Singular spectrum analysis of near-infrared spectroscopy signal classification for mental arithmetic and rest state

1 Introduction

The brain–computer interface (BCI) allows human and computers to communicate directly. In contrast to traditional equipment (such as stylus, microphones, and keyboards), the BCI is a recent development (Stegman et al., 2020). Near-infrared spectroscopy (NIRS) is a non-invasive technique with advantages over electroencephalograms (EEGs) since it does not have any side effects and also has no need for the subject to be immobilized (Ferreri et al., 2014). Instead of passive BCI, active BCI users execute a mental task without using any external stimuli (Sereshkeh et al., 2018). Some of the most common examples of these mental tasks are motor imagery (Wierzgała et al., 2018), mental arithmetic (Choi et al., 2018), and word generation-P300 speller (Oralhan, 2019).

Several researchers have examined EEG signals for brain-state detection using mental arithmetic tasks versus during rest state. EEG signals were optimized based on Bayesian and K-nearest neighbor (KNN) using a 40-feature set (Sharma et al., 2021), capturing different cognitive spatiotemporal processes using functional magnetic resonance imaging (fMRI) and EEG (Bréchet et al., 2019). The technique proposed in this study is based on the NIRS signal; the idea behind NIRS is similar to that of fMRI, in that it is predicated upon the idea that increased oxygen consumption from brain activity causes an increase in cerebral blood flow because of neurovascular coupling. As a result, local oxyhemoglobin (HbO₂) changes, which NIRS may non-invasively detect. Most biological tissue, including skin and bone, is transparent to near-infrared (NIR) light; however, oxyhemoglobin (Oxy-Hb) and deoxyhemoglobin (Deoxy-Hb) are superior absorbers of NIR in the 700–900 nm range. This is the basic premise of fNIRS (Leeb et al., 2015). NIRS is a new approach for measuring the dynamics of cerebral blood volume that is intimately linked to brain neuronal activity (Shin et al., 2016; Sehgal et al., 2014).

Since time-series classification of Oxy-Hb and Deoxy-Hb parameters are transient and non-stationary (Liu et al., 2019), this study proposes singular spectrum analysis (SSA) for NIRS analysis. These non-stationary NIRS signals are analyzed by extracting the principal components of Oxy-Hb as a function of time. SSA has been proposed for multi-channel EEG processing and was found to be a good method for removing muscular artifacts (Petrantonakis and Kompatsiaris, 2018). Based on the Oxy-Hb pattern of the signal, in this study the signal is classified into two states of the brain: mental arithmetic and rest. Previous results observed, while taking into consideration parameters (Oxy-Hb and Deoxy-Hb) as features, accuracy of 86.96% (Petrantonakis and Kompatsiaris, 2018). A graph NIRS (GNIRS) technique has also been proposed based on the spatial information of NIRS, with an accuracy of 92.52% (Petrantonakis and Kompatsiaris, 2018). Research in this field is yet to investigate the use of SSA for NIRS analysis. Petrantonakis and Kompatsiaris (2018) made the first comparative study of currently available NIRS technologies.

In biomedical signal analysis, data complexity and privacy are becoming more significant factors, especially for applications based on NIRS and the BCI that involve sensitive neural data. Effective signal decomposition and feature extraction strategies are motivated by the large dimensionality, non-stationarity, and subject-specific variability of brain signals, which have been shown to have a considerable impact on classification performance and computing complexity. Popescu et al. (2021) explored privacy-preserving techniques (homomorphic encryption) for machine-learning-based EEG signal categorization. They discussed data privacy and computational difficulty in sensitive biological signal processing, which is comparable to issues with NIRS/BCI systems. Riess et al. (2024) provided insights related to privacy and complexity in biomedical signal classification by demonstrating how to combine (differential) privacy with learning from complex-valued biomedical signals utilizing federated techniques. Torkzadehmahani et al. (2022) outlined important strategies and trade-offs in data privacy and learning complexity in an organized summary of privacy-preserving AI techniques in biomedical applications (such as secure computation and federated learning).

In biomedical signal analysis, feature extraction is a critical step, as raw physiological signals often contain high-dimensional, noisy, and overlapping information that can obscure relevant patterns. Effective feature extraction transforms these complex signals into compact and discriminative representations, facilitating accurate classification and enhancing the interpretability of machine learning models. In related domains, advanced feature extraction techniques have demonstrated significant performance improvements. Cheltha et al. (2024) applied edge detection and backbone-modified Mask R-CNN to extract structural and spatial features from human motion data, enabling robust detection under occlusions. Khan et al. (2022) employed a fine-tuned AlexNet to extract hierarchical convolutional features from video frames which improved anomaly detection performance in crowded scenes. Similarly, Khan et al. (2023a) combined multiple acoustic features, including Mel-frequency cepstral coefficients (MFCCs), delta MFCCs, and spectral contrast, to enhance text-independent speaker identification using a Transformer model. These studies collectively highlight that carefully designed feature extraction, whether through handcrafted preprocessing or deep-learning-based representation learning, is essential for improving model robustness and accuracy, and the same principles can be effectively applied to biomedical signal processing tasks.

In our study, the high-dimensional nature of the input features presents a challenge in terms of data complexity, as redundant or noisy features can reduce model accuracy. Recent attention-based mechanisms have been shown to mitigate such challenges by selectively focusing on the most informative components of complex data. For example, Aung and Khan (2025) demonstrated that an enhanced U-Net hybrid model integrating multiple attention mechanisms significantly improves feature representation and segmentation precision for small and irregular pulmonary nodules in complex imaging backgrounds. Inspired by this, our feature selection process emphasizes the most discriminative features, ensuring effective learning even in complex signal environments. In addition to complexity, data privacy is a critical concern, particularly when handling sensitive physiological or personal information. Decentralized and privacy-preserving learning approaches, such as those illustrated by Khan et al. (2023b) in a smartphone recommendation system, allow models to learn without direct access to raw user data, protecting privacy while maintaining predictive performance. In our framework, similar strategies such as anonymization, federated learning, or local processing of sensitive signals could be adopted in future implementations to ensure that privacy is preserved without affecting model efficacy. By explicitly addressing both feature complexity and data privacy, our approach aligns with contemporary best practice in machine learning, ensuring that robust performance is achieved while adhering to ethical standards for sensitive data usage. In this study, we consider Oxy-Hb as a parameter and apply SSA and extract several features. Since NIRS is analogous to time series, the features were motivated based on the characteristic of the raw signal we passed from different preprocessing techniques, filtering at 0.1 Hz to remove artifacts and detrending operation to evaluate the best fit of line. Decomposition has advantages over other methods as it yields modal information in time-domain signals. SSA was used for decomposition to decompose signals into five different principal components, each of which was then analyzed. Each signal was analyzed on the basis of several features which were ranked using the chi-squared test, the highest-ranked features being used to train the model. Accuracy, precision, F1-score, and recall value have impactful meanings, but they fail to consider all four parameters of the confusion matrix at once. Therefore, we present our model efficiency based on Matthew’s correlation coefficient. The following key information in this study is listed as follows.

• We propose a filtering-detrending singular spectrum analysis (FD-SSA) framework tailored for near-infrared spectroscopy (NIRS) signals, enabling robust separation of principal components and improved characterization of brain states during mental arithmetic and rest.
• We extracted and selected discriminative features, such as power bandwidth, entropy, and complexity, from Oxy-Hb signals. A mean classification accuracy of 98.4% was attained, surpassing state-of-the-art techniques (79.4% and 86.9%).
• The proposed approach was validated using a variety of statistical measures (Matthews correlation coefficient: 0.889; precision: 0.968; F1-score: 0.966; recall: 0.963) which demonstrated both accuracy and reliability. This underscored the method’s potential for real-world brain–computer interface applications.

Beyond the technical improvements, the clinical applicability and practical significance of achieving high classification accuracy for mental arithmetic are substantial. The reliable distinction of cognitive states enables more responsive BCI systems, supporting applications such as cognitive workload monitoring, neurofeedback therapy, and assistive communication for individuals with motor impairments. In real-time deployment, our method could allow the continuous monitoring of and rapid adaptation to a user’s cognitive state, reducing false activations and improving system usability. Compared to existing approaches, this enhanced performance opens possibilities for BCIs to operate effectively in naturalistic environments, providing more precise control, reduced mental fatigue, and a more intuitive user experience, thus bridging the gap between laboratory research and practical clinical or everyday applications. The following sections comprise the remainder of this paper: Section 2 presents the methodology used in this study’s experimentation. The outcomes of the experimentation are discussed in Section 3. The conclusion is presented in Section 4.

2 Methodology

In this paper, we propose filtering detrending singular spectrum analysis (FD-SSA). This section describes our proposed approach for the detection of mental arithmetic and rest states. Figure 1 shows the overall framework of mental arithmetic and the rest states.

2.1 NIRS dataset

In this study, we consider a publicly available functional near-infrared spectroscopy (fNIRS) dataset consisting of eight healthy adult participants (three women, five men; mean age 26 years, standard deviation 2.8 years). Brain hemodynamic activity was recorded using a continuous-wave ETG-4000 system from Hitachi Medical Co., Japan, which provides multi-channel measurements of oxyhemoglobin (Oxy-Hb), deoxyhemoglobin (Deoxy-Hb), and total hemoglobin (Total-Hb) expressed in millimeters. The participants were instructed to perform a cue-guided mental arithmetic (MA) task which involved sequentially subtracting a single-digit number from a visually presented two-digit number as quickly and accurately as possible. Each task block lasted 12 s, followed by a 28-s rest period to allow hemodynamic responses to return to baseline. Each participant completed 18 to 24 trials of the task. The dataset is organized into individual MATLAB files for each participant, including recorded hemoglobin signals and trial information, making it suitable for studies on cognitive workload, brain–computer interface (BCI) design, and signal classification. The hemoglobin signals are stored in time-series matrices, where rows correspond to time samples and columns correspond to individual channels. Each trial is labeled to indicate the experimental condition—mental arithmetic (task) or rest (baseline). Additional variables include the sampling frequency (10 Hz), the number of trials, and trial onset/offset indices that segment the continuous recordings into discrete task and rest blocks. This dataset allows the straightforward extraction of trial-specific features, application of signal preprocessing (e.g., filtering and baseline correction), and trial-wise classification for cognitive task detection or BCI research (Bauernfeind et al., 2014).

2.2 Preprocessing of NIRS data

The NIRS signals were preprocessed using a four-stage pipeline designed to optimize the extraction of task-related hemodynamic responses. A second-order Butterworth high-pass filter with a cutoff frequency of 0.1 Hz was selected because it provides a smooth frequency response with minimal phase distortion, effectively removing slow baseline drifts and low-frequency physiological artifacts, such as Mayer waves, while preserving cognitive-related signal components. Linear detrending was applied to further remove any residual slow trends. To reduce high-frequency noise while maintaining the temporal fidelity of the hemodynamic response, a moving average filter with a window size of five samples was chosen. This size was empirically determined through pilot testing and aligns with standard practice in NIRS studies, balancing noise suppression with the preservation of rapid task-related fluctuations. These parameter choices were carefully optimized to ensure that the preprocessing pipeline enhanced signal quality without distorting the discriminative features critical for accurate cognitive state classification. The signal was preprocessed using the following four stages. where Z I is the input signal, M a x ( Z I ) is the maximum value of input signal, and M i n ( Z I ) is the minimum value of input signal. where μ X ( t ) is trend and σ X is the autocorrelation error.

• High pass filtering. Signals were filtered using a high pass filter at a cutoff frequency of 0.1 Hz to remove baseline artifacts. When employing proper cutting frequencies, high throughput filters are effective in eliminating NIRS contamination and removing artifacts from compound NIRS.
• Normalization reorganized data such that there was no redundancy of data, and it eliminated modification anomalies in a dataset from the data. This helped improve classification performance and reduce machine learning algorithm runtime. The proportion of features decreased using feature selection and feature weighting procedures are used to assess the influence of data normalization on feature relevance (Singh and Singh, 2020), as define in Equation 1,

Z N M = Z I − M i n Z I / M a x Z I − M i n Z I , (1)

• Detrending removes the trend’s effects from a data collection, revealing just the variations in values from the trend; this allows cyclical and other patterns to be discovered. Regression analysis and other statistical approaches can be used to detrend data. When addressing generally referred “baseline shifts,” the impact of detrending on ERPs and subsequent decoding becomes obvious, which are then often corrected by detrending analysis (van Driel et al., 2021), as define in Equation 2,

Z D T = μ X t + σ X t , (2)

• Data smoothing is a technique for removing noise from a dataset so that essential patterns may be perceived. It is used to forecast patterns, such as those in stock prices. The random approach and the use of moving averages are two alternative data-smoothing models. In several datasets, data smoothing produced the same or higher effort estimation accuracy than the non-smoothing scenario, whereas the standard outlier elimination method produced lower accuracy (Singh and Singh, 2020), as define in Equation 3,

Z S M = 1 n ∑ i = 0 n − 1 Z I N − i . (3)

Figure 2 shows preprocessed data for all the four stages as the signal after filtering in Stage 1 is normalized because of a change in amplitude, followed by detrending to remove certain trends from the signal. Finally, smoothing is performed using moving averages to remove noise.

2.3 Determining principal components using singular spectrum analysis

The process of SSA involves eliminating noise from a signal and breaking it down into multiple principal components, each of which has a different degree of significance based on the embedding dimension. If the series class is not too large, SSA is a model-free, exploratory technique that may be used in batch or real-time mode. It is also well-defined, allowing the rule for choosing the right parameters to be fixed. A time-series model must be defined to perform statistical testing (van Driel et al., 2021). Assuming Z(t) is a time-series signal, like in Equation 4,SSA is divided into two stages: decomposition and reconstruction. During the decomposition phase, singular value decomposition (SVD) is performed after a time-delay embedding technique. A one-dimensional time series is mapped into multi-dimensional space to form the Hankel/trajectory matrix (Widom, 1966) during the time-delay embedding stage (Equation 5), where M denotes the embedding dimension, which should be between 2 < M N 2 < K, and K denotes the indices of time-delay vectors, which should be between N − M + 1.

Z t : t = 1,2,3 … … … … . N . (4)

H a n k e l M a t r i x H M = z 11 z 12 z 13 z 21 z 13 z 23 z 13 z 32 z 33 . (5)

The trajectory matrix is decomposed into a sequence of rank-one basic matrices using SVD (Golyandina et al., 2018). The trajectory matrices in Equation 6 are not Hankel matrices after the grouping stage and must be rebuilt. Diagonal averaging is necessary to fit this matrix into the Hankel structure. To do this, we substitute the antidiagonal components of the trajectory matrix with their average. Finally, using the reverse embedding approach, we may reconstruct the original single-channel signal with a duration of N samples. We must apply the diagonal averaging and reconstruction steps to each sub trajectory matrix in order to decompose a single-channel signal into L decomposed signals, a process referred to as “elementary grouping”.where U is the left singular vector, D is the singular value, and V is the right singular vector. In this study, the signal was decomposed into five principal components that expressed zero correlation between them for all the subjects. The decomposed signals and their correlation plot are shown in Figures 3, 4, respectively. SSA was applied to the preprocessed NIRS signals for denoising and trend extraction, the embedding dimension (window length) M was set to 50, corresponding to approximately one-tenth of the total number of samples in each analyzed segment, following standard practice to balance resolution and noise suppression in non-stationary biosignals. The number of components L retained for reconstruction was selected based on the eigenvalue spectrum and cumulative explained variance, with the first five components capturing over 90% of the signal energy. These parameters were empirically optimized through a grid search on a subset of the data to maximize signal-to-noise ratio while preserving the hemodynamic dynamics relevant to mental arithmetic tasks. Sensitivity analysis confirmed that small variations in M (±10 samples) and L (±1 component) did not significantly affect classification performance, indicating that the selected parameters provide a robust decomposition for the NIRS signals used in this study.

A = U D V T , (6)

2.4 Feature analysis

Numerous features, including those from the time and spectral domains, comprising power bandwidth, approximate entropy, peak-to-peak, complexity, and mobility were retrieved and assessed using the chi-squared test, which determines whether two variables are independent of one another. This is useful for determining whether two distinct variables are correlated (Korenaga et al., 2019). Stronger independence is indicated by a larger chi-squared value, while greater variance between the two variables indicates less correlation.

The 14 features extracted were sorted according to the chi-squared test results. Due to the binary classification nature of the problem, goodness-of-fit was assessed by comparing categories from a random sample. Statistical features such as features 2, 3, 5, 10, and 11 were chosen for training. Table 2 shows the results of the chi-squared test acquired by all the features for classification between the binary of mental arithmetic and rest states. Different order moments, variance, and signal range were statistical variables that outperformed power bandwidth, Lyapunov exponent, and kurtosis. The chi-squared test is evaluated in Equation 7.where Z O i is the observed value and E i is the expected value.

x 1 2 = ∑ Z O i − E i 2 / E i , (7)

The scores of the chi-squared test are discussed next in Section 3.

3 Results and discussion

We used the proposed FD-SSA technique on an NIRS dataset. NIRS is a method for measuring the antagonistic hemodynamic response of oxygen levels in the brain with the parameters Oxy-Hb and Deoxy-Hb. All living entities significantly scattered near-infrared light by means of light emitting diodes and photodetectors placed 3 cm apart. The dataset contained eight individuals (three men and five women; average age 26 years, standard variation 2.8 years). An ETG-4000 continuous-wave device was used to monitor brain oxygenation. The multi-channel system recorded changes in Oxy-Hb, Deoxy-Hb, and Total-Hb in millimeters. The participants were prompted to do mental computations guided by cues. They needed to swiftly subtract a one-digit number from a two-digit number for 12 s (e.g., 97 − 4 = 93; 93 − 4 = 89), then rest 28 s. Participants 1–3 finished 18 MA trials, and participants 4–8 performed 24 trials (Pfurtscheller et al., 2010).

Table 1 reports statistics of the preprocessed signal by four stages. First, a high-pass filtered signal passing through a cutoff frequency of 0.1 Hz removed motion artifacts; it was used since larger drifts were present in the input signal. In the second stage, the signal was normalized because the data recording trials were distinguished for subjects; the number of trials was 18 for participants 1–3 and 24 for participants 4–8. The purpose of data normalization was to minimize and remove data duplication. The standard deviation changed from 0.1229 to 5.7260 after normalization. Normalization was followed by detrending to remove the effects of the trend from a data collection. Revealing just the deviations in values from the trend, it enabled the detection of cyclical and other patterns. The signal after detrending changed mean value from 3.0880 to −1.4480, removing the trends. Finally, smoothing was used as a data preprocessing technique to remove noise from a dataset. The overall range was changed from 5.7310 from 4.0660 and the standard deviation from 0.1229 to 0.8625 after preprocessing.

TABLE 1. Statistical parameters of pre-processed data.

As mentioned for region of interest (ROI)-1 in [14], the channel individual principal components accuracy is presented in Table 2 for all eight subjects, having an accuracy of five principal components along with a mean accuracy 91.91% and standard deviation of 2.42.

TABLE 2. Accuracy of all subjects in ROI-1 along with five principal components (PC1–PC5).

A receiver operating characteristics (ROC) curve displays the true positive rate (TPR, or sensitivity) against the false-positive rate for various categorization score thresholds (FPR, or 1 − specificity). Each point on a ROC curve corresponds to a pair of TPR and FPR values for a specific threshold value. By altering the threshold value, alternative pairings of TPR and FPR values may be discovered, and the pairs can then be used to construct an ROC curve. TPR rises fast with little change in FPR in the resulting ROC curve (Figure 5), and the area under the ROC curve (AUC) is also much greater (0.96). These ROC characteristics demonstrate that the KNN classifier could properly distinguish between mental arithmetic and rest states.

In this study, fine, medium, and coarse tree, and fine, medium, coarse, cosine, cubic, and weighted K-nearest neighbor (KNN) algorithms were used to classify NIRS data. The parameters used in the classifiers are shown in Table 3. The overall classification accuracy reported in Table 3 (first column) corresponds to the reconstructed signal obtained after the FD-SSA feature fusion step. The additional columns (PC1–PC5) are shown only to indicate the percentage variance and contribution of each principal component rather than classification results on individual components.

TABLE 3. Accuracy of all subjects in ROI-1 along with five principal components (PC1–PC5).

Time-domain and frequency-domain features were assessed using the chi-squared test to determine their relevance for the classification task. The chi-squared scores for all 14 features are reported in Table 4. Features with scores exceeding a threshold of 500 were considered significant for model training. Using this criterion, five features were selected: standard deviation, variance, peak-to-peak, variance of the first-order derivative, and variance of the second-order derivative. As shown in Table 4, these exhibit scores well above 597, while the remaining features score below 400, indicating minimal contribution. The threshold of 500 thus effectively separates highly discriminative features from weak or irrelevant ones. This selection reduces dimensionality, mitigates overfitting, and ensures that the classifier receives the most relevant information. Furthermore, empirical validation confirmed that including features below this threshold did not improve accuracy, supporting the choice of 500 as a principled cutoff for feature selection. These features capture complementary characteristics of the signal: standard deviation and variance describe the overall signal variability, peak-to-peak indicates the range of amplitude fluctuations, variance of the first-order derivative reflects the signal’s rate-of-change, and variance of the second-order derivative represents its acceleration or curvature. The selected features were then combined in various permutations to train the classification model, allowing identification of the combination that achieves highest accuracy. This approach ensures that the model leverages the most informative and discriminative features, effectively reducing dimensionality while maintaining classification performance.

TABLE 4. Chi-squared test values for all 14 features.

To rigorously evaluate model generalization and mitigate the risk of data leakage, we adopted an 80:20 train–test split at the subject level, wherein 80% of the participant data were randomly assigned to the training set and the remaining 20% were reserved for testing. By partitioning the dataset along subject boundaries rather than individual trials or temporal slices, we ensured that no subject-specific information influenced both training and testing phases, thereby preventing overfitting to individual hemodynamic patterns. The random assignment was repeated across multiple iterations to obtain statistically robust performance metrics, and the corresponding subject IDs for each split are provided in the supplementary material to facilitate reproducibility. This methodology allows the reported classification performance to more accurately reflect inter-subject generalization, which is critical for assessing the translational potential of NIRS-based cognitive state decoding in heterogeneous populations.

Figure 6 represents the chi-squared test score on which features were selected in different parameters to train the model by optimizing the time to train and observe the best feature set for the binary classification of states.

Table 5 shows the accuracy of ROI-1 and ROI-2 in the prefrontal cortex of the brain, obtaining mean accuracy of 97.03% and Std deviation of 1.21 for ROI-1 and mean accuracy of 97.57% and standard deviation of 1.60 for ROI-2.

TABLE 5. Accuracy of all subjects in ROI-1 and ROI-2.

The existing method results in a form of classification accuracy based only on three parameters, excluding the false-negative value which is also an important parameter of the confusion matrix. MCC considers the entire confusion matrix as one entity and calculates a score based on all four parameters of the confusion matrix. The reported MCC score along with other performance parameters are presented in Table 6.

TABLE 6. Model evaluation parameters with MCC score for ROI-1.

SSA has significantly improved accuracy for the binary classification of mental arithmetic and rest compared to the already existing method, with an increment of 11.5% in the previous GNIRS [15] method in Table 7. The mean accuracy reported in existing methods were 79.4% for the antagonistic pattern (Bauernfeind et al., 2011) and 87.27 for the GNIRS method (Petrantonakis and Kompatsiaris, 2018); however, in this study, the proposed mean accuracy for FD-SSA is 98.1%.

TABLE 7. Comparison of mean accuracy with existing methods and proposed FD-SSA method.

When assessing binary classification, MCC yields a more truthful and informative score than the accuracy and F1-scores. Therefore, it should be the preferred parameter for evaluating the model. This study obtained better accuracy, and a new model parameter of Matthew’s correlation coefficient is proposed for NIRS classification for mental arithmetic and rest states.

4 Conclusion

This research explored singular spectrum analysis (SSA) for decomposing fNIRS signals. Selecting an optimal combination of features and classifiers is critical for single-trial classification in fNIRS-based BCI mental arithmetic (MA) tasks. Among the features considered, statistical features significantly outperformed frequency-domain features, such as power bandwidth and Lyapunov exponent. The extracted features minimize misclassification costs and enhance Matthew’s correlation coefficient (MCC) score. The obtained MCC values ranged from 0.77 to 0.99 across subjects, with a mean of 0.88, highlighting the robustness of the proposed approach. Overall, the application of SSA combined with optimized feature-classifier selection substantially improved classification outcomes.

For future research, integrating adaptive feature selection strategies with hybrid decomposition techniques can further improve MA task classification performance. This can also extend to multi-channel MA task classification. Real-time fNIRS-BCI scenarios can be explored for practical neuro-rehabilitation and cognitive monitoring applications.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Ruisen Huang, Shenzhen Technology University, China

Reviewed by: Arfat Ahmad Khan, Khon Kaen University, Thailand

Mathi R, SASTRA University, India