EEG Signal Analysis of Writing and Typing between Adults with Dyslexia and Normal Controls

EEG is one of the most useful techniques used to represent behaviours of the brain and helps explore valuable insights through the measurement of brain electrical activity. Hence, plays a vital role in detecting neurological conditions. In this paper, we identify some unique EEG patterns pertaining to dyslexia, which is a learning disability with a neurological origin. Although EEG signals hold important insights of brain behaviours, uncovering these insights are not always straightforward due to its complexity. We tackle this using machine learning and uncover unique EEG signals generated in adults with dyslexia during writing and typing as well as optimal EEG electrodes and brain regions for classification. This study revealed that the greater level of difficulties seen in individuals with dyslexia during writing and typing compared to normal controls are reflected in the brainwave signal patterns


II. EEG Signal Acquisition
The EEG headset used for this research was the Cognionics 32-channel dry EEG headset, and the EEG was recorded at a sampling rate of 300Hz.The EEG channel map is depicted in Fig. 1 where the channels used on this specific EEG headset are indicated in grey.This research was carried out with a total of 32 participants, where 17 participants were individuals with dyslexia (7 males and 10 females) and 15 participants were normal controls (8 males and 7 females).The number of participants was determined using the Altman's Nomogram sample size calculation as shown in Fig. 2. Therefore, for a power of 0.80 (p-value significance of 0.05) and a standardised difference value between 0.8 and 1.0 (Cohen's d effect size), the total amount of participants would range between 30-50.Hence, the number of participants per group would range between 15-25.The inclusion and exclusion criteria of participants was 18 years and above, right-handed, fluent in English, have a normal or corrected-to-normal vision and normal hearing.The participants in the group with dyslexia had to be diagnosed by a psychologist as having dyslexia and the control group had to be free from motor and neurological conditions such as dyslexia, ADHD and autism.The participants with dyslexia were recruited with the help of DSF Literacy and Clinical Services in Western Australia (The Dyslexia-SPELD Foundation).The participants were given a simple writing and typing task, which was designed similar to the standardised psychometric tests used in the dyslexia diagnosis process under the supervision of a psychologist specialised in dyslexia assessments.The EEG was acquired while the participants were performing the task with the EEG headset setup on his/her head.The EEG device was wirelessly paired to a computer which had the EEG data acquisition software installed.The EEG was also recorded in the relaxed state where the participants were instructed to stay seated and relaxed with their eyes closed, avoiding body movements including jaw clenches for 60 seconds at a stretch.
• Writing Task -The participants were given a topic to write a simple short paragraph.They were provided with paper and a pen, the topic given was 'My family'.
• Typing Task -This task is similar to the writing task, where the participants were given a topic to type a simple short paragraph using a standard QWERTY keyboard.The topic given was 'How I spent my weekend'.

III. EEG Signal Processing
The EEG signals collected from each participant were processed in multiple phases prior to the classification as depicted in Fig. 3; namely preprocessing, sub-band decomposition and feature extraction.Each phase includes sub-phases as shown in the pseudocode in Fig. 4, which will be explained in detail in the following sections.

A. Preprocessing
The EEG signals were preprocessed in order to reduce unwanted artefacts such as eye blinks, body movements and electric power noise.Eye blinks and body movements were filtered using Artefact Subspace Reconstruction (ASR) which 'relies on a sliding-window Principal Component Analysis, which statistically interpolates any high-variance signal components exceeding a threshold relative to the covariance of the calibration dataset.Each affected time point of EEG is then linearly reconstructed from the retained signal subspace based on the correlation structure observed in the calibration data' [1,9].This was performed using the EEGLAB ASR plugin where the inputs were the relaxed state EEG that was the calibration dataset and the actual experiment task EEG.Shown below in Fig. 5 is a raw experiment EEG with the unwanted artefacts and in Fig. 6 the ASR filtered EEG.
Next, the electric power noise of 50Hz as shown in Fig. 7 was filtered out using a band-stop IIR Butterworth digital filter by removing at least half the power of the frequency between 49Hz to 51Hz as shown in Fig. 8.

B. Sub-band Decomposition
In this research, the EEG signals are analysed by decomposing the EEG signals into pre-defined sub-bands (Fig. 9).The sub-bands are namely delta, theta, alpha, beta and gamma.The sub-band decomposition was performed using band-pass FIR digital filters.Next, the frequency domain transformation was performed using MATLAB's FFT function.This function returns the Discrete Fourier Transform (DST) computed using a FFT algorithm.

C. Feature Extraction
A total of 8 features mean, median, mode, standard deviation, maximum, minimum, skewness and kurtosis were calculated for each participant, for each task, at each of the 5 frequency sub-bands (delta, theta, alpha, beta and gamma) in each of the 32 channels.All of these features collectively represent important characteristics of the EEG signal datasets.This adds up to a total of 1280 predictors per participant, which will be the input for the classifiers.

IV. EEG Classification
Previous studies [1,10,11] show that Support Vector Machines (SVM) is one of the most suitable classifiers to be used for EEG classifications.Hence, in this research we perform the classification of EEG using Cubic Support Vector Machines.Further, in addition to creating classifiers with all the EEG channels as a whole, classifiers were also created for different segments of the brain as illustrated in   The classifier outputs were measured based on the Validation Accuracy (VA), Sensitivity/True Positive Rate (TPR) and Specificity/ True Negative Rate (TNR) that were calculated using the resulting confusion matrix as shown in Fig. 10 and ( 1), ( 2) and (3).(1) (2) (3)

V. Results and Discussion
Poor writing skills are one of the commonly seen difficulties in individuals with dyslexia.The classifier results from the writing task, which is summarised in Table II, verify that adults with dyslexia produce unique brainwave signal patterns compared to normal controls.The peak VA of 71.88%, a sensitivity of 76.47% and specificity of 66.67% was produced from the anterior frontal classifier, which included the EEG electrodes AF7, AF3, AF4 and AF8.However, this outcome has not previously been reported in previous similar studies, and a possible explanation for this might be that because those studies had not used the EEG electrodes AF7, AF3, AF4 and AF8.The channels used in these similar studies were C3, C4, P3 and P4 [12][13][14].Therefore, these results contribute towards to the pool of knowledge as a new finding.Fig. 11 depicts the positions of AF7, AF3, AF4 and AF8.Typing can be considered as the modern-day replacement to writing and is yet another task found more challenging by individuals with dyslexia.Table III illustrates the behaviour of seventeen classifiers built to analyse the typing task.We examined the left hemisphere, right hemisphere, frontal lobe, central lobe, parietal lobe and the occipital lobe.Except for the parietal lobe, others showed a substantial difference between the sensitivity and specificity rates, which is not preferable.The classifiers from parietal and parieto-occipital performed fairly well.The frontal classifier showed the top VA of 78.13% with a fairly balanced specificity and sensitivity.Interestingly, this was close to the most significant region identified for writing, which was the anteriorfrontal.The most significant EEG channels responsible for producing unique brainwave signals in individuals with dyslexia compared to normal controls were F5, F3, Fz, F4 and F6.Fig. 12 depicts the position of these four channels.All these findings show that EEG signals generated while typing produce unique brainwave signal patterns in adults with dyslexia compared to normal controls.Further, comparison of EEG signal patterns between persons with and without dyslexia during typing is a gap to be filled in the literature; therefore, we did not find any research results that could be directly compared against our results.

VI. Conclusions and Future Work
In this paper, we conducted research to identify whether adults with dyslexia produced unique brainwave signal patterns during writing and typing.The results show that adults with dyslexia show unique brainwave activation patterns during each task compared to normal controls.Although similar writing tasks had been investigated in past studies, the current research was conducted with additional EEG sensors and discovered a new optimal brain region anterior frontal, which has not been reported in past studies.On the other hand, the research results also uncovered novel findings for typing as this task that had not been analysed in past similar studies.This research contributes vital insights to the pool of knowledge about the unique brainwave patterns of adults with dyslexia, which could serve as a base for future studies, and could even one day, help complement the conventional dyslexia diagnosis process by giving a better view of the disability through the introduction of neurological aspects.
These preliminary findings can be further examined by making variations in parameters such as input features, channels, frequency sub-bands, kernels and more advanced classifiers such as Fuzzy SVM.This could perhaps lead towards the enhancement of result accuracies similar to how the current research obtained better results by making variations in the EEG sensors used for each classifier.The scope of this research was limited to right-handed adults.Further studies can be carried out in order to compare EEG signals of individuals below 18 years and left-handed.Comparisons of the EEG signals could also be made between the genders male and female.Further, this research can be expanded in order to identify unique brainwave signal patterns of other specific learning disabilities such as dysgraphia and dyscalculia.Lastly, the function of each brain region needs to be compared with the result outcomes in order to identify the neurological reason behind each discovery.

Table I .
This helps to identify sections of the brain that have more prominent EEG activation patterns.

TABLE I .
Feature Grouping

Table II .
Writing Task Classifier Results

Table III .
Typing Task Classifier Results