Changes in Functional Connectivity of Electroencephalography While Learning to Touch-Type

Abstract

The functional brain connectivity of electroencephalography (EEG) data that was acquired during the process of learning how to touch-type using the Colemak keyboard distribution is analyzed in this paper. The partial directed coherence (PDC) of the EEG in alpha, beta, and gamma rhythms was used to assess the functional brain connectivity at different learning stages. As a result, connectivity patterns common to the volunteers of the learning process are found to be representative of underlying brain processes. In particular, functional connectivity within the alpha brain rhythm in low-difficulty learning tasks exhibits the greatest desynchronization in the parietal lobes, which may be an indication of good performance during those tests. Widespread increase in fronto-central brain connectivity in the alpha band during the high-difficulty lesson is shown as a reflection of refined attention allocation and effective motor program processing. Beta modulation during motor planning is also reflected through an increase in frontal functional connectivity, as well as repetition suppression by a decrease in gamma connectivity. Metrics from complex network theory were used to associate channels P4, F4, Cz, and C4 as relevant in processes such as the execution of motor sequences, cognitive performance, and focused attention. These results add insight to previous analysis performed on the same database and further prove the feasibility of monitoring a learning process with EEG.

1. Introduction

There is a growing emphasis in neuroscience on developing physiological and behavioral measures that objectively quantify the development of new skills. The emerging domain of neurocognitive phenomics aims to bridge cognitive psychology and brain processes [1], seeking to understand how learning and cognition manifest through physiological signals, particularly electroencephalography (EEG). Historically, spectral analysis of EEG has been instrumental in establishing neurophysiological relationships and studying the correlation between specific EEG patterns and mental activities.

Despite advances in understanding EEG data during various mental tasks, there has been limited focus on how the spectral properties evolve throughout the entire learning process. Past studies have generally examined static or pre- and post-learning states rather than tracking dynamic changes over time. Early investigations suggested a link between active learning phases and EEG desynchronization, but technical limitations constrained the ability to analyze continuous changes related to skill acquisition.

Motivated by these gaps, an increasing number of studies have explored the application of multivariate functional analysis for cognitive monitoring using EEG [2]. For example, the assessment of a learning process through EEG was proposed in [3]. There, EEG measurements were obtained, and the effect that training had on different brain rhythms was analyzed through the use of functional analysis of variance (FANOVA) of the EEG’s power spectral density (PSD). The main finding in [3] was that the power in $\beta$ and $\gamma$ brain rhythms decreased from the beginning to the end of a training lesson, and this effect was explained by the decrease in engagement and mental effort as the volunteers became more proficient in typing. Those findings contributed to the understanding of the learning process of touch-typing using the Colemak keyboard layout and provided insights into the relationship between brain activity and skill acquisition.

Similar observations were made in [4]. There, the effects of a sound quality visual feedback system (SQVFS) in violin learning were investigated. For that purpose, the EEG activity of participants with no previous violin playing experience was studied while they learned how to play the violin. The results in [4] showed that SQVFS led to a significant improvement in violin beginners’ ability to produce a stable sound using the bow. Significant $\gamma$ band desynchronizations were observed in beginners across blocks, showing correlation with the amount of improvement in the task. This suggests that the $\gamma$ band could be a potential biomarker of motor learning in violin beginners. Such results could be interpreted in the context of the temporal binding model, which associates the $\gamma$ band with the integration of information processed in distributed cortical areas. The demand for cognitive resources and information processing in complex tasks may lead to $\gamma$ band power enhancement, which may reduce as the task becomes automated. Furthermore, significant differences were found between the amount of $\gamma$ band desynchronization among the beginner groups in [4] and the volunteers in [3].

Up to this point, no evidence of significant participation of the $\alpha$ brain rhythm was found in [3] or [4]. Nevertheless, a decrease in $\alpha$ band power at posterior sites (parietal and occipital) has been linked to cognitive processes, and may be a reflection of cognitive load during learning tasks. In that line, larger $\alpha$ event-related desynchronization (ERD) has been associated with good-performing and skilled volunteers in comparison to novice ones [5]. Perhaps such an effect was noticed in the early stages of the Colemak touch-typing experiment, as 32.8% of the tests performed in [6] showed significant changes (increase or decrease) in $\alpha$ band power. Yet, such results were not regarded as valuable as the 61.9% and 72.5% significant changes observed in the $\beta$ and $\gamma$ bands, respectively. Furthermore, FANOVA-based analysis in [3] did not find significant involvement of the $\alpha$ rhythm in the learning process.

Analysis of functional brain connectivity using EEG data has gained popularity as different metrics have shown to be reliable indicators of the directed influences among neural signals [7]. Based on that, a closer look at the involvement of the $\alpha$, $\beta$, and $\gamma$ bands is proposed here for the data in the original experiment of learning to touch-type in Colemak. This analysis is highly relevant because it addresses the fundamental challenge of understanding how brain activity dynamically evolves during learning processes. Analyzing changes of functional connectivity during skill acquisition has practical implications for educational strategies, neuro-rehabilitation, and brain-computer interfaces. This research contributes to developing tools for real-time monitoring of cognitive engagement and learning efficiency, which are increasingly relevant in an era focused on personalized education and neurotechnology applications.

2. Materials and Methods

In this section, a concise overview of the experiment and its database is provided. Next, the theoretical background on the method of analysis is fully addressed, as well as how it is used to assess the functional connectivity from the EEG data. In-depth information about the EEG data is provided in [3], while the database itself is freely available online [8].

2.1. Preliminaries

For this study, a database comprised of the EEG measurements of ten volunteers (six female and four male, all of them right-handed) was used. The mean age of the volunteers was 29.3 years with a standard deviation of 5.7 years. All participants provided written informed consent before participating in this study. All experiments were conducted in an ethical and responsible manner. All volunteers took twelve lessons (one daily) of Colemak touch-typing. Those lessons are provided freely online at https://colemak.com/Typing_lessons (accessed on 2 October 2025). During each lesson, the volunteers were seated in front of a computer in a room free of noise and distractions at the Laboratory of Biomedical Signal Processing, Center of Research and Advanced Studies (Cinvestav), Monterrey’s Unit. Each lesson had increased difficulty in comparison to the previous one by adding new keys to the words to be typed. However, only at the fourth, eighth, and eleventh lessons, one, two, and three complete horizontal lines of keys in the keyboard were used, respectively. Therefore, for comparison purposes, those three lessons were assumed as representative of low-, medium-, and high-difficulty tasks. Note that EEG measurements were acquired only during those three lessons. The volunteers were asked to repeat each of the typing lessons five times, with resting intervals of 2 min between repetitions.

2.2. EEG Data Acquisition and Preprocessing

EEG measurements with fixed gain referenced to linked mastoids were acquired with gel-based electrodes at positions F3, Fz, F4, C3, Cz, C4, P3, POz, and P4 (according to 10–20 international reference system) using the B-Alert X10 wireless system from Advanced Brain Monitoring (https://advancedbrainmonitoring.com (accessed on 2 October 2025)), Carlsbad, CA, USA. All measurements were acquired at a sampling rate of 256 Hz.

Prior to the analysis, the raw signals are processed through B-Alert Live Software [9], which includes different signal analysis techniques that identify and decontaminate eye blinks, as well as identify and reject data points that were contaminated with electromyography (EMG) signals, amplifier saturation, and/or excursions attributable to movement artifacts. For a detailed description of the artifact decontamination procedures, see [10]. A summary of those techniques is presented next:

• Excursion and amplifier saturation: contaminated periods are replaced with zero values, starting and ending at zero crossing before and after each event.
• Spikes caused by artifacts are identified, and the signal value is interpolated.
• Invalid epochs: if more than 128 zero values are inserted for an overlay (1 s epoch with 50% overlap), the current epoch is excluded from analysis.
• EMG: For each EEG channel, overlays with exceeding power within a combination of high frequency (based on 70–128 Hz bins for each overlay) and low frequency (based on 35–40 Hz) are labeled as periods with excessive EMG contamination. If only one overlay has EMG, posterior analysis is based on the average of the remaining two overlays. If excessive EMG is detected in two overlays, the second is classified as EMG, and it is excluded from analysis.
• Electrooculogram (EOG) due to eye blinks: Identification of eye blinks in the EEG without the use of a reference EOG channel is achieved through wavelet transforms that deconstruct the fast component of the bipolar Fz-POz signal, then a regression equation is used to identify the EEG regions contaminated with eye blinks. Representative EEG preceding the eye blink is inserted in the contaminated region.

Finally, decontaminated data is filtered with a fifth-order Butterworth bandpass filter with cutoff frequencies at 0.1 and 50 Hz. The previously described artifact decontamination procedures have been widely validated in different studies, and they have been found to be well-suited for processing brain activity in real-time settings, especially those related to analyzing cognition (see, e.g., [10,11,12,13]).

Next, for each volunteer, the data is arranged in spatio-temporal matrices as follows:

$$X_{l,k}=\left[\begin{array}{cccc}{x}_{l,k,1}\left(1\right)& x_{l,k,1}\left(2\right)& \cdots & x_{l,k,1}\left(N_{l,k}\right)\\ x_{l,k,2}\left(1\right)& x_{l,k,2}\left(2\right)& \cdots & x_{l,k,2}\left(N_{l,k}\right)\\ ⋮& ⋮& & ⋮\\ x_{l,k,m}\left(1\right)& x_{l,k,m}\left(2\right)& \cdots & x_{l,k,m}\left(N_{l,k}\right)\end{array}\right],$$

(1)

where $l=1,2,3$ denotes low, medium, and high level of difficulty of the measured lesson, respectively, and $k=1,2,3,4,5$ is the time each lesson is repeated (attempt), for $m=1,\dots ,9$ channels, and $n=1,\dots ,N_{l,k}$ time samples. Note that the value of $N_{l,k}$ depends on the time volunteers took to complete each lesson. In general, $N_{l,k}>N_{l,k+1}$ as the volunteers improved in their ability to perform the task, but such improvement is smaller as the difficulty of the tasks increases (see [3] for details).

2.3. Proposed Method

Our proposed analysis is based on the partial directed coherence (PDC), which was introduced in [14] as a frequency-domain measurement of relationships (direction of information flow) between time series. The PDC has been used to assess functional brain connectivity through a rigorous determination of its statistical significance [15]. Let measurements in (1) be rewritten by defining

$$\mathit{x}\left(n\right)={[x_1\left(n\right),x_2\left(n\right),\dots ,x_M\left(n\right)]}^T,$$

(2)

where the subindices $\left\{l,k\right\}$ have been dropped for notational convenience. Then, $\mathit{x}\left(n\right)$ can be fitted to a multivariate autoregressive (MVAR) model such that

$$\mathit{x}\left(n\right)=\sum _{p=1}^P{A}_p\mathit{x}(n-p)+\mathit{e}\left(n\right),$$

(3)

where $A_p\in {\mathbb{R}}^{M\times M}$, for $p=1,2,\dots ,P$, are the matrices containing the $a_{m,p}$ coefficients (linear parameters) of the MVAR model of order P, and $\mathit{e}\left(n\right)={[e_1\left(n\right),e_2\left(n\right),\dots ,e_M\left(n\right)]}^T$, such that $e_m\left(n\right)$ is a white-noise input (uncorrelated error process) to the m-th channel. Next, the total transfer function from each noise source to each variable in the model can be determined by z-transforming (3):

$$\mathit{X}\left(z\right)=\sum _{p=1}^P{A}_p\mathit{X}\left(z\right)z^{-p}+\mathit{E}\left(z\right).$$

(4)

Therefore, the total transfer function $\mathit{H}\left(z\right)=\mathit{X}\left(z\right)/\mathit{E}\left(z\right)$ is given by

$$\mathit{H}\left(z\right)={\left[I-\sum _{p=1}^P{A}_p\mathit{X}\left(z\right)z^{-p}\right]}^{-1}=\mathit{A}{\left(z\right)}^{-1},$$

(5)

where I is an $M\times M$ identity matrix. Note that the total transfer function in the frequency domain is given by ${\mathit{H}\left(f\right)=\mathit{H}\left(z\right)|}_{z=e^{j2\pi f}}=\mathit{A}{\left(f\right)}^{-1}$.

Under these conditions, a measure of the direct causal relations (directional connectivity) of $x_j$ to $x_i$, for $\left\{i,j\right\}\in \left\{1,2,\dots ,M\right\}$ is given by the PDC defined as

$${\pi }_{i\leftarrow j}\left(f\right)={\displaystyle \frac{A_{ij}\left(f\right)}{\sqrt{{\mathit{a}}_j\left(f\right){\mathit{a}}_j^T\left(f\right)}}},$$

(6)

where $A_{ij}\left(f\right)$ and ${\mathit{a}}_j\left(f\right)$ are, respectively, the $i,j$ element and the j-th column of $\mathit{A}\left(f\right)$. In practice, only the magnitude of (6) is considered.

In this paper, PDC values are computed according to (6) per volunteer and per attempt by using the asympPDC package version 3.0.1, which consists of a collection of MATLAB/Octave routines and functions for the analysis of multiple time series, such as EEG. AsympPDC is freely available at https://github.com/asymppdc/asympPDC/releases/tag/v3.0.1 (accessed on 2 October 2025). Its latest version already includes the fast asymptotic PDC calculation algorithm proposed in [16].

While the analysis in [3] was performed by using the PSD values estimated by B-Alert’s software, here PDC values are computed from the preprocessed EEG data of each lesson attempt. For those calculations, data was fitted to MVAR models using the Vieira–Morf algorithm already implemented in asympPDC, with a maximum order allowance of $P=10$. However, models never exceeded $P=4$ (based on the Schwarz Bayesian optimality criterion), and the median model order was $P=3$. In order to keep some similarity to the original analysis in [3], PDC values are calculated on three-second windows, with a sliding-window overlap of one second, which is in the same way that B-Alert’s software estimates its PSD epochs.

Under those conditions, PDC values for each volunteer are computed at each attempt for the $\alpha$ band at $f=8,9,\dots ,13$ Hz, $\beta$ band at $f=14,15,\dots ,29$ Hz, and $\gamma$ band at $f=30,31,\dots ,40$ Hz. Next, each rhythm is averaged along its corresponding bandwidth in order to have a unique averaged representation for $\alpha$, $\beta$, and $\gamma$ bands, respectively. Note that, in all calculations, only those PDC values regarded as statistically significant ($p<0.05$) are considered. Finally, a comparison between the PDC values at the last and the first attempt of a lesson with a given difficulty is performed among all volunteers in all EEG measuring channels, and then those with the greatest differences are discussed.

3. Results

All data was preprocessed as indicated in Section 2.2, and PDC values were computed as previously explained in Section 2.3. The results are presented in the following forms:

(a)

A spatial representation of the largest changes in connectivities that are preserved after thresholding the differences in PDC values between the last and first attempt of a lesson. The threshold used for binarization was $\eta =5$% change in PDC value. Hence, changes greater than $\eta$ or less than $-\eta$ were considered as a relevant increase or decrease, respectively. Such binarization allows for a compact representation of the most significant connectivities for a posterior analysis through measures of network topology (see, e.g., [17]), in which EEG measuring channels are considered as network nodes;

(b)

Density distributions of the change (increase or decrease) of the in-degree and out-degree of the directed networks. Those measures of degree indicate the number of connections being directed towards a node or getting out of a node, respectively. Then, the density is simply computed by dividing the number of nodes presenting a specific degree by the total number of nodes.

Based on that, Figure 1, Figure 2 and Figure 3 show the changes that training had on functional brain connectivity for the $\alpha$ band at the three different levels of difficulty that the measured lessons had. Similarly, changes in $\beta$ are shown in Figure 4, Figure 5 and Figure 6, and for $\gamma$ in Figure 7, Figure 8 and Figure 9. A discussion of the results is presented in the next section.

4. Discussion

The main focus of this study was on brain rhythms, as their involvement in learning processes has been previously reported. Yet, the original analysis proposed in [3] did not reveal important involvement of the $\alpha$ band. It is possible that the choice of a PSD-based analysis turned out to be a limiting factor. For that reason, this study decided to treat the original EEG data (instead of the PSD estimated values) using a metric of directed connectivity based on the PDC for analysis. Next, a more detailed discussion of the changes in connectivity in each brain rhythm is presented.

4.1. $\alpha$ Rhythm

Parietal and occipital areas are known to represent sensory–motor integration and visual interpretation, and those processes are involved in the touch-typing task under study. Specifically, the reduction in functional connectivity in the parietal lobe during the low-difficulty lesson (see Figure 1) may be a reflection of the processing of perceptual information and motor behavior [18]. Furthermore, such reduction of functional connectivity in the parietal area supports the notion of an $\alpha$ desynchronization associated to good performance, as proposed in [5], given that the best performance that our group of volunteers showed was during the low-difficulty lessons, with a 30% average improvement in execution time in the last attempt of the lesson in comparison to the first one.

The overall response of our volunteers during the medium-difficulty training session, shown in Figure 2, reveals an increase in frontal to parietal connectivity that may be related to an activation of the default mode network (DMN). This is likely due to the activation of past outcomes in memory being evoked as a mechanism during the repetitive process of learning to touch-type. Such activation of the DMN has already been reported during other learning tasks (see, e.g., [19]).

The more widespread reduction of functional connectivity with the right parietal area observed when attempting the high-difficulty lesson (see Figure 3) could indicate that a larger neural network is involved in information processing, thereby facilitating and optimizing learning [20]. A proposed contributing factor to such widespread reduction could be related to increased task complexity [21]. On the other hand, the widespread increase in fronto-central brain connectivity during the high-difficulty lesson might be a reflection of refined attention allocation and effective motor program processing that the volunteers achieved by the final training lessons, which has already been reported by means of a co-activation of sensory–motor rhythm and fronto-central $\alpha$ power [22].

The results obtained for $\alpha$ rhythm connectivity suggest the interaction of different areas of the brain at that frequency during the learning process, while PSD-based analysis proposed in [3] could not even reveal $\alpha$ desynchronization during the learning process on the same data, even though this phenomenon has been previously reported [23].

4.2. $\beta$ Rhythm

Changes in functional connectivity in $\beta$ rhythm follow similar patterns to those in $\alpha$ but with added contribution from the left hemisphere. Specifically, during the medium-difficulty task (Figure 5), such contributions aim to modulate central cortical regions. This phenomenon is in line with previously seen $\beta$ modulation during motor planning [24].

The frontal modulation of $\beta$ during the high-difficulty task (see Figure 6) can be explained not only by means of the actual movement, but also by adaptation, in which extended motor practice leaves traces in the movement-related $\beta$ modulation of a subsequent motor test [25].

4.3. $\gamma$ Rhythm

As suggested in [4], the higher demand of cognitive resources posed by the medium-difficulty task is reflected in an increase in functional connectivity in the $\gamma$ band (Figure 8, which later seems to be highly reduced in the high-difficulty task (Figure 9) as the typing task becomes automated. This phenomenon has been referred to as repetition suppression where a decrease in neural activity upon repeated presentations of the same stimulus leads to memory encoding [26].

4.4. Distribution of Connectivity Degree

The distribution of connectivity’s degree revealed that overall changes mostly happened in high degrees (greater than four) for the in-degree, independently of difficulty and rhythm. Such a result indicates that changes in connectivity are more likely to happen in the form of a particularly increased functional dependency in a node that receives information from many other areas of the brain. Low degrees (one or two) are most seen in the out-degree, which is a reflection of all nodes being mostly equally likely to contribute to changes of connectivity during the task.

The most recurrent nodes with increased functional dependency during the learning process were the nodes Fz, C4, and P4. This last one presents consistent increments in all frequencies during the medium-difficulty task (see Figure 2, Figure 5 and Figure 8). Such behavior may be related to the fact that for the first time the typing task involved letters in the upper and mid-row of the keyboard, which produced an increased need for motor control from the right parietal lobe [27]. The same effect is not seen in the low-difficulty task as it involves only the middle row of the keyboard, nor in the high-difficulty task as going back and forth between rows is no longer a new task. The increments of connectivity in P4 seem to be coupled to a bilateral decrease in connectivity towards Cz. Such behavior has already been associated with changes in cortical activity during the period of execution of motor sequences [28].

Furthermore, a consistent increase in connectivity towards Fz for $\gamma$ rhythm (Figure 7, Figure 8 and Figure 9) may be evidence of reaching peak concentration during the task, as frontal lobe function is highly correlated with cognitive performance and focused attention [29]. Finally, increased connectivity in C4 during almost all tasks and frequencies may be related to an increase in attentional suppression [30].

4.5. Changes in Connectivity with Increasing Task Difficulty

An overall result that deserves attention is the greater change (increase or decrease) in functional connectivity as the difficulty of the task increases. This is relevant as some studies support the idea that task demand can affect measures of functional connectivity during the execution of a task [31], and the results presented here are significant because many studies until now have not considered the effect of task demand on measures of functional connectivity [32].

In our case, a greater number of functional connectivities change in $\alpha$ and $\beta$ rhythms as the difficulty of the task is higher, and for the $\gamma$ band, the statement is true when comparing low- and medium-difficulty tasks. Nevertheless, our results are representative only of very specific moments of the learning process, and a more detailed analysis that includes a measure of functional connectivity at every lesson is needed.

4.6. Limitations of This Study

The database used in this paper was obtained with a relatively small sensor array, which restricts spatial resolution. This limits the ability to accurately localize neural activity to specific brain regions and to comprehensively understand the spatial distribution of brain rhythms during learning. While our results need to be interpreted with caution due to the small sample size, and do not allow for population inference, they show consistency across participants.

Nevertheless, the proposed functional connectivity analysis has the potential of giving further insight into the role of $\alpha$ rhythm during the learning process that the previous PSD-based analysis could not reveal, which is in line with the current challenges in neurosciences [33]. Therefore, the proposed analysis, together with expanding the sensor array, could help elucidate the neural networks engaged during learning. For that purpose, this paper exemplifies the use of connectivity degrees as a metric of complex network analysis applied in the analysis of functional brain connectivity.

4.7. Future Work

Based on the results and discussions presented in the study, several open research questions and avenues for future work emerge:

1.

How do interactions between different EEG rhythms, such as $\alpha$, $\beta$, and $\gamma$ bands, influence the process of learning, and can analyzing these couplings provide more comprehensive insights into neurocognitive mechanisms? Future research could explore cross-frequency coupling measures during skill acquisition [34].

2.

How do individual differences, such as age, cognitive capacity, or prior experience, affect functional connectivity dynamics during learning? Personalized models and larger sample sizes could help tailor neurofeedback or educational interventions.

3.

How do the observed changes in functional connectivity relate to long-term retention and transfer of learned skills? The experimental setup used to obtain the database involved a finite number of lessons and repetitions, which may not capture the full dynamics of longer-term learning processes and non-stationary brain activity, thus limiting external validity for more complex or extended training scenarios. It is clear that longitudinal studies assessing EEG dynamics beyond immediate training could evaluate the neural correlates of durable learning outcomes.

4.

What are the neuroplasticity mechanisms underlying the observed decreases in specific rhythms like beta and gamma during learning, and can targeted interventions (e.g., neurofeedback, transcranial stimulation) enhance or accelerate the learning process?

5. Conclusions

This paper shows how EEG-based neurophysiological monitoring may be used to objectively assess and optimize learning and training processes. Specifically, it suggests that:

1.

EEG analytics can leverage the ability to monitor learners’ engagement and cognitive states in real-time, allowing for dynamic adjustments to training programs—such as modifying difficulty levels or providing targeted feedback—to improve skill acquisition and retention.

2.

The proposed methodology enables identification of individual differences in functional brain connectivity related to learning, facilitating tailored training approaches that cater to each learner’s neurophysiological profile, thereby increasing efficiency and effectiveness.

3.

Incorporating EEG-based measures of learning assessment provides an objective layer of evaluation—potentially leading to better understanding of how training translates to neurocognitive change.

4.

Insights from the study can inform the design of neuroadaptive learning platforms that autonomously adjust content based on the learner’s brain activity, optimizing resource allocation and reducing time-to-competency.

Overall, this research is aligned with creating neuroscience-driven assessment tools to support better decision-making in educational and workplace training environments, ultimately enhancing workforce skills development and organizational performance.