Extraction of discriminative features from EEG signals of dyslexic children; before and after the treatment

Abstract

Dyslexia is a neurological disorder manifested as difficulty reading and writing. It can occur despite adequate instruction, intelligence, and intact sensory abilities. Different electroencephalogram (EEG) patterns have been demonstrated between dyslexic and healthy subjects in previous studies. This study focuses on the difference between patients before and after treatment. The main goal is to identify the subset of features that adequately discriminate subjects before and after a specific treatment plan. The treatment consists of Transcranial Direct Current Stimulation (tDCS) and occupational therapy using the BrainWare SAFARI software. The EEG signals of sixteen dyslexic children were recorded during the eyes-closed resting state before and after treatment. The preprocessing step was followed by the extraction of a wide range of features to investigate the differences related to the treatment. An optimal subset of features extracted from recorded EEG signals was determined using Principal Component Analysis (PCA) in conjunction with the Sequential Floating Forward Selection (SFFS) algorithm. The results showed that treatment leads to significant changes in EEG features like spectral and phase-related EEG features, in various regions. It has been demonstrated that the extracted subset of discriminative features can be useful for classification applications in treatment assessment. The most discriminative subset of features could classify the data with an accuracy of 92% with SVM classifier. The above result confirms the efficacy of the treatment plans in improving dyslexic children's cognitive skills.

Introduction

The ability to read and write is a fundamental skill that children must master in their early schooling years in order to succeed academically. These skills can be affected by any problem, which emphasizes the importance of studying dyslexia (Usman et al. 2021a). Scientists propose several definitions for dyslexia, including verbal or learning disability, principally related to reading despite having age-appropriate cognitive abilities, motivation, and receiving a satisfactory education (Snowling et al. 2020; Ulrike et al. 2020; Gialluisi et al. 2021). Dyslexia can occur because of a lack of phonemic awareness that causes problems with reading and spelling (Ziegler et al. 2020; Share 2021; Medina and Guimarães 2021). Recent magnetic resonance imaging (MRI) researches have confirmed decreased gray matter volume in temporoparietal and occipitotemporal areas in patients with developmental dyslexia (Ulrike et al. 2020; Usman et al. 2021b; Ostertag et al. 2021; Sihvonen et al. 2021; Kronbichler et al. 2008). Also, other structural irregularities were found in the cerebellum and lingual gyrus (Kronbichler et al. 2008). These structural abnormalities have shown a positive correlation with phonological processing and spelling performance (Kronbichler et al. 2008; Eckert et al. 2005; Pernet et al. 2009). Moreover, previous studies have employed various imaging methods like MRI, Positron Emission Tomography (PET) (Conant et al. 2020; Paulesu et al. 2014; Shaywitz et al. 2003), Magnetoencephalography (MEG) (Lizarazu et al. 2021), and Electroencephalography (EEG) to categorize this disorder (Ortiz et al. 2020; Bosch-Bayard et al. 2020; Erdogdu et al. 2019). As EEG recordings pose no radiation risks and are more cost-effective than other neuroimaging methods, it is a reliable method for acquiring data in various fields (Tuncer et al. 2021). Several EEG studies have focused on eyes open and eyes closed tasks, Event-Related Potentials (ERP) recording, and memory processing tasks (Bosch-Bayard et al. 2020; Christoforou et al. 2021; Fraga González et al. 2018; Stankova et al. 2019; Xue et al. 2020; Coch 2021). These studies revealed abnormal brain processes related to impairments of information processing in the visual and auditory modalities in dyslexic subjects. Previous researches have shown various changes in spectral features of the EEG signals like abnormal alpha activity in dyslexic children (Bosch-Bayard et al. 2020; Wang et al. 2017).

The majority of previous EEG studies in dyslexic cases have focused on finding differences between patients and controls by measuring EEG abnormalities (specifically, EEG power in individual frequency bands), although there have been a few recent studies that examined the changes in EEG that occurred before and after neurofeedback treatment (Eroğlu et al. 2020, 2018). Here the aim is to investigate dyslexic children before and after a specific treatment plan. This study investigates EEG signals' discriminative features related to a specific treatment process using tDCS and occupational therapy. In this research, EEG signals were recorded from dyslexic children both before and after the treatment. After gathering the data, various EEG features were investigated. In the proposed method, a combined method consists of the Principal Component Analysis (PCA) and Sequential Floating Forward Selection (SFFS) search algorithms has been used alongside the statistical analysis to extract a suboptimal subset of discriminative features after the treatment process. Finally, the discriminative features were used for classifying the subjects before and after the treatment. The classification stage results showed the strength of the proposed method in extracting the useful features. It is worth mentioning that this is, to the best of our knowledge, the first study that examined a wide range of EEG features for finding discriminative features for a group of dyslexic children undergoing treatment. The rest of this paper has been organized into four sections: First, in the Method and analysis section, subjects participated in this research are described. EEG data recording and preprocessing procedures are explained in the following. The method of extracting and selecting discriminative EEG features is described following by explanation of the classification process in the last section of the first part. In the results section, the discriminative features and classification results are presented. The obtained results are discussed in the discussion section. Finally, this paper is concluded in the conclusion section.

Method and analysis

Participants

Subjects in this study were 16 right-handed dyslexic children (7 females + 9 males) with reading and writing problems before treatment. The subjects attended the Atieh Clinical Neuroscience Center, with the initial reports of low school dictation scores. All subjects were native Persian speakers with normal vision and did not differ with regard to chronological age (Average = 7.76 ± 1.01), IQ (Average = 106 ± 13) test score. To mention, the subjects had no history of neurological disease or clinical reports of hyperactivity disorder. These subjects received a diagnosis of dyslexia, according to the Atieh Clinical Neuroscience Center using the Wechsler Intelligence Scale test (version-5) (Wechsler 1949) with aveage of 88.70 ± 13.6. Then, they underwent tDCS treatment and cognitive training. All participants’ families provided written informed consent and agreed with the treatment procedure, and the ethics committee of Iran Medical University approved the study. The participation process was tailored to be convenient for the children, and all subjects reported any possible negative side effects of the treatment process.

Treatment plan

Each subject attended 20 sessions over six weeks, with a minimum interval of two days between sessions. The tDCS is a form of neurostimulation where low levels of constant current are delivered to the brain's targeted areas, often producing great treatment results (Morya et al. 2019; Costanzo et al. 2019; Rios, et al. 2018). The subjects underwent tDCS treatment on the left anode in each session for 20 min. The current amplitude in each stimulation session was 1 mA using a pair of scalp electrodes (5 × 7 cm) made of conductive rubber and covered with saline-soaked synthetic sponges. The anodal electrode was positioned midway between T3 and T5. The cathodal electrode was placed midway between T4 and T6. For Occupational therapy, the BrainWare SAFARI (Kline et al. xxxx) was used. This program teaches the underlying skills that facilitate students to learn and demonstrate learning. The skills developed in this program were chosen because they are decisive in learning. The BrainWare SAFARI application offers various practices. In this study, according to the treatment plan, six practices were chosen for all the subjects. Each treatment session included occupational therapy, including ‘Visual Spatial Working Memory’, ‘Visual Sustained Attention’, ‘Visual Selective Attention’, ‘Visual Figure Ground’, ‘Visual Processing Speed’ and ‘Visual Motor Integration’. Subjects participated in three standard dictation tests akin to the dictation tests that students face in their school, including an average of 100 words with a general and understandable subject before and after the treatment. The dictation test average scores from the Atieh Clinical Neuroscience Center improved after the treatment, and the participants' reading and writing abilities after the treatment, in addition to the results of the Wechsler Intelligence Scale test, confirmed the effectiveness of treatment.

EEG recordings

The resting-state eyes-closed EEG was recorded from all participants in a specialized clinical environment before and after the treatment. The data was gathered for 5 min using a 19-channel EEG cap with a sampling rate of 500 Hz with a Mitsar EEG recorder, which has FDA and CE approval. Electrodes were positioned according to the international 10–20 system. Figure 1 shows the EEG channels’ spatial setup. Prior to each recording session, the operator adjusted the channels' impedance to ensure the best impedance.

Fig. 1.

Location of the EEG channels and tDCS electrodes (The anodal electrode midway between T3 and T5 and the cathodal electrode midway between T4 and T6.)

Preprocessing and segmentation

The EEG preprocessing was performed through The EEGLAB toolbox (Delorme and Makeig 2004). The data were inspected for each subject in time; then, the unfavorable parts were removed. Due to the noisy behavior of F7 and FP2 channels in all subjects, they were removed during preprocessing step making the total number of channels for each subject equal to 17. For excluding the low and high-frequency noises, a band-pass FIR filter was used to omit frequencies lower than 0.5 Hz and higher than 100 Hz from the data; moreover, a notch filter of 50 Hz with 5 Hz bandwidth was used reducing the line power noise. In the following step, Independent Component Analysis (ICA) (Comon 1994) was used to examine the extracted components and remove those related to artifacts. To specify artifact infected independent components, both the Multiple Artifact Rejection Algorithm (MARA) (Winkler et al. 2011) and the ICLable (Pion-Tonachini et al. 2019) were applied. Considering MARA and ICLable results, the power spectrum of the extracted components was visually inspected. Then the non-brain activity was removed from ICs so that there was no common information between variables in the processed data and those eliminated from the data. After preprocessing, EEG signals were segmented into 2-s windows, and the features were extracted for each window individually. After extracting different features for each window, wrapper feature selection methods were used to select satisfactory features for classification. Using wrapper methods creates feedback from the classification step based on the precision of results to choose the best possible features.

Feature extraction

A wide range of various features was extracted from the data to investigate the discriminative features of EEG signals before and after the treatment. Here, these features are briefly explained.

Spectral features

EEG contains various frequency bands particularly important to characterize different brain states. The sub-bands of interest are: delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–15 Hz), beta (15–30 Hz), and gamma (higher than 30 Hz). The relative Power Spectrum Density (PSD) can be obtained by dividing each frequency band's PSD by the total PSD of the whole frequency band. As it has been shown in previous researches, there is a relation between the power of individual frequency bands and subjects’ various conditions like the level of cognitive impairments, attention, or brain mental functioning intensity (Klimesch 1999; Aftanas and Golocheikine 2001; Hlinka et al. 2010). Recent researches have investigated the discriminative role of EEG power spectra between healthy and dyslexic children (Martínez-Briones et al. 2020). In this study, each frequency band's relative power was extracted as a potential discriminative feature. In addition to PSD, the coherence, which investigates the relationship between these signals in the frequency domain, has been measured (Bosch-Bayard et al. 2020).

Harmonic features

A large number of previous EEG-based studies have investigated harmonic features for various applications like sleep stage detection (XXX 2004; Khalighi et al. 2013). The first Harmonic feature is the central frequency which indicates the power intensity in each band, has been indicated in Eq. (1). Equation (2) shows the second harmonic feature which is the frequency bandwidth, indicating the distribution of power alteration for each frequency band. Finally, the third harmonic feature is the spectrum's value in the central frequency that shows the power spectrum density in each frequency band. Here, $p_{\mathit{xx}}(f)$ is the power spectral density in frequency $f$ and $f_c$ is the central frequency of each band. Also, $f_l$ and $f_h$ are the lower and upper limits of each individual frequency band. $f_{\sigma }$ is bandwidth which shows power distribution in each frequency band. $S_{f_c}$ indicates the power spectral density of central frequency in each individual band.

$$f_{\sigma }=\frac{{\sum }_{f_l}^{f_h}{\left(f,-,f_c\right)}^2{p}_{\mathit{xx}}(f)}{{\sum }_{f_l}^{f_h}{p}_{\mathit{xx}}(f)}$$ 1

$$S_{f_c}=p_{\mathit{xx}}(f_c)$$ 2

Slow wave index

For Delta, Theta, and Alpha frequency bands that are considered as slow waves, the slow-wave index can be defined in Eqs. (3) to (5):

$$\mathrm{Delta}\mathrm{Slow}\mathrm{Wave}\mathrm{Index}(\mathrm{DSI})=\frac{{\mathit{BSP}}_{\mathit{delta}}}{{\mathit{BSP}}_{\mathit{theta}}+{\mathit{BSP}}_{\mathit{alpha}}}$$ 3

$$\mathrm{Theta}\mathrm{Slow}\mathrm{wave}\mathrm{Index}(\mathrm{TSI})=\frac{{\mathit{BSP}}_{\mathit{theta}}}{{\mathit{BSP}}_{\mathit{delta}}+{\mathit{BSP}}_{\mathit{alpha}}}$$ 4

$$\mathrm{Alpha}\mathrm{Slow}\mathrm{Wave}\mathrm{Index}(\mathrm{ASI})=\frac{{\mathit{BSP}}_{\mathit{alpha}}}{{\mathit{BSP}}_{\mathit{theta}}+{\mathit{BSP}}_{\mathit{delta}}}$$ 5

ASI is a measure that has been developed to detect fluctuations of vigilance in daytime pharmaco-electroencephalogram studies (Jobert et al. 1994). Here BSP stands for Band spectral power of each frequency band.

Entropy

Applying the entropy to the EEG signals is a way to quantify the amount of uncertainty or randomness in the pattern, comparable to the amount of information in the signal. Better saying entropy represents the amount of irregularity in a signal and is defined as follows. The entropy of EEG signals has been used in a wide range of previous researches (Bruce et al. 2009; Phung, et al. 2014; Kang et al. 2019). Entropy can be calculated from Eq. (6).

$$Entropy=-\sum _{i=1}^N\frac{n_i}{n}\ln (\frac{n_i}{n})$$ 6

In equation number 6, $n$ indicates the number of samples in a selected period of signal, N is the number of bins in each period and $n_i$ is the number of samples in each bin. The size for each bin is equal to $\lfloor \sqrt{n}\rfloor$.

Phase related features

The first phase related feature to be investigated is the Phase Locking Index (PLI).

This criterion is used to measure the synchronicity of two signals. In equation number 7, which is the PLI formula calculation, the “sign” is the sign function and “$<>$” measures the mean value.

$$PLI=\left|⟨,s,i,g,n,\left[s,i,n,\left({\mathrm{\Delta }}_{\phi },(t_k)\right)\right],⟩\right|$$ 7

Here, ${\mathrm{\Delta }}_{\phi }$ is the phase difference between the two signals in time $t_k$.

EMD-based features

Empirical Mode Decomposition (EMD) breaks down a signal into a number of components without leaving the time domain called IMFs. In EMD, due to the decomposition algorithm, ${\mathit{IMF}}_i$ has a higher frequency than ${\mathit{IMF}}_{i+1}$. These IMFs form a complete and nearly orthogonal basis for the original signal (Eq. 8). It is a useful method for analyzing non-stationary and non-linear signals like EEG (Zhuang et al. 2017). In this study, the IMFs and various related extracted features were investigated. First, the time derivative of the signal was examined through Eq. (9). In the following equation, ${\mathit{IMF}}_i$ indicates the $i$th IMF, and $r$ is the residual of the signal.

$$x\left[n\right]=\sum _{i=1}^M{\mathit{IMF}}_i+r$$ 8

$$D_t=\frac{1}{N-1}\sum _{n=1}^{N-1}\left|{\mathit{IMF}}_i,(n+1),{-IMF}_i,(n)\right|$$ 9

In Eq. (9), ${\mathit{IMF}}_i(n)$ is the $i$th IMF at the moment $n$. Moreover, $D_t$ shows the first derivative of signal.

Secondly, the signal phase derivative, which measures the intensity of phase changes, was investigated. The Hilbert transform was used for measuring the phase of signal. Here, z(n) is the Hilbert transform of an IMF with the length of N, which A(n) is the amplitude and $\phi (n)$ is the phase, as illustrated in Eq. (10).

$$\mathrm{z}(\mathrm{n})=\mathrm{A}(\mathrm{n})e^{j\phi (n)}$$ 10

Equation 11 is the first derivative of the phase.

$$D_p=\frac{1}{N-1}\sum _{n=1}^N\left|\phi ,\left(n,+,1\right),-,\phi ,(n)\right|$$ 11

Finally, the normalized energy of the signal was measured for the available data. The Eq. 12 shows the normalized energy for each IMF of the signal $x$.

$$E_{\mathit{norm}}=\frac{{\sum }_{n=1}^N{\mathit{IMF}}^2[n]}{{\sum }_{n=1}^N{x}^2[n]}$$ 12

Feature selection

A machine learning project relies heavily on feature selection. Features selection is necessary for many reasons in a study. It allows the machine learning algorithm to train faster while reducing a model's complexity and making it easier to evaluate. If the right subset is selected, overfitting is reduced and accuracy is improved. In this study, the number of extracted features is much more in comparison with the available dataset. Reducing the dimensions in order to find the most critical features related to the treatment plan and increase the classifier accuracy, following the selection of the best group of discriminative features, a combined method of feature selection has been used.

Statistical test

The Kolmogorov–Smirnov test (KS-test) (Massey 1951) was used to extract the changed features after the treatment in the feature selection step. The p-value for selected features was less than 0.01.

PCA

To reduce the extracted features’ dimension, the PCA (Wold et al. 1987) was used. PCA is a technique for reducing such datasets’ dimensionality, increasing interpretability, and minimizing information loss. It does so by creating new uncorrelated variables that successively maximize variance (Jolliffe and Cadima 2016). PCA produces linear combinations of the original variables, features in this study, to generate the components, also known as principal components, or PCs. These PCs were selected in a way that covers 95% of the total input features variance.

SFFS

To choose the best subset of features, SFFS algorithm was used. The SFFS algorithm finds an optimum subset of features by insertions (i.e., by appending a new feature to the subset of previously selected features) and deletions (i.e., by discarding a feature from the subset of already selected features) to achieve the highest criterion (Ververidis and Kotropoulos 2008). The criterion employed in SFFS is the accuracy of the classifier. SFFS is directly used after PCA. Hence the SFFS selects the best subset of PCs with the highest accuracy rate.

In this pipeline, SFFS uses Support Vector Machine (SVM) with RBF kernel. Since the available data for the classification stage was limited, the Leave-One-Out cross-validation was used. In this method, each time, one data is used as test data, while others are used for training the classifier. Finally, the accuracy is reported as the average accuracy of all steps.

Back projection

After selecting the most desirable subset of PCs by using SFFS and SVM, it is time to return to the feature space. Each PC consists of a linear combination of features with a specific weight. In this study, these weights are used to select the features. The absolute values of these weights were summed and used to rank features. Here, the most compelling features in classification could be distinguished.

Classification

After distinguishing the most discriminative features, the data were classified with these features using an SVM classifier. For selecting the most effective subset, features were added to the classifier one by one sequentially according to their weight. In this way, the subset of features with the most accuracy in classification can be identified. For better demonstration, Fig. 2 shows the flowchart of the selected pipeline for this study.

Fig. 2.

Data processing pipeline

Results

Statistical analysis results

First of all, the performance of subjects examined after treatment. The Wechsler test results changed to the average of 96.61 ± 10.3, which shows a significant increase based on t-test. As mentioned in the previous section, the first step for finding the discriminative features was using statistical analysis. Here, 442 spectral-features, 425 phase-related features and 510 EMD-based features were analyzed using KS-test. Afterwards, 33 features with a p-value of 0.01 or less were selected. Spectral, phase-related, EMD, and entropy features shown meaningful differences for these cases before and after the treatment, and for spectral features, the differences of the delta, theta, alpha, beta and gamma band. Amongst these frequency bands, the beta band showed the most variations. Figure 3 shows the number of discriminative features for each individual frequency band after statistical test.

Fig. 3.

The number of features with a considerable change after the treatment for each individual frequency bands. As it is shown in this figure the majority of changes are related to the Beta band

Statistical analysis of the data from this study revealed that 13 of the 17 electrodes had exhibited different behavior after treatment. In the C3 electrode, the most changes have been observed. Figure 4 demonstrates the number of discriminative features for each EEG channel.

Fig. 4.

The number of features that have shown a significant difference (P-value of less than 0.01 for the KS-test) after the treatment for each EEG channel

Combined approach results

As explained in the feature selection section, following the selection of extracted features by means of the KS-test with a p-value of less than 0.01, PCA algorithm was performed to reduce the number of selected features selecting the best combination possible. PCA and SFFS were used to select the PC subset with highest accuracy. Considering the feature weights in PCs, 12 first features having the most significant weight among all features were selected from the 33 features according to their weights. Table 1 shows the 12 selected features for classification. Here, 25% of the discriminative features are related to EMD-based features, 33.33% of features are spectral features, 33.33% are phase-related, and 8.33% are entropy related features, which shows these features have valuable information in dyslexic patients’ treatment applications. In phase-related features, the most discriminative feature was the PLI between P3 and F8 channels showing a considerable increase after the treatment and got closer to the normal subjects (Seshadri et al. 2022). EMD-based features have been discriminated in first and second IMFs, which have a high-frequency. The results of the statistical analysis revealed that discriminative features mostly occurred in high-frequency bands.

Table 1.

The selected subset of discriminative features after performing PCA using the SFFS algorithm

Feature number (according to weights)	Channel	Feature
1	C3	Phase changes of first IMF
2	O1	Normalized power of the second IMF
3	Cz	Spectrum amplitude in central frequency of alpha band
4	C3	Time changes of first IMF
5	C3-P3	Phase synchronization
6	O1	Power spectrum in the central frequency of the beta band
7	Pz-F4	Coherence
8	C3	Power spectrum density of alpha band
9	Fp1	Entropy
10	C3	The relative spectral power of the delta band
11	P3-F8	PLI
12	Pz-T5	Phase synchronization

These features are arranged according to their weight so that feature number 1 has the highest weight and the feature number 12 has the lowest. As shown in the table the majority of the selected features are from EEG channels located on the occipital and partial regions

With the 12 first discriminative features with the highest weight, a classification process combining the leave-one-out method with an SVM classifier was used. As shown in Fig. 5 and mentioned earlier, features were added to the classifier one by one in the order mentioned in Table 1, the SVM classifier's best accuracy was achieved while using the 10 first features. It has to be mentioned that these numbers are related to the size of the data available in this study, and with recording more data from other subjects, these numbers might change.

Fig. 5.

The accuracy value for each subset of features. The variance of values is smaller than 0.07. As shown in this figure, the accuracy value increases by considering more features for classification. The highest value for accuracy is achieved by considering 10 features for classification before the accuracy decreases again by considering a higher number of features

Discussion

Dyslexia is a neurodevelopmental disorder that affects various populations across the globe. The disorder is often referred to as a learning disability and is distinguished by core cognitive deficits in the linguistic system, specific to phonological processing (Peters et al. 2020) and functions specific to the magnocellular visual pathways (Vilhena et al. 2021; Stein 2019; Mascheretti et al. 2021). Previous research has shown that dyslexia causes changes in some features extracted by EEG signals (Ortiz et al. 2020) compared to the normal group. In this study, in order to find the discriminative features the subjects were screened after and before a specific treatment plan. According to the statistical analysis results, spectral feature changes are related to all of the delta, theta, alpha, beta and gamma frequency bands. Most changes were related to beta bend. Also, the results from EMD-based features showed first IMFs were among the most discriminative features indicating the importance of higher frequencies. Previous studies have shown a higher power spectrum amplitude of EEG signals in healthy groups than the dyslexic subjects (Chiarenza 2021). This study reveales that the power spectrum has increased, especially in the beta frequency band after the treatment, and this observation correlates with the previous studies (Penolazzi et al. 2010). A considerable point about discriminative features in this study is that the results show that after the treatment, EEG features' behavior is becoming more similar to the normal behavior that is expected from a normal subject. This study shows a general global increase in PSD of EEG signals indicating that similar to healthy subjects, and the PSD is generally higher in dyslexic children after the treatment process. In addition, the results have shown a lower entropy in subjects after the treatment. It needs to be mentioned that previous studies have shown contrary results about entropy (Eroğlu et al. 2018). Moreover, the results shows an increase in synchronization between P3 and F8 channels. The PLI for P3 and F8 channels has increased. This piece of work supported the hypothesis that using tDCS stimulation alongside an occupational therapy plan can improve the condition of dyslexic children. Previous researches have shown that multiple sessions of left anodal tDCS treatment over the parietal-temporal regions that mediate phonological processing and phoneme-grapheme conversion (Jobard et al. 2007; Pugh et al. 2000; Pugh et al. 2001) combined with other treatments can be useful for improving the reading ability and efficiency in dyslexic patients due to the increase of neural excitability caused by tDCS (Rios, et al. 2018; Costanzo et al. 2016a; Costanzo et al. 2016b). Moreover, new evidence indicates that in addition to changes in neural excitability, other mechanisms contribute to the long-term outcomes of tDCS (Lang et al. 2011). Studies have exhibited a direct connection between tDCS and synaptic plasticity, wherein tDCS provokes long-term potentiation (Fritsch et al. 2010) and modulates postsynaptic connections (Meinzer et al. 2014; Stagg and Nitsche 2011).

It is well-known that EEG channels can represent the activity of different brain regions. According to the results of this study, most discriminative features are related to C3 and F3 channels, which can be related to the Broca area. Previous works have shown that these areas play an essential role in reading, writing, and language-related functions in the brain and abnormalities in dyslexic cases (Paulesu et al. 1996). This can be the result of occupational therapy (using the BrainWare SAFARI software). The main aim of using this software is to develop and strengthen children's cognitive skills that can help improve their condition. The results showed discriminative features related to the channels in the frontal region that is involved in cognition. Also, some of the discriminative features are related to the left parietal-temporal regions (i.e., Pz, P3, T5, and T3 channels), which is possibly the result of tDCS combination with the occupational therapy. This result follows data reporting evidence of left- parietal-temporal cortex involvement in phonological processing and grapheme-to-phoneme mapping in typical readers (Jobard et al. 2007; Costanzo et al. 2016a; Valdois et al. 2006). Previous neuroimaging studies have shown various findings, which can be explained by the variability of target brain regions' location among subjects (Müller-Axt et al. 2017; Cao et al. 2017). Dyslexia is a heterogeneous condition, and studies have shown variability in the location of neuroanatomical abnormalities in the human brain (Rios, et al. 2018).

Several important points need to be considered. First, the EEG data for each subject is limited to 19 (17 after pre-processing) channels, and therefore, the spatial resolution is low compared to high-density EEG recording setups. Having a higher number of EEG channels in future studies helps to have a higher spatial resolution and, therefore, reports the results concerning brain regions with higher accuracy. Secondly, the results of the feature extraction search algorithm and classifications are based on the available data, and recording more data from a higher number of subjects can help optimize the extracted subset of discriminative features related to the treatment in these subjects. Finally, as mentioned earlier, this study's main goal was to use a combined method of extracting the discriminative features related to a specific treatment plan used in a clinical environment and prove its performance for classification. The treatment for the subjects included both tDCS and occupational therapy using the BrainWare SAFARI software. This study did not have a sham group, and therefore it is challenging to attribute the results to tDCS or occupational therapy individually. It was impossible to run the processing pipeline using just tDCS or occupational therapy due to the agreement between the Atieh Clinical Neuroscience Center and the parents, and it was vital to run the original treatment procedure.

Conclusion

This study has shown that treatment of dyslexic children causes various changes in a wide range of features extracted from recorded EEG signals. It is demonstrated that by using feature reduction methods like PCA alongside SFFS algorithms, we can select a near-optimal subset of discriminative features according to the data available to classify the dyslexic subjects before and after the treatment with high accuracy. The results have demonstrated the value of EMD-based, Spectral, and Phase-related features information, especially in partial and posterior regions of the brain, in treating dyslexic children using tDCS and occupational therapy. The combined feature selection method used in this study highlights the discriminative EEG features that can help assess a treatment plan's effectiveness in dyslexic cases and other related applications.

Declarations

Conflict of interest

The authors declare that they have no conflict of interest.