the evaluation of the model relied solely on a single 80/20 split for training and test sets, which may constrain the generalizability of its performance; (c) controlled laboratory settings may not fully replicate real-world olfactory environments, thereby reducing the ecological validity of the results.

5. Conclusions

This study explored the effects of inhaling SEO or BEO on emotions from both subjective (self-reporting) and objective (EEG) perspectives. In addition, this study utilized EEG to assess emotions in odor-induced emotion recognition. By incorporating the β/α ratio and subdivided EEG frequency bands (1–30 Hz), features were extracted and integrated into a machine learning framework, resulting in the development of three emotion recognition models. The experimental results confirmed that the inhalation of SEO or BEO effectively induced low or high arousal emotions, ensuring the authenticity of the emotional labels in the emotion recognition process. Furthermore, the inclusion of the β/α ratio markedly improved the performance of all models (DA, SVM, and RF). For instance, the accuracy of the DA and SVM models increased by 5%, while the accuracy of RF model improved by 10%. Notably, the RF model demonstrated a distinct advantage over the other two models in handling emotion recognition based on EEG signals induced by olfactory stimuli, with RF-1 and RF-2 achieving test accuracies of 85% and 95%, respectively, which may be due to the high nonlinear correlation of EEG characteristics, while RF model has excellent ability to capture nonlinear characteristics.
This study validated the effectiveness of the β/α ratio as an objective emotional indicator and highlighted its importance in EEG-based emotion recognition models. These models are capable of classifying emotions with different levels of arousal and contribute to a deeper understanding of how essential oils influence brain activity and emotions, thereby providing scientific support and new research approaches for olfactory aromatherapy.
However, there are still several limitations in the future. For example, different individuals have different sensitivity to odors, which may lead to inconsistencies in emotional responses, which, in turn, affects the generalization ability of the model. In addition, the limited variety of essential oils used in the current study will have led to insufficient coverage of the emotions induced and limit the application scenarios of the model. Future research should further explore factors, such as individual differences and essential oil types, to improve the practicality and robustness of the emotion recognition model.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/app15094980/s1, Figure S1: cross-validation diagram; Table S1: rating and statistical analysis of participants inhaling SEO or BEO in valence dimension; Table S2: p-values and t-values of paired t-test between RS and inhaling SEO in topography analysis; Table S3: p-values and t-values of paired t-test between RS and inhaling BEO in topography analysis; Table S4: t-values of paired t-test between RS and inhaling SEO in spectrogram analysis; Table S5: t-values of paired t-test between RS and inhaling BEO in spectrogram analysis.

Author Contributions

Conceptualization, J.F., G.Z. and F.Y.; Methodology, J.F., S.L., G.Z. and F.Y.; Software, J.F.; Investigation, J.F., G.Y., S.L., S.Z., G.Z. and F.Y.; Writing—Original Draft Preparation, J.F.; Writing—Review and Editing, G.Y., S.L., S.Z., G.Z. and F.Y.; Visualization, J.F.; Supervision, G.Y., G.Z. and F.Y.; Funding Acquisition, F.Y. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by the Special Funds for Higher Education at the Municipal Level in Shanghai (Applied High-level-Interdisciplinary) [1021GK210006149-B20]; the Collaborative Innovation Fund of Shanghai Institute of Technology [XTCX2022-12]; and the Key Research Project on Camphor tree (KRPCT) of Jiangxi Forest Department [2020CXZX07].

Institutional Review Board Statement

This study adhered to the guidelines of the Declaration of Helsinki and received approval from the Ethics Committee of Shanghai Jiao Tong University, approval number #B2021153I.

Informed Consent Statement

Written informed consent has been obtained from the patients to publish this paper.

Data Availability Statement

EEG source data and essential oils in this study are available from the corresponding author on reasonable request. All statistical results are provided in the Supplementary Materials. All analysis code has been available at https://github.com/FNT0126/EEG_MATLAB.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Li, X.; Zhang, Y.; Tiwari, P.; Song, D.; Hu, B.; Yang, M.; Zhao, Z.; Kumar, N.; Marttinen, P. EEG Based Emotion Recognition: A Tutorial and Review. ACM Comput. Surv. 2022, 55, 1–57. [Google Scholar] [CrossRef]
  2. Jaiswal, A.; Raju, A.K.; Deb, S. Facial Emotion Detection Using Deep Learning. In Proceedings of the 2020 International Conference for Emerging Technology (INCET), Belgaum, India, 5–7 June 2020; pp. 1–5. [Google Scholar]
  3. Luck, S.J. An Introduction to the Event-Related Potential Technique; MIT Press: Cambridge, MA, USA, 2014. [Google Scholar]
  4. Abdumalikov, S.; Kim, J.; Yoon, Y. Performance Analysis and Improvement of Machine Learning with Various Feature Selection Methods for EEG-Based Emotion Classification. Appl. Sci. 2024, 14, 10511. [Google Scholar] [CrossRef]
  5. Can, Y.S.; Mahesh, B.; André, E. Approaches, Applications, and Challenges in Physiological Emotion Recognition—A Tutorial Overview. Proc. IEEE 2023, 111, 1287–1313. [Google Scholar] [CrossRef]
  6. Zhang, X.; Cheng, X.; Liu, H. TPRO-NET: An EEG-based emotion recognition method reflecting subtle changes in emotion. Sci. Rep. 2024, 14, 13491. [Google Scholar] [CrossRef]
  7. Mohamed, A.F.; Jusas, V. Developing Innovative Feature Extraction Techniques from the Emotion Recognition Field on Motor Imagery Using Brain–Computer Interface EEG Signals. Appl. Sci. 2024, 14, 11323. [Google Scholar] [CrossRef]
  8. Diaz, H.; Cid, F.; Otarola, J.; Rojas, R.; Caete, L. EEG Beta band frequency domain evaluation for assessing stress and anxiety in resting, eyes closed, basal conditions. Procedia Comput. Sci. 2019, 162, 974–981. [Google Scholar] [CrossRef]
  9. Karrasch, M.; Krause, C.M.; Laine, M.; Lang, A.H.; Lehto, M. Event-related desynchronization and synchronization during an auditory lexical matching task. Electroencephalogr. Clin. Neurophysiol. 1998, 107, 112–121. [Google Scholar] [CrossRef]
  10. Klimesch, W.; Vogt, F.; Doppelmayr, M. Interindividual differences in alpha and theta power reflect memory performance. Intelligence 1999, 27, 347–362. [Google Scholar] [CrossRef]
  11. Li, J.; Zhang, Z.; He, H. Hierarchical Convolutional Neural Networks for EEG-Based Emotion Recognition. Cogn. Comput. 2017, 10, 368–380. [Google Scholar] [CrossRef]
  12. Wei-Long, Z.; Bao-Liang, L. Investigating Critical Frequency Bands and Channels for EEG-Based Emotion Recognition with Deep Neural Networks. IEEE Trans. Auton. Ment. Dev. 2015, 7, 162–175. [Google Scholar] [CrossRef]
  13. Freeman, F.G.; Mikulka, P.J.; Prinzel, L.J.; Scerbo, M.W. Evaluation of an adaptive automation system using three EEG indices with a visual tracking task. Biol. Psychol. 1999, 50, 61–76. [Google Scholar] [CrossRef]
  14. Moon, S.A.; Bae, J.; Kim, K.; Cho, S.Y.; Kwon, G.; Lee, R.; Ko, S.H.; Lim, S.; Moon, C. EEG Revealed That Fragrances Positively Affect Menopausal Symptoms in Mid-life Women. Exp. Neurobiol. 2020, 29, 389–401. [Google Scholar] [CrossRef] [PubMed]
  15. Rajendran, V.G.; Jayalalitha, S.; Adalarasu, K. EEG Based Evaluation of Examination Stress and Test Anxiety Among College Students. IRBM 2022, 43, 349–361. [Google Scholar] [CrossRef]
  16. Farrar, A.J.; Farrar, F.C. Clinical Aromatherapy. Nurs. Clin. N. Am. 2020, 55, 489–504. [Google Scholar] [CrossRef] [PubMed]
  17. Satou, T.; Ogawa, Y.; Koike, K. Relationship Between Emotional Behavior in Mice and the Concentration of (+)-α-Santalol in the Brain. Phytother. Res. 2015, 29, 1246–1250. [Google Scholar] [CrossRef]
  18. Chang, J.; Yang, H.; Shan, X.; Zhao, L.; Li, Y.; Zhang, Z.; Abankwah, J.K.; Zhang, M.; Bian, Y.; Guo, Y. Bergamot essential oil improves CUMS-induced depression-like behaviour in rats by protecting the plasticity of hippocampal neurons. J. Cell. Mol. Med. 2024, 28, e18178. [Google Scholar] [CrossRef]
  19. Hou, H.R.; Zhang, X.N.; Meng, Q.H. Odor-induced emotion recognition based on average frequency band division of EEG signals. J. Neurosci. Methods 2020, 334, 108599. [Google Scholar] [CrossRef]
  20. Xing, M.; Hu, S.; Wei, B.; Lv, Z. Spatial-frequency-temporal convolutional recurrent network for olfactory-enhanced EEG emotion recognition. J. Neurosci. Methods 2022, 376, 109624. [Google Scholar] [CrossRef]
  21. Huang, Z.-Y.; Chiang, C.-C.; Chen, J.-H.; Chen, Y.-C.; Chung, H.-L.; Cai, Y.-P.; Hsu, H.-C. A study on computer vision for facial emotion recognition. Sci. Rep. 2023, 13, 8425. [Google Scholar] [CrossRef]
  22. Russell, J.A. An Approach to Environmental Psychology; MIT Press: Cambridge, MA, USA, 1974. [Google Scholar]
  23. Ekman, P. Basic Emotions. In Handbook of Cognition and Emotion; Dalgleish, T., Power, M.J., Eds.; Wiley: Hoboken, NJ, USA, 1999; pp. 45–60. [Google Scholar]
  24. Mendivil, S.J.A.; Marquez, B.Y.; Esqueda, E.J.J. Emotion Classification from Electroencephalographic Signals Using Machine Learning. Brain Sci. 2024, 14, 1211. [Google Scholar] [CrossRef]
  25. Kroupi, E.; Vesin, J.-M.; Ebrahimi, T. Subject-Independent Odor Pleasantness Classification Using Brain and Peripheral Signals. IEEE Trans. Affect. Comput. 2016, 7, 422–434. [Google Scholar] [CrossRef]
  26. Serrano, A.C.B.; Romero, E.C.; González, I.B. Desarrollo de la comprensión lectora a través de la colaboración y el diálogo: Una intervención en el contexto natural del aula desde el Modelo PASS. Investig. Sobre Lect. 2023, 18, 88–114. [Google Scholar] [CrossRef]
  27. Posner, J.; Russell, J.A.; Peterson, B.S. The circumplex model of affect: An integrative approach to affective neuroscience, cognitive development, and psychopathology. Dev. Psychopathol. 2005, 17, 715–734. [Google Scholar] [CrossRef]
  28. Hu, S.; Lai, Y.; Valdes-Sosa, P.A.; Bringas-Vega, M.L.; Yao, D. How do reference montage and electrodes setup affect the measured scalp EEG potentials? J. Neural Eng. 2018, 15, 026013. [Google Scholar] [CrossRef] [PubMed]
  29. Delorme, A.; Makeig, S. EEGLAB: An open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J. Neurosci. Methods 2004, 134, 9–21. [Google Scholar] [CrossRef]
  30. Belouchrani, A.; Cichocki, A. Robust whitening procedure in blind source separation context. Electron. Lett. 2000, 36, 2050–2051. [Google Scholar] [CrossRef]
  31. McMenamin, B.W.; Shackman, A.J.; Greischar, L.L.; Davidson, R.J. Electromyogenic artifacts and electroencephalographic inferences revisited. NeuroImage 2011, 54, 4–9. [Google Scholar] [CrossRef]
  32. He, X.; Qin, S.; Yu, G.; Zhang, S.; Yi, F. Study on the Effect of Dalbergia pinnata (Lour.) Prain Essential Oil on Electroencephalography upon Stimulation with Different Auditory Effects. Molecules 2024, 29, 1584. [Google Scholar] [CrossRef]
  33. Sowndhararajan, K.; Kim, S. Influence of Fragrances on Human Psychophysiological Activity: With Special Reference to Human Electroencephalographic Response. Sci. Pharm. 2016, 84, 724–751. [Google Scholar] [CrossRef]
  34. Verstraeten, E.; Cluydts, R. Attentional switching-related human EEG alpha oscillations. Neuroreport 2002, 13, 681–684. [Google Scholar] [CrossRef]
  35. Travis, F. Temporal and spatial characteristics of meditation EEG. Psychol. Trauma Theory Res. Pract. Policy 2020, 12, 111–115. [Google Scholar] [CrossRef]
  36. Gasser, T.; Bächer, P.; Steinberg, H. Test-retest reliability of spectral parameters of the EEG. Electroencephalogr. Clin. Neurophysiol. 1985, 60, 312–319. [Google Scholar] [CrossRef]
  37. Chang, C.-C.; Lin, C.-J. LIBSVM: A library for support vector machines. ACM Trans. Intell. Syst. Technol. 2011, 2, 1–27. [Google Scholar] [CrossRef]
  38. Sachs, M.C. plotROC: A Tool for Plotting ROC Curves. J. Stat. Softw. 2017, 79, 1–19. [Google Scholar] [CrossRef]
  39. Jackson, A.F.; Bolger, D.J. The neurophysiological bases of EEG and EEG measurement: A review for the rest of us. Psychophysiology 2014, 51, 1061–1071. [Google Scholar] [CrossRef] [PubMed]
  40. Höferl, M.; Hütter, C.; Buchbauer, G. A pilot study on the physiological effects of three essential oils in humans. Nat. Prod. Commun. 2016, 11, 1934578X1601101034. [Google Scholar] [CrossRef]
  41. Watanabe, E.; Kuchta, K.; Kimura, M.; Rauwald, H.W.; Kamei, T.; Imanishi, J. Effects of Bergamot (Citrus bergamia (Risso) Wright & Arn.) Essential Oil Aromatherapy on Mood States, Parasympathetic Nervous System Activity, and Salivary Cortisol Levels in 41 Healthy Females. Forsch. Komplementärmedizin Res. Complement. Med. 2015, 22, 43–49. [Google Scholar] [CrossRef]
  42. Vieira, A.J.; Beserra, F.P.; Souza, M.C.; Totti, B.M.; Rozza, A.L. Limonene: Aroma of innovation in health and disease. Chem.-Biol. Interact. 2018, 283, 97–106. [Google Scholar] [CrossRef] [PubMed]
  43. Zhang, Z. Spectral and Time-Frequency Analysis. In EEG Signal Processing and Feature Extraction; Hu, L., Zhang, Z., Eds.; Springer: Singapore, 2019; pp. 89–116. [Google Scholar]
  44. Bazanova, O.M.; Vernon, D. Interpreting EEG alpha activity. Neurosci. Biobehav. Rev. 2014, 44, 94–110. [Google Scholar] [CrossRef]
  45. Neuper, C.; Pfurtscheller, G. Event-related dynamics of cortical rhythms: Frequency-specific features and functional correlates. Int. J. Psychophysiol. 2001, 43, 41–58. [Google Scholar] [CrossRef]
  46. Zatorre, R.J.; Jones-Gotman, M.; Evans, A.C.; Meyer, E. Functional localization and lateralization of human olfactory cortex. Nature 1992, 360, 339–340. [Google Scholar] [PubMed]
  47. Ackermann, P.; Kohlschein, C.; Bitsch, J.Á.; Wehrle, K.; Jeschke, S. EEG-based automatic emotion recognition: Feature extraction, selection and classification methods. In Proceedings of the 2016 IEEE 18th International Conference on e-Health Networking, Applications and Services (Healthcom), Munich, Germany, 14–16 September 2016; pp. 1–6. [Google Scholar]
  48. Rigatti, S.J. Random Forest. J. Insur. Med. 2017, 47, 31–39. [Google Scholar] [CrossRef] [PubMed]
  49. Awad, M.; Khanna, R. Support Vector Machines for Classification. In Efficient Learning Machines: Theories, Concepts, and Applications for Engineers and System Designers; Apress: Berkeley, CA, USA, 2015; pp. 39–66. [Google Scholar]
  50. Momade, M.H.; Shahid, S.; Hainin, M.R.b.; Nashwan, M.S.; Tahir Umar, A. Modelling labour productivity using SVM and RF: A comparative study on classifiers performance. Int. J. Constr. Manag. 2022, 22, 1924–1934. [Google Scholar] [CrossRef]
  51. Wang, Q.; Sourina, O. Real-time mental arithmetic task recognition from EEG signals. IEEE Trans. Neural Syst. Rehabil. Eng. 2013, 21, 225–232. [Google Scholar] [CrossRef]
  52. Sakalle, A.; Tomar, P.; Bhardwaj, H.; Bhardwaj, A. Emotion recognition using portable eeg device. In Proceedings of the International Conference on Artificial Intelligence and Sustainable Computing, Greater Noida, India, 22–23 March 2021; pp. 17–30. [Google Scholar]
  53. Yan, N.; Cheng, H.; Liu, X.; Chen, F.; Wang, M. Lightweight privacy-preserving feature extraction for EEG signals under edge computing. IEEE Internet Things J. 2023, 11, 2520–2533. [Google Scholar]
Applsci 15 04980 g001
Figure 1. Experimental procedure and overview of data analysis. (A) Experimental procedure overview diagram. “S” denotes the subject, with a total of 50 participants (n = 50). SEO refers to sandalwood essential oil, while BEO denotes bergamot essential oil. The EEG experiment consists of the following three stages: (1) the resting stage, (2) the recording of EEG signals, and (3) the completion of the questionnaire. The resting state group did not receive olfactory stimulation when recording their EEG signals (as opposed to the SEO and BEO groups), and no questionnaire was required. (B) Data analysis summary diagram. The “Channel Locations” module illustrates the electrode placement; “S” in the “EEG Raw Data” module signifies the subject count (n = 50), while in the “Data Set” module, it indicates the sample size (k = 100). Each cube symbolizes a distinct feature, with varying partitions highlighted in different colors. The abbreviations F, T, C, P, and O correspond to the frontal (F-ROI), temporal (T-ROI), central (C-ROI), parietal (P-ROI), and occipital (O-ROI) regions, respectively. Subscripts 1 through 7 represent the δ, θ, α1, α2, β1, β2, and β/α ratio, respectively. The EEG data from participants inhaling SEO were categorized as low arousal (label: 0), whereas data from BEO inhalation were classified as high arousal (label: 1).
Figure 1. Experimental procedure and overview of data analysis. (A) Experimental procedure overview diagram. “S” denotes the subject, with a total of 50 participants (n = 50). SEO refers to sandalwood essential oil, while BEO denotes bergamot essential oil. The EEG experiment consists of the following three stages: (1) the resting stage, (2) the recording of EEG signals, and (3) the completion of the questionnaire. The resting state group did not receive olfactory stimulation when recording their EEG signals (as opposed to the SEO and BEO groups), and no questionnaire was required. (B) Data analysis summary diagram. The “Channel Locations” module illustrates the electrode placement; “S” in the “EEG Raw Data” module signifies the subject count (n = 50), while in the “Data Set” module, it indicates the sample size (k = 100). Each cube symbolizes a distinct feature, with varying partitions highlighted in different colors. The abbreviations F, T, C, P, and O correspond to the frontal (F-ROI), temporal (T-ROI), central (C-ROI), parietal (P-ROI), and occipital (O-ROI) regions, respectively. Subscripts 1 through 7 represent the δ, θ, α1, α2, β1, β2, and β/α ratio, respectively. The EEG data from participants inhaling SEO were categorized as low arousal (label: 0), whereas data from BEO inhalation were classified as high arousal (label: 1).
Applsci 15 04980 g001
Applsci 15 04980 g002
Figure 2. Self-report scores of inhaling SEO or BEO. Participants evaluated the effects of inhaling sandalwood essential oil (SEO) or bergamot essential oil (BEO) on arousal (A) and valence (B) dimensions using a 1–9 scale, where 1 signifies low levels and 5 denotes neutrality. Statistical outcomes are sequentially displayed in descending order for the entire cohort, followed by female and male participants. There were twenty-eight female and twenty-two male participants in total.
Figure 2. Self-report scores of inhaling SEO or BEO. Participants evaluated the effects of inhaling sandalwood essential oil (SEO) or bergamot essential oil (BEO) on arousal (A) and valence (B) dimensions using a 1–9 scale, where 1 signifies low levels and 5 denotes neutrality. Statistical outcomes are sequentially displayed in descending order for the entire cohort, followed by female and male participants. There were twenty-eight female and twenty-two male participants in total.
Applsci 15 04980 g002
Applsci 15 04980 g003
Figure 3. Comparison of topographic maps and spectrum maps under different conditions. (A–C) Topographic comparison of the average power spectral density (PSD) between the three groups of subjects in the resting state (RS), those who inhaled sandalwood essential oil (SEO), and those who inhaled bergamot essential oil (BEO), where PSD increases from blue to red. (D,E) RS was compared with inhaled SEO or BEO, respectively. The average PSD between the two groups of subjects was subjected to a one-tailed paired t-test to obtain a topographic map of the p-value, with a significance level of 0.05. For (D), p < 0.05 indicates that the average PSD of RS is significantly smaller than that of the SEO group (α1, α2), and p < 0.05 indicates that the average PSD of RS is significantly larger than that of the SEO group (β/α ratio). For (E), the mean PSD of RS was significantly smaller than that of the BEO group (p < 0.05). (F,G) Topographic maps of the t-value based on the statistical test method used for (D,E). When t < 0, the topographic map color is red; the darker the color, the greater the difference, and vice versa. For F, the average PSD of RS is smaller than that of the SEO group (α1, α2) when t < 0, and the average PSD of RS is significantly larger than that of the SEO group (β/α ratio) when t < 0. For (G), the mean PSD of RS is smaller than that of the BEO group when t < 0. See Supplementary Tables S2 and S3 for specific p-values and t-values. (H,I) RS, SEO, and BEO are represented by three different colored lines, indicating intersubject α1 and α2 average PSD in different regions of interest (ROIs) under different conditions. The F-ROI, T-ROI, C-ROI, P-ROI, and O-ROI correspond to the frontal lobe area, temporal lobe area, central area, parietal lobe area, and occipital lobe area, respectively. The shaded areas of different colors represent the standard error (SE) of the mean PSD between subjects in different groups. A one-tailed paired t-test was performed on the average PSD between subjects in different ROIs for each frequency band of RS and SEO (marked in orange font) or BEO (marked in red font), with a significance level of 0.05. It is noteworthy that when p < 0.05, the font is bolded. * p < 0.05, ** p < 0.01. See Supplementary Tables S4 and S5 for specific t-values.
Figure 3. Comparison of topographic maps and spectrum maps under different conditions. (A–C) Topographic comparison of the average power spectral density (PSD) between the three groups of subjects in the resting state (RS), those who inhaled sandalwood essential oil (SEO), and those who inhaled bergamot essential oil (BEO), where PSD increases from blue to red. (D,E) RS was compared with inhaled SEO or BEO, respectively. The average PSD between the two groups of subjects was subjected to a one-tailed paired t-test to obtain a topographic map of the p-value, with a significance level of 0.05. For (D), p < 0.05 indicates that the average PSD of RS is significantly smaller than that of the SEO group (α1, α2), and p < 0.05 indicates that the average PSD of RS is significantly larger than that of the SEO group (β/α ratio). For (E), the mean PSD of RS was significantly smaller than that of the BEO group (p < 0.05). (F,G) Topographic maps of the t-value based on the statistical test method used for (D,E). When t < 0, the topographic map color is red; the darker the color, the greater the difference, and vice versa. For F, the average PSD of RS is smaller than that of the SEO group (α1, α2) when t < 0, and the average PSD of RS is significantly larger than that of the SEO group (β/α ratio) when t < 0. For (G), the mean PSD of RS is smaller than that of the BEO group when t < 0. See Supplementary Tables S2 and S3 for specific p-values and t-values. (H,I) RS, SEO, and BEO are represented by three different colored lines, indicating intersubject α1 and α2 average PSD in different regions of interest (ROIs) under different conditions. The F-ROI, T-ROI, C-ROI, P-ROI, and O-ROI correspond to the frontal lobe area, temporal lobe area, central area, parietal lobe area, and occipital lobe area, respectively. The shaded areas of different colors represent the standard error (SE) of the mean PSD between subjects in different groups. A one-tailed paired t-test was performed on the average PSD between subjects in different ROIs for each frequency band of RS and SEO (marked in orange font) or BEO (marked in red font), with a significance level of 0.05. It is noteworthy that when p < 0.05, the font is bolded. * p < 0.05, ** p < 0.01. See Supplementary Tables S4 and S5 for specific t-values.
Applsci 15 04980 g003
Applsci 15 04980 g004
Figure 4. Schematic diagram of hyperparameter optimization of the support vector machine (SVM) model. (A) Hyperparameter optimization for SVM-1 involves constructing features based on traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2). SVM-1 uses a grid search method and a validation set accuracy rate obtained through 5-fold cross-validation of the training set (Supplementary Figure S1) to identify the optimal hyperparameters. (B) For SVM-2 hyperparameter optimization, the β/α ratio indicator is introduced during feature construction. All other conditions remain the same as in SVM-1.
Figure 4. Schematic diagram of hyperparameter optimization of the support vector machine (SVM) model. (A) Hyperparameter optimization for SVM-1 involves constructing features based on traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2). SVM-1 uses a grid search method and a validation set accuracy rate obtained through 5-fold cross-validation of the training set (Supplementary Figure S1) to identify the optimal hyperparameters. (B) For SVM-2 hyperparameter optimization, the β/α ratio indicator is introduced during feature construction. All other conditions remain the same as in SVM-1.
Applsci 15 04980 g004
Applsci 15 04980 g005
Figure 5. Schematic diagram of hyperparameter optimization of the random forest (RF) model. (A) RF-1 hyperparameter optimization. RF-1 constructs features based on traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2). RF-1 uses a grid search method and uses the sum of the validation set accuracy and the area under the ROC curve (AUC) obtained by 5-fold cross-validation of the training set as the target value to find the best hyperparameters. The size of the circle represents the accuracy of the validation set, and the color mapping of the circle represents the value of the area under the ROC curve. The best observation point is marked by a red box. (B) RF-2 hyperparameter optimization. Compared with RF-1, the β/α ratio is introduced in the feature construction, and the other conditions are the same.
Figure 5. Schematic diagram of hyperparameter optimization of the random forest (RF) model. (A) RF-1 hyperparameter optimization. RF-1 constructs features based on traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2). RF-1 uses a grid search method and uses the sum of the validation set accuracy and the area under the ROC curve (AUC) obtained by 5-fold cross-validation of the training set as the target value to find the best hyperparameters. The size of the circle represents the accuracy of the validation set, and the color mapping of the circle represents the value of the area under the ROC curve. The best observation point is marked by a red box. (B) RF-2 hyperparameter optimization. Compared with RF-1, the β/α ratio is introduced in the feature construction, and the other conditions are the same.
Applsci 15 04980 g005
Applsci 15 04980 g006
Figure 6. Evaluation of test set indicators of different models. (A) Discriminant analysis-1 (DA-1), support vector machine-1 (SVM-1), and random forest-1 (RF-1) construct features based on traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2). DA-2, SVM-2, and RF-2 introduce the β/α ratio indicator when constructing features, and the other conditions are the same. Comparison of the six indicators of DA-1, SVM-1, and RF-1. The value range of these indicators is between 0 and 1. The closer to 1, the better the model performance (see Section 2.10 for details). (B) Receiver operating characteristic (ROC) curves of the six models. The area under the ROC curve (AUC) of each model is calculated. The dotted line represents the random guessing of the model. The value range of AUC is between 0 and 1. The closer to 1, the better the model performance.
Figure 6. Evaluation of test set indicators of different models. (A) Discriminant analysis-1 (DA-1), support vector machine-1 (SVM-1), and random forest-1 (RF-1) construct features based on traditional EEG frequency bands (δ, θ, α1, α2, β1, and β2). DA-2, SVM-2, and RF-2 introduce the β/α ratio indicator when constructing features, and the other conditions are the same. Comparison of the six indicators of DA-1, SVM-1, and RF-1. The value range of these indicators is between 0 and 1. The closer to 1, the better the model performance (see Section 2.10 for details). (B) Receiver operating characteristic (ROC) curves of the six models. The area under the ROC curve (AUC) of each model is calculated. The dotted line represents the random guessing of the model. The value range of AUC is between 0 and 1. The closer to 1, the better the model performance.
Applsci 15 04980 g006
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Fang, J.; Yu, G.; Liao, S.; Zhang, S.; Zhu, G.; Yi, F. Using the β/α Ratio to Enhance Odor-Induced EEG Emotion Recognition. Appl. Sci. 2025, 15, 4980. https://doi.org/10.3390/app15094980

AMA Style

Fang J, Yu G, Liao S, Zhang S, Zhu G, Yi F. Using the β/α Ratio to Enhance Odor-Induced EEG Emotion Recognition. Applied Sciences. 2025; 15(9):4980. https://doi.org/10.3390/app15094980

Chicago/Turabian Style

Fang, Jiayi, Genfa Yu, Shengliang Liao, Songxing Zhang, Guangyong Zhu, and Fengping Yi. 2025. "Using the β/α Ratio to Enhance Odor-Induced EEG Emotion Recognition" Applied Sciences 15, no. 9: 4980. https://doi.org/10.3390/app15094980

APA Style

Fang, J., Yu, G., Liao, S., Zhang, S., Zhu, G., & Yi, F. (2025). Using the β/α Ratio to Enhance Odor-Induced EEG Emotion Recognition. Applied Sciences, 15(9), 4980. https://doi.org/10.3390/app15094980

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop