The raw EEG signals were first filtered by a finite impulse response (FIR) band-pass filter between 0.5 Hz and 50 Hz to reduce noise and extract relevant information. Next, 15,000-point data (60s) for each trial in each channel were equally cut into 6 epochs of equal length, each containing 2,500-point data (10 s), as illustrated in Figure 3f. By the above operations, the data of each subject containing 20 trials were transformed into 120 epochs, of which 60 epochs corresponded to the labels of attention states and the other 60 epochs corresponded to the labels of non-attention states. To ensure the quality of the data, epochs with high amplitude or significant myoelectricity were removed from the dataset. After preprocessing, a total of 10,188 epochs were obtained from 85 subjects.

It is now generally accepted that advanced cognition in the brain is often associated with time–frequency and non-linear dynamic features of EEG (Klimesch et al., 1993, 1998; Chun et al., 2011). To comprehensively investigate the relationship between these features and attention states, we exhaustively enumerated and extracted time domain features, frequency domain features, and non-linear dynamic features from the preprocessed EEG data, as described in Table 1. To simplify notation, we utilized T_l, F_l, and D_l to represent a time domain feature, a frequency domain feature, and a non-linear dynamics feature, respectively, with the subscript l denoting the order of the feature in Table 1. Additionally, we used Tlm, Flm, and Dlm to signify the m-th dimension of the time domain feature T_l, frequency doamin feature F_l, and non-linear dynamic features D_l, respectively. For example, F43 denotes the third dimension of power spectrum for frequency bands F₄.

To valid the effectiveness of these extracted features, we used three common classification methods, including random forest (RF), decision tree (DT), and support vector machine (SVM) to build our attention recognition framework.

RF classifiers have been shown to be highly effective in small EEG data sets as demonstrated by the studies conducted by Amin et al. (2017), Lotte et al. (2018). In this study, the random forest classifier held 100 evaluators and used Gini impurity to measure the quality of a split as criterion. For DT, the criterion and splitter were set to information entropy and best. Linear kernel with a penalty parameter C of 2 and the kernel function coefficients of 0.2 was used for constructing the SVM classifiers.

We employed 5-fold cross-validation (CV) for intra-subject classification, dividing the training set and testing set strictly in the order of time. For inter-subject classification, leave-one-subject-out (LOSO) CV was employed.

In this section, we compared four types of features, namely time domain features, frequency domain features, non-linear dynamic features, and fusion features, in three different classifiers: RF, DT, and SVM. It should be noted that fusion features involve time domain, frequency domain, and non-linear dynamic features.

For each type of feature and each classifier, we calculated the mean accuracy based on a 5-fold CV approach, resulting in 85 mean accuracies in total. We then performed paired-sample t-tests to compare the accuracies of fusion features with those of the other types of features.

Figure 4 displays the mean and standard deviation of intra-subject accuracies for four types of features and three classifiers among 85 subjects. Fusion features demonstrate excellent performance, achieving accuracies (%) of 85.1, 78.7, and 79.8 using RF, DT, and SVM, respectively. For all classifiers (RF, DT, and SVM), the accuracies of fusion features are significantly greater than time domain features, frequency domain features, and non-linear dynamic features. RF performs the best among the three classifiers with different features, which is consistent with previous research findings that RF performs well on small datasets (Amin et al., 2017). The average accuracies (%) using RF are 81.4, 84.8, 84.0, and 85.1 for time domain features, frequency domain features, non-linear dynamic features, and fusion features, respectively.

In terms of inter-subject analysis, we fixed features and classifiers as those in the intra-subject analysis. For each feature type and each classifier, we obtained 85 accuracies using LOSO CV and performed paired-sample t-tests to compare the accuracies of fusion features with those of other types of features. Figure 5 presents the inter-subject results obtained using different types of features and different classifiers, showing that RF and SVM were found to perform relatively well. Using RF, the average inter-subject accuracies (%) are 75.6, 78.7, 78.1, and 80.0 for the time domain features, frequency domain features, non-linear dynamic features, and fusion features, respectively. The best result (81.6%) of the average accuracies is achieved by SVM for distinguishing fusion features across subjects.

Our study demonstrates the superior performance of our proposed method for the attention task, as compared to three baseline methods: PSD-SVM (Huang H. et al., 2021), Dynamical Complexity-XGBoost (Wan et al., 2021), and STFT-SVM (Acı et al., 2019). While PSD-SVM uses PSD features in the δ, θ, α, β, and γ bands and applies SVM for classification, Dynamical Complexity-XGBoost employs Multiscale Approximate Entropy, Sample Entropy, and Fuzzy Entropy as features and uses Extreme Gradient Boosting (XGBoost) for classification. Additionally, STFT-SVM utilizes the short-time Fourier transform (STFT) with a Blackman window to calculate smoothed time-dependent power spectra as features, which are then classified using SVM.

Our methods surpass the performance of baseline methods in both intra-subject and inter-subject classification, as evidenced by the results presented in Table 2. These findings demonstrate that our methods outperform the baseline methods.

To determine the effectiveness of features, we conducted an individual feature analysis by evaluating the classification performance. The results were calculated for intra-subject classification by RF and inter-subject classification by SVM, as depicted in Figure 6.

In the intra-subject experiment, numerous time domain, frequency domain, and non-linear dynamic features demonstrate an accuracy exceeding 70%. This finding suggests that the majority of the features calculated in Section 2 are effective within subjects. In the inter-subject experiment, only a few features exhibit a accuracy exceeding 70%, such as power spectrum for frequency bands F₄, wavelet entropy D₁₃, and Teager–kaiser energy D₁₄. These features demonstrate robustness across subjects, indicating their effectiveness as more reliable features. The highest inter-subject accuracy (%) achieved by an individual feature is 75.24 ± 11.53. However, our proposed fusion feature demonstrates a significantly higher accuracy (%) of 81.60 ± 9.93. This result further confirms the effectiveness of our fusion feature method.

First, we classified PSD features from different frequency bands (δ, θ, α, β, and γ) within and across subjects, respectively. The inter-subject and intra-subject average accuracies with features from different frequency bands are shown in Tables 3, 4, respectively. The average accuracies in Tables 3, 4 indicate that the PSD calculated from δ, θ, α, β, and γ bands exhibits distinct separability, which suggests that δ, θ, α, β, and γ oscillations of brain activity are related to the processing of attention states (Brown, 1970; Ray and Cole, 1985; Klimesch et al., 1993; Klimesch, 1999; Prinzel et al., 2001).

Additionally, time–frequency analysis with Morlet wavelets Cohen (2019) was employed in electrode position Fz in an experiment, as depicted in Figure 7. This figure illustrates the distinct patterns observed for different attention states. Notably, the analysis reveals that frequencies below 30 Hz exhibit significantly lower response energy during the attention state compared to the non-attention state. These findings further support the relationship between attention states and frequency bands such as α, β, and θ.

To further explore neural patterns associated with attention and non-attention states across all participants, we calculated the topographical maps of power features by averaging the power features over all epochs in all subjects for each frequency band between attention states and non-attention states. We then normalized the features by Z-Score for all epochs within each frequency band for each subject. Figure 8 depicts the topographical maps of the power features corresponding to the attention state an the non-attention state. The results demonstrate the existence of neural patterns associated with attention and non-attention states.

Although the neural patterns of attention states and non-attention states are similar, greater activation iobserved in the prefrontal areas for both. In the α, β, and θ bands, the lateral prefrontal and occipital areas exhibit less activation during attention states than non-attention states. During non-attention states, there are significant higher β responses in both prefrontal and occipital regions. The existing studies (Klimesch et al., 1998; Egner and Gruzelier, 2004) have showed that the changes in EEG features are closely related to the degree of attention, with varying degrees of amplitude and power of individual rhythmic brain waves. For instance, when participants were in the state of attention, their EEG signals exhibited a significant decrease in α and β waves (Prinzel et al., 2001). Conversely, during non-attention processing, the energy of β and α responses was increased. These findings on neural patterns are consistent with previous attention studies (Ray and Cole, 1985; Klimesch, 1999; Kelly et al., 2003; Swartwood et al., 2003; K Binienda et al., 2011).

In our study, we conducted an analysis of the differences in the different types of features across channels between attention and non-attention states at the group level. Specifically, we collected features extracted from each channel of different epochs for each subject. We then divided the features into two parts for attention and non-attention states, respectively. After removing outliers, we averaged each part to obtain the average features, which represented the average level of a specific feature for a given channel and subject for attention and non-attention states.

To analyze whether there were significant differences in the average features between attention and non-attention states at the group level, paired t-tests were implemented on the average features in the time domain, frequency domain, and non-linear dynamic features, respectively. It is assumed that the average features of attention and non-attention do not significantly differ at the group level. The p-value of 0.05 is used as a significance level. This means that if the p>0.05, the null hypothesis is true; otherwise false. The results of the t-tests in terms of time domain features, frequency domain features, and non-linear dynamic features are, respectively, shown in Figures 9–11.

Regarding the average features of time domain features, it was found that for most of the channels corresponding to the feature skewness T₄ and decorrelation time T₁₁, the p-values were much less than 0.05, indicating a relatively significant difference. This finding is consistent with the results of previous studies that have investigated the relationship between skewness and cognition in EEG signals Davis et al. (2020). Additionally, Figure 9 shows that prefrontal channels such as Fp1, Fp2, F7, F3, Fz, and F4 exhibit more significant features relative to other channels.

Compared to the average features of time domain features, the average features of frequency domain features performed better in the t-test. The p-values of the average features of F₄ (power spectrum density) of θ, α, β, and power spectral density ratio δ/γ, δ/θ, δ/α, δ/β, δ/γ, θ/α, θ/β, θ/γ, α/β, and α/γ were less than 0.05 in frontal, occipital, and temporal brain regions, as shown in Figure 10. These significant differences at the group level explain the superior performance of frequency domain features compared to time domain features in Figures 4, 5. This finding is consistent with previous studies that cognitive tasks can enhance the power of eeg, particularly in visual cortex Fitzgibbon et al. (2004).

As indicated in Figure 11, the average features of non-linear dynamic features are very special, and most of the non-linear features are significant for part of the channels. This finding is consistent with previous attention studies that have found that the value of entropy decreases with a decrease of attention states (Li et al., 2012). For instance, the features D₁₄ (Teager–kaiser energy) exhibit a highly significant difference between attention and non-attention states, as shown in Figure 11, which indicates that Teager–kaiser energy D₁₄ is an excellent feature. Furthermore, the first eight dimensions of Teager–kaiser energy exhibit more significant channels and a deeper degree than the last six dimensions. According to the content in Section 2, the first eight dimensions correspond to Teager–kaiser energy of the low-frequency portion of wavelet decomposition, while the last six dimensions correspond to Teager–kaiser energy of the high-frequency portion of wavelet decomposition. This highlights that the contrast between attention and non-attention states in low-frequency portion is much more pronounced than in high-frequency portion, which is consistent with previous research (Fiebelkorn and Kastner, 2019).

Taken together, some average features from time domain, frequency domain, and non-linear dynamics exhibit significant differences at the group level, which further validates the effectiveness of our framework for attention recognition based on EEG.

Brain functional connectivity (FC) elucidating the statistical dependencies and directed information flows unveils the functions and intricate interactions of diverse brain regions (Cao et al., 2022). We estimated correlations between different channels for each epoch of each participant to construct FC matrices and separately averaged the FC matrices of attention and non-attention states for each participant. This allowed us to obtain the average FC matrices for attention and non-attention states. We then conducted paired-sample t-tests and corrected using the false discovery rate (FDR) method for the average FC matrices corresponding to the two states. Figure 12 shows significant differences in brain FCs between attention and non-attention states across different frequency bands at the group level.

From Figure 12, it is evident that the significant differences in brain FCs between attention and non-attention states are primarily concentrated in the low-frequency bands, including δ, θ, lower α, and upper α (Zoefel et al., 2011) frequency bands. Within δ and θ frequency bands, there are significantly different connectivities distributed in various brain regions, including the left and right temporal lobes, parietal lobes, prefrontal regions, and occipital lobes. Within lower α and upper α frequency bands, the connectivities are mainly concentrated in the left and right temporal lobes. Overall, as the frequency range increases, there is a decrease in the number of significant connectivities, which further demonstrates that lower frequencies are more capable of characterizing changes in attention states. This is consistent with our discussion in section 4.1.

Furthermore, by comparing the differences in relatively low-frequency connectivities between the left and right hemispheres, we found that the number of significant connectivities in the right hemisphere at the group level was significantly higher than that in the left hemisphere. This indicates that the right hemisphere interacts more closely with information during attention changes, which is consistent with previous studies showing significant hemispheric asymmetry and lateralization toward the right hemisphere in the attention process of individuals (Bartolomeo and Malkinson, 2019).

In Section 3, we observed that neural patterns for attention and non-attention states exist according to Tables 3, 4, Figure 8. In this section, we analyzed these neural patterns of attention and non-attention states in greater detail.

Figure 13 shows the time–frequency analysis using Morlet wavelets for all epochs recorded from electrodes Fp1, Fp2, Fz, F3, F4, T3, T4, FT7, FT8, O1, Oz, and O2 in an experiment. As demonstrated by Figures 7 and 13, the time–frequency analysis reveals different patterns for different attention states. Specifically, the responses of low-frequency oscillations during attention states are lower than during non-attention states, especially in the temporal lobes and prefrontal regions. Additionally, the neural patterns remain relatively stable over time for each epoch within the experiment.

As shown in Figure 8, the average PSD in the θ, α, and β bands across all subjects exhibits distinct differences between attention and non-attention states, whereas this in the δ do not. This implies that the 4–30 Hz frequency bands are more closely associated with attention than the other frequency bands in EEG signals, which is consistent with prior research (Prinzel et al., 2001). For the δ band, compared with non-attention states, the neural patterns of attention states have a significantly lower response in the prefrontal. For the θ, β, and α bands, compared with non-attention states, the neural patterns of attention states have a significantly lower response in the lateral prefrontal and occipital areas.

As described in Section 1, the tasks in paradigms of attention recognition include the Stroop test, breath counting, and reading comprehension. The Stroop test is inherently short in duration (Kawashima et al., 2023), which makes it difficult for participants to sustain their attention over time. The task of counting the number of breaths can easily lead to mental wandering (Braboszcz and Delorme, 2011). Additionally, reading tasks are influenced by different materials, leading to variations in attention and concentration levels among individuals (Li et al., 2011). To address these limitations, we proposed mental arithmetic tasks as an attention task in our paradigm. Mental arithmetic tasks have a longer duration and are not affected by different materials. Our framework's results provide strong evidence for the effective characterization of attention and non-attention states under our proposed paradigm. First, the excellent classification results in Figures 4, 5 demonstrate that attention states and non-attention states can be distinguished accurately with our paradigm. Second, different activation maps in Figure 8 between attention and non-attention states further support the efficacy of our proposed paradigm in characterizing attention and non-attention states. Together, these findings highlight the potential value of our paradigm in advancing research and understanding of attention states.

First, our method can not only be used for intra-subject attention recognition but also for inter-subject attention recognition, making the application more convenient. Second, we can see that our proposed fusion feature method achieves the accuracy (%) of 85.05 ± 6.87 in intra-subject attention recognition and 81.60 ± 9.93 in inter-subject attention recognition. Compared with other attention classification methods, it achieves better classification results both within and across subjects, making the results of attention recognition more accurate.

There are three limitations to this study. First, this study analyzed the EEG features between attention states and non-attention states offline and did not perform some online validation. Thus, we will perform online validation for the framework presented in this study in future article. Second, in this study, all EEG channels (30 channels) were used to calculate the features, and it is difficult to collect 30 channels of EEG data for the complex variety of application scenarios of attention recognition (such as hospitals and schools). Therefore, reducing the number of channels for attention recognition is also an important direction for our future study. Third, in this study, for attention and non-attention states, we analyzed and classified the manual features based on EEG wthin and across subjects, and it is necessary to design an end-to-end network framework on large dataset in future study.