Cloud Hosting

Participants

Twenty-seven healthy right-handed volunteers participated in the study. Participants were recruited via social networks after they filled an online screening form for neurological and psychiatric disorders. Data from one individual were excluded from the analysis due to extensive EEG artifacts. The final study sample consisted of 26 participants (age 25.7 4.49, range 19-38; 12 females, 14 males) divided into two groups: M group (12 students or specialists with professional math education and experience), and H group (14 students or specialists in humanities). The principle of dividing individuals into groups based on education was as follows: the participants of the M group were either students of at least the third year of universities in mathematical specialties or working alumni of these universities. The same applies to the participants of the H group, but in humanitarian specialties (history, philology, law). This study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the Institute of Higher Nervous Activity and Neurophysiology of the Russian Academy of Sciences (protocol No3 from 24 August 2017). Volunteers have given written informed consent to their participation in the study, after the procedure was explained to them.

Participants were comfortably seated in a sound-shielded room, 1 meter away from a 19-inch square monitor. During EEG recording, participants were presented with tasks and were asked to mentally solve them, giving equal priority to accuracy and the shortest solving time. The tasks were presented in light gray in the center of a black screen, with the sizes of letters and digits being equal for all tasks. We presented three types of tasks (Table 2) in a pseudorandom order: 60 verbal, 60 arithmetic, and 60 logical tasks. The trial sequence consisted of task instructions (2 s), a fixation cross (0.5 s), the task (< 40 s), and a black screen during the participants response (4 s). Tasks were presented with a limited duration of 40 s. Participants clicked the left PC mouse button when they were ready to provide an answer. If no response was given within 40 s, the task was considered unsolved, disappeared from the screen, and the next trial began. The decision time (DT) was calculated between the task onset and the participants response. Additional long breaks were provided every 20-30 minutes or upon the participants request. The entire experiment typically lasted 2-2.5 hours.

The EEG was recorded using a 128-channel Geodesic Sensor Nets (Electrical Geodesics, Inc (EGI), Eugene, OR, USA) system based on the 1010 electrode montage. The recordings were band-pass filtered with a 0.1 Hz70 Hz analog filter, notch filtered at 50 Hz, and sampled at 1000 Hz with online re-referencing to the average using Net Station software. Impedance was kept below 50K(Omega).

All preprocessing was done using MNE software50. We excluded 46 skirt channels (defined as channels with EGI polar coordinate r > 0.5) near the periphery of the EEG net that are particularly sensitive to noise and muscle artifact. The remaining 83 channels retained for analysis is provided in the supplementary materials and labeled with their EGI channel number.The number of discarded channels was consistent across conditions and participants.

Remaining data were downsampled to 250 Hz. EOG artifacts were removed using automatic ICA in MNE. ICA components were found on high-passed filtered at 1 Hz signal and applied to unfiltered signal. The EEG data were analyzed in epochs of 2 s without overlap starting 5 s after the presentation of each task. As we were more interested in the recognition of mental operations during task solving rather than in the visual perception of the tasks, we assumed that during the first 5 s of cognitive task presentation, EEG could reflect visual retrieval and early (stimulus-driven bottom-up) cognitive stages related to semantic and lexical numerical representations of the stimuli. EEG, accompanied by the visual perception of complex visual stimuli such as math and verbal tasks, has a strong impact on eye-movement activity. Epochs with low signal-to-noise ratio were rejected using minimum and maximum peak-to-peak amplitudes. 94.12.2 % of epochs remained after preprocessing. Following the epoch rejection, there were no statistically significant differences in the number of epochs between the groups (M and H) and conditions.

A mixed-model ANOVA (GROUP, TASK) was applied to compare results on behavioral performance between the groups and 3 types of the tasks. As the distribution of behavioral variables deviated from normal we applied cluster based permutation tests with Spearman coefficient to perform a correlation analysis of EEG patterns and behavioral performance. We examined the mean PSD value of 96-channel data for each frequency range in correlation with behavioral results.

We performed cross-subject group classification to recognize the group type to which the individual EEG data belonged. Additionally, we compared the performance of three pipelines: supervised and unsupervised projections with Logistic Regression and handcrafted power features with LightGBM.

The frequency bands used for feature estimation were as follows: (theta)1 (theta1): 46 Hz, (theta)2 (theta2): 68 Hz, (alpha)1 (alpha1): 810 Hz, (alpha)2 (alpha2): 1012 Hz, (beta)1 (beta1): 1216 Hz, (beta)2 (beta2): 1620 Hz, (beta)3 (beta3): 2024 Hz. For the first two pipelines, we decomposed EEG into these bands using a set of filter banks with Butterworth Bandpass filters, and for each epoch in each band, the covariance spatial matrix was estimated. The feature space before projection and vectorization was ({varvec{X}}in {{mathbb {R}}}^{Ktimes Ntimes N}), where K is the number of frequency bands and N is the number of EEG channels.

For the third pipeline with handcrafted features, power spectral density (PSD) for each epoch and frequency band was estimated using the multitaper method51 that calculates spectral density for orthogonal tapers and then averages them together for each channel. Relative power, equal to the power in each frequency band divided by the total power, was used to form the feature space for classification. Thus, the feature space in this pipeline was ({varvec{X}}in {{mathbb {R}}}^{Ntimes K}), where K is the number of frequency bands and N is the number of EEG channels.

In the space of symmetric positive-definite (SPD) matrices29, the covariance matrices of the CSP filtered signal will take the form of (1), where ({varvec{W}}in {{mathbb {R}}}^{Ntimes J}) and (J le N) is the number of CSP filters sorted by decreasing eigenvalues. Thus, the feature vector will take the form of (2) with the number of components equal to J, where ({varvec{Sigma _{1}}}) is the mean covariance matrix of the first class.

$$begin{aligned} {varvec{Sigma _{{varvec{Z}}_{i}}}}= & {} {varvec{W^{T} Sigma _{{varvec{X}}_{i}} W}} end{aligned}$$

(1)

$$begin{aligned} {varvec{F}}_{i}= & {} begin{bmatrix} log ({varvec{Sigma _{Z_{i}}}}left[ 1,1right] ) - log ({varvec{Sigma _{1}}}left[ 1,1right] ) \ vdots \ log ({varvec{Sigma _{Z_{i}}}}left[ J,Jright] ) - log ({varvec{Sigma _{1}}}left[ J,Jright] ) end{bmatrix} end{aligned}$$

(2)

In the space of SPD matrices52, the covariance matrix of the signal projected with Principal Component Analysis (PCA) will take the form of (3), where ({varvec{W}}in {{mathbb {R}}}^{Ntimes J}) and (J le N) is the number of PCA components sorted by decreasing eigenvalues. Therefore, the minimal representation of a matrix in the Riemannian space will take the form of (4), where ({varvec{F}}{i}) is a feature vector with the number of components equal to (J*(J+1)/2), and (overline{{varvec{Sigma }}}{Z}^{-1/2}) is the mean of (3) matrices according to the Riemannian metric.

$$begin{aligned} {varvec{Sigma }}_{{Z}_{i}}= & {} {varvec{W}}_{UNSUP}^{T} {varvec{Sigma }}_{{X}_{i}}{varvec{W}}_{UNSUP} end{aligned}$$

(3)

$$begin{aligned} {varvec{F}}_{i}= & {} {Upper}(log (overline{{varvec{Sigma }}}_{Z}^{-1/2}{varvec{Sigma }}_{{Z}_{i}}overline{{varvec{Sigma }}}_{Z}^{-1/2})) end{aligned}$$

(4)

Logistic regression is a linear model employed to estimate the likelihood of a specific class, and it can serve as a supervised binary classification algorithm. To avoid overfitting, logistic regression was trained with (L_2) regularization, and the regularization parameter for (L_2) was specified within the range of (left[ 1e^{-10},1e^{9}right]).

LightGBM is a gradient boosting framework that uses a decision tree algorithm with leaf-wise split, and is known for its high performance. To optimize its performance, a cross-validation grid search was used to tune its hyperparameters.

To perform subject-independent classification of the participant group (M versus H), EEG epochs were labeled according to the group to which the subject belonged. For testing, epochs from two randomly selected participants from different groups were chosen, while epochs from two randomly selected participants (one from each group) were chosen for validation. The remaining participants epochs were used for training. This process was repeated ten times, resulting in ten different folds for each type of task.

Balanced accuracy (BA), receiver operating characteristic (ROC) curve and area under the curve (AUC) were used to assess the performance of the models.

One advantage of linear models is their interpretability, which allows for the identification of the strength and direction of specific effects in the features30.

In classification tasks, the backward model transforms the feature space ({varvec{X}}{i} in {{mathbb {R}}}^{Ntimes 1}) into a new representation that maximizes the discriminability between the two classes using the filter ({varvec{W}}in {{mathbb {R}}}^{Ntimes 1}) as shown in Eq. (5). On the other hand, the forward models in Eq. (6) describe sample generation as a multiplication of the activation pattern ({varvec{A}} in {{mathbb {R}}}^{Ntimes 1}) by the factor (s_{i}). The activation factor can be obtained using Eq. (7), where the covariance matrix is ({varvec{Sigma _X}} = mathbb Eleft[ {varvec{X}}{i},{varvec{X}}{i}^{T}right] _{i}).

$$begin{aligned} {{varvec{W}}}^{T}{varvec{X_{i}}}= & {} {hat{s}}_{i} end{aligned}$$

(5)

$$begin{aligned} {varvec{x_{i}}}= & {} s_{i}{varvec{A}}+{{varvec{varepsilon }}_{i}} end{aligned}$$

(6)

$$begin{aligned} {varvec{A}}= & {} {varvec{Sigma _{{varvec{X}}} W Sigma _{hat{{varvec{s}}}}^{-1}}} = {varvec{Sigma _{{varvec{X}}} W}} = Covleft[ {varvec{X}}_{i}, s_iright] end{aligned}$$

(7)

When using CSP, the feature space takes the form of ({varvec{X}}_{i} in {{mathbb {R}}}^{NKtimes 1}) and the filter takes the form of ({varvec{W}} in {{mathbb {R}}}^{NKtimes 1}). This filter has full column rank, which is proven by using the Sylvester rank inequality in Eq. (8).

$$begin{aligned} rank({varvec{W_{1}}}) + rank({varvec{W_{2}}}) - M le rank({varvec{W}}) le min(rank({varvec{W_{1}}}),rank({varvec{W_{2}}})) end{aligned}$$

(8)

For examining the feature importance in the case of LightGBM, we estimated the SHapley Additive exPlanations (SHAP) values32. SHAP values assign an importance value to each feature in a model. Features with positive SHAP values positively impact the prediction, while those with negative values have a negative impact. The magnitude is a measure of how strong the effect is.

A cluster-based permutation test53 was used to investigate differences in EEG power spectral density (PSD) between groups. A two-sided T-statistic with a threshold of 6 was applied and corrected for multiple comparisons using N=1024 permutations. Cluster-level correction based on spatial adjacency was also performed.

See the article here:
Detecting cognitive traits and occupational proficiency using EEG and statistical inference | Scientific Reports - Nature.com