1 Introduction

Dyslexia is identified as one of the neurological disorders causing disability in reading and writing capabilities, without any intellectual deficit [20, 67, 102]. Dyslexia is also regarded as a language impairment disorder where patients find it difficult to read, communicate and write [94]. Such developmental neurological disorder is often characterized by reading and writing disorder, with psychological manifestations in children, adolescents, and adults [94]. Multiple varying symptoms can be observed in the person affected with dyslexia like slow reading and writing skills, confusion among words, the erratic spelling of words, struggle with scheduling and establishing tasks, an incorrect perception of alphabets, as shown in Fig. 1. The primary cause of dyslexia is misbalanced left hemisphere functions as a weakness in the zones concerned with language processing, speed, short-time memory, auditory, visual perceptions, speaking, and related motor skills [2, 44]. Several techniques have been identified to detect and diagnose dyslexia [14, 16, 33]. The conventional method of disease detection is behavioural symptoms study, which involves the utilization of highly recognized standard tests performed by the psychologist to evaluate the dyslexic state of an individual [23]. Other techniques used for disease identification are scanning of distinctive brain behaviours, known as brain imaging techniques like functional magnetic resonance imaging (fMRI) [93], magneto-encephalography (MEG) [84], and electroencephalography (EEG) [30, 82]. Nowadays, eye tracking is also exploited as an imaging technique for detecting dyslexia. These brain characteristics can be studied using EEG scanning as one of the main approaches.

Fig. 1
figure 1

Problems associated with dyslexia disorder

Full size image

One of the cost-effective and practical approaches for analyzing neurological defects is monitoring EEG signals as an exploratory cue. The common brain imaging techniques are costly and use ionizing radiations; therefore, they are unpractical and cannot be repeated to diagnose learning disability detection [1]. EEG is based on scanning stimulating brain responses using specific electrodes attached to the scalp. The process of EEG scanning for dyslexia detection involves optimization of input parameters, pre-processing of signals, feature selection and extraction and classification using machine learning algorithms. Input parameters like age, gender, and EEG channels specific to people with dyslexia need to be optimized for better study and analysis and inclusion/exclusion criteria and experimental setup. Once the EEG signals are recorded, the unwanted noises or artefacts need to be removed during pre-processing of signals as they can limit the accuracy of results. EEG creates a large amount of data that need to be minimized without loss of useful information for which specific feature extraction techniques like discrete wavelet transform (DWT) [24] and independent component analysis. DWT allows specific feature extraction and also categorize the EEG signals in five major frequency bands as θ (theta band - 4–8 Hz), δ (delta band - 0.5–4 Hz), α (alpha band - 8–13 Hz), β (beta band - 13–30 Hz) [34] and γ (gamma band - >30 Hz). Manifestly, neural processing is accompanied by alteration in the EEG frequency, which can be in the form of spatial coherence, change in amplitude or power, modification of envelope or consecutive combination of these. Studies have reported that arousal and motor actions are concomitant to low-frequency EEG signals. In contrast, attention, memory, expression, and emotions are typically correlated with high-frequency EEG signals [96]. Every EEG signal frequency band is concerned with a specific functional domain like deep sleep (delta band), drowsiness (theta band), relax state (alpha band), learning activities (beta band) and hyperactivity or stress (gamma band) [52].

After feature selection, the classification step is one of the most important steps in pattern identification of dyslexia. The classification step is generally performed using a specific machine learning algorithm as one of the main approaches. Learning and identifying new patterns from the enormous amount of data is the major advantage of using artificial intelligence approaches, but selecting a particular machine learning algorithm facilitates the accuracy and precision of the results. In multiple types of research, different types of algorithms have been studied like support vector machine (SVM) [63], Naïve Bayes [76], K – nearest neighbour (KNN) [73], ELM [101] and many others. The various algorithm for the classification of EEG signals is reviewed for different scenarios to identify the superior machine learning algorithms based on the performance evaluation.

The learning of children with dyslexia disorders requires special efforts as the patient unable to describe the scenarios and impairment took long time for the problem identification. Several machine learning based modules are available for learning and understanding of symbols and hand eye coordination [57, 61]. An interactive and machine learning based platform which could training itself with the input data received from patients affected with dyslexia could help in enhance their learning at their own pace. The key to the prevention of learning disorders like dyslexia is the early detection of disorder that could be performed through intervening with children’s in their early phase of reading, writing and recognising words as well as in speech recognition reduce the risk of spreading of disorder from 20% to below 5% of children [80].

This manuscript is focused on analyzing EEG specific patterns for dyslexia and process involved in the detection and diagnosis of dyslexia using machine learning techniques for the early detection of disorder to minimize the risk involved. The treatment of disorder could be achieved if detected at early stage with proper learning activities, approaches and techniques. Machine learning approach for the detection of disorder involves four steps; optimization of input parameters, pre-processing of EEG signals, feature selection and extraction and classification using machine learning algorithms with performance evaluation. The research papers are divided into three sections: the study of EEG signals and pattern identification, feature selection and extraction studies, and machine learning techniques for data classification. The total number of papers reviewed after screening is shown in Fig. 2a, and b indicates the distribution of machine learning techniques applied by the researcher in papers used for the review. A separate section is designed for each study area to cover the pros and cons of each area. Section 2 of this paper describes the EEG acquisition process and an overview of its applications. Section 3 highlights the scope of using EEG as a benchmark for Dyslexia diagnosis, and section 4 reviews the studies carried out by various researchers using machine learning techniques and performance evaluation of various techniques for the diagnosis of dyslexia, and section 5 contains the conclusion obtained from the review of all the studies performed by the researchers on techniques used for the diagnostic of dyslexia.

Fig. 2
figure 2

a Distribution of papers based on different areas of study and b distribution of papers based on the machine learning technique used for data analysis

Full size image

2 Electroencephalography (EEG) signals

Multiple electro-biological measurements like ECG (Electrocardiography), EEG, EMG (Electromyography), and many others are commonly employed as body imaging techniques for the detailed study of specific organs and related disorders. One medical imaging technique, which reads electrical activity stimulated by brain structures using metal electrodes and conductive media, and presents as an electroencephalogram, defines electroencephalography. EEG reading is a non-invasive procedure that incorporates the study of event-related potentials and identifies specific brain signals. Activating neurons as a response to the stimulus generates local current flows (ion exchanges) across the neuronal membrane, creating action potentials. EEG working is based on measuring local currents produced during synaptic transmissions across neurons as responses to the stimulus. EEG scanning is considered one of the influential techniques in neurosciences because of its proficiency in reflecting both normal and abnormal activities [30]. This section describes the working of EEG in capturing specific brain waves specific to the concerned active zone of the brain.

2.1 Human brain and brain waves classification

Neurons constitute the structural and functional unit of the brain, responsible for the transmission of impulse (synapses) and generation of action potential concerning specific stimulus-induced to the human body. The human brain is categorized in three main structural divisions: cerebrum, cerebellum, and brain stem. The cerebrum (left and right hemisphere) comprises learning, reading, emotions and behaviour, thinking and voluntary movements. Each cerebral hemisphere is divided into four lobes as frontal lobe (problem-solving and judgment skills), the parietal lobe (handwriting, reading, arithmetic and taste), the occipital lobe (visual processing system) and the temporal lobe (memory and hearing). The cerebellum is at the base and back of the brain, playing a crucial role in coordination and balance. EEG portray the electrical activity of the cerebral cortex related to specific lobes and hemisphere. EEG states the brain waves or signals concerned with the specific zone [28]. Generally, brain wave patterns are sinusoidal and can be measured over a small amplitude range from 0.5 to 100 μV. EEG records the brain waves and uses different extraction techniques; four significant brain sub-bands waves are extracted, as alpha (8–13 Hz), beta (>13 Hz), theta (4–8 Hz) and delta (0.5–4 Hz). Every frequency sub-band is concerned with a specific domain like deep sleep (delta band), drowsiness (theta band), relax state (alpha band) and learning activities (beta band) [52].

Sensitivity of EEG to variant states like stressed, alert, resting, hypnosis and sleep, define its specificity in recording explicit brain wave patterns. Hence, EEG signal data comprises wave patterns of different characteristics and even unique brain wave nature of every individual allow characterization of precise wave patterns, reflecting brain activities following electrode placement and related brain zone. This application of EEG makes it suited for detecting specific brain functioning abnormalities as learning disorders, as shown in Fig. 3 [68]. Other advantages of EEG over other brain imaging techniques are its high speed as complex pattern of brain can be recorded in fraction of seconds once stimulated, less spatial resolution, and low cost. These advantages mentioned above allow the selection of EEG scanning for detecting and diagnosing learning disabilities like dyslexia [58].

Fig. 3
figure 3

EEG Signals [68]

Full size image

2.2 EEG signal recording system

EEG Recording system involves the employment of electrodes with conductive media, amplifiers with filters, A/D converter and recording device for encephalographic measurements. Reciting neuronal signalling is made possible by electrodes followed by amplifying signals through amplifiers. Analogue signals are converted to digital form through the converter, after which data is then stored and displayed (M. [51, 87]). During the mono-channel EEG measurement process, three electrodes as the active electrode, reference electrode and ground electrode are utilized, where signal conduction between signal and reference electrode is recorded as fluctuations in neuronal potential gradients and ground electrode works for deriving differential voltage. On the contrary, multiple signal (active) electrodes as 128 or 256 are engaged in multi-channel configurations.

Amplification of EEG signals indicates its compatibility with devices like recorders, A/D converters and others. Amplifiers and filters ensure the removal of unwanted noises and prevent the distortion of desired signals. Once the signals get amplified, they are converted to digital, stored, and displayed [85]. According to various researches carried out using EEG sampling, a specific EEG recording system should comprise of electrode cap with conducting jelly, an amplifier with amplification gain between 100 and 100,000, input impedances as 100 M Ohms, and a common-mode rejection ratio of at least 100 dB, and analogue filters with high pass and low pass filters. All such features dominate the study of EEG signals over other electro-biological measurements for the diagnostics of dyslexia.

3 EEG and learning disabilities (dyslexia)

Alterations in brain functionalities, influencing cognitive processes associated with learning abilities, define learning disabilities (LD). Specific genetic, neurobiological, or cellular factors contribute to developing these disabilities, causing impairments in the child’s intellectual skills like reading, writing, or speaking. Such disorders lead to the malfunctioning of brain processing activities, which can be predicted from the difference between expectations of one’s intelligence and related performance. Various learning disabilities involve reading disabilities, written language disabilities, math disabilities, attention defects and autism disabilities [12, 48]. They can be categorized as dyslexia (inability to read and comprehend the text accurately and fluently), Dysgraphia (difficulty in writing), Dyscalculia (maths and calculation difficulties), ADHD (Attention-Deficit/Hyperactivity Disorder), and ASD (autism spectrum disorder). The study will focus on dyslexia, epidemiology, detection methods, and related studies.

3.1 Dyslexia

Rudolf Berlin coined the term ‘Dyslexia’ in 1887 by identifying the disability in learning in Oswald Berkhan (the first person to be identified with the disorder). It is a special learning difficulty (SLD) in which a person faces difficulty in reading, word-spelling, writing, reasoning, word decoding with accuracy and fluency and many other neurological traits irrespective of average or above-average intelligence [29]. The reading accuracy is one of the standard criteria for diagnosing dyslexia where, if the accuracy is more than 1.5 SD (standard deviation) below the mean; then it defines the 7% proportion of the population to be dyslexic [64]. The disorder marked itself as complications in phonological [13], orthographic [13], working memory [29], brain systems asynchrony [13], poor executive function skills [70] and rapid naming processing [29]. If dyslexia gets detected in early childhood, then stated as the developmental while acquired due to brain injury or stroke, then referred to as acquired dyslexia.

Behavioural symptoms are generally considered as the study approach for the conventional dyslexia detection techniques, involving valuation through standardized tests like Wechsler Individual Achievement Test (WIAT), Comprehensive Test of Phonological Processing (CTOPP), Oral and Written Language Scales (OWLS), Woodcock-Johnson (WJ), etc. These tests are based on assessments of the person’s reading, writing, IQ, memory and phonological processing abilities and conclude the state of dyslexia. These conventional methods based on behavioural aspects are highly time-consuming and tiresome, and even the variability in symptoms among individuals make the analysis a challenging task. Several techniques have been proposed by researchers for detecting developmental dyslexia like reading/writing text [49], web-based word games [72], eye tracking [8], MRI scans [65], EEG scans [73], video and image capturing [32], etc. Usage of different detecting methods depends on varying attributes to be diagnosed for dyslexia like grey matter deficit using structural magnetic resonance imaging and reduced neural activities in specific brain zone can be demonstrated using MRI scans and EEG scans [84]. EEG scans are successfully exploited for dyslexia detection by identifying unique brain activation patterns specific to brain activities. The section focuses on analyzing unique brain structures, dynamics, and behaviour through EEG, which come out as efficient techniques and potentially reveal the indicators specified in the dyslexic’s brain.

3.2 Analysis of EEG signals

EEG scanning and analysis can be potentially used for detecting dyslexia and the sequential steps for the recorded data processing involve optimization of input parameters for EEG signal recordings, pre-processing of signals followed by feature selection and extraction, and classification using machine learning algorithms. The manuscript provides the insights of multiple studies dealing with EEG scan analysis for dyslexia and analyzes an improved framework to extemporize the classifier’s performance and accuracy in discriminating between dyslexics and controls. Steps for input data processing using EEG signal recording for dyslexia are:

3.2.1 Functionality domain and channel selection

Multiple kinds of research identified unique brainwave patterns for dyslexia using EEG, focusing on optimizing various parameters like the number of participants, age group, EEG channels, Inclusion and exclusion criteria of subjects, domain selection among multidisciplinary challenges in dyslexia metrics and experimental setup protocols.

  1. A.

    Number of participants, Age group and Gender: EEG is often considered a signature of cognitive activities. In order to understand the underlying cognitive of learning disabilities, it is essential to evaluate the electroencephalogram parameters and their relation to the results of the Wechsler Intelligence Scale. The sample size includes test and control groups, variable numbers, as shown in Table 1. Accuracy in results can be obtained after evaluating the mean sample size of multiple studies (Table 1). The average number of participants considered for the classification of dyslexia and non-dyslexia persons for the test and control groups is 18 and 12, respectively. Both age groups as child and adult groups can be possibly evaluated for dyslexia as concluding from the studies, defining minimum age of 2 years to maximum age of 40 years can be selected, irrespective of the gender.

    Table 1 Studies demonstrating the input parameters for EEG scan analysis
    Full size table
  1. B.

    Domain selection: Dyslexia is a syndrome involving a disability across a range of activities and severity, which generate several challenges in developing dyslexic metrics to be measured for all possible discriminant. These observations anticipate the extent of impairment towards hearing, language; vision, reading, and spelling; writing and speed; arithmetic calculations and time management; memory and cognition; and behaviour, health, development and personality. Previously, communication of brain systems and related functional networks can be evaluated using task-independent EEG activity [92]. However, the domain specificity allows pattern recognition and increase the classifier efficiency. Some studies focused on the reading domain for classifying EEG signals specific to dyslexia, like evaluating Real word vs non-sense word reading-related differences in Dyslexics and controls [62]. The studies performed on writing controls as the specific domain for the dyslexic characterization incorporated writing as the explicit domain as it requires special attention from learners defining the signal conduction and classification from the frontal lobe to the motor cortex [19, 26, 52, 53]. Studies may include the selection of specific domains like as followed in Selvi and Saravanan [76] study, where 16 most frequent characteristics of dyslexia were taken into consideration, including difficulty in reading, writing, spelling, handwriting and speed management memory and concentration, and others.

  1. C.

    EEG channels: During EEG measurements, electrodes are placed near specific brain centres like Occipital (O), Frontal (F), Temporal (T), Parietal (P) and Central (C). In different studies, as listed in Table 1, selecting a variable number of channels depends on the target challenge related to dyslexia like reading, writing, speech or mental disabilities. Most of the studies selected the writing domain as the study zone for categorizing unique brainwave patterns in people with dyslexia [19, 24, 98]. These studies investigated the central and parietal lobe channel-specific for writing analysis as C3, C4, P3 and P4. Anterior-Frontal lobe channels were studied as the unique brain signal pattern generating lobes as AF7, AF3, AF4 and AF8; concluding with the classifier results validation through sensitivity (76.47%) and specificity (66.7%) analysis [63]. Typing challenge study is now rising as one of the novel approaches and modern-day replacement to writing, distinguishing the brainwave pattern of dyslexics and controls. Frontal lobe channels (F5, F3, Fz, F4 and F6) came out as the most significant EEG channels for producing unique brainwave patterns specific to typing difficulties in dyslexic individuals, as validated based on sensitivity and specificity analysis [63]. Based on studies, popular EEG channels reported are Fp1, F3, Fz, F4, F6, F7, F8, T3, C3, Cz, C4, T4, PHz, AF3, AF4, TP7, P7 for identifying unique brain waves pattern for dyslexia. Hence, depending on specific domain selection, specific EEG channels need to be investigated to generate brainwave patterns.

  1. D.

    Inclusion and Exclusion Criteria: During the research, evaluation criteria involving inclusion and exclusion measures allow appropriate selection of participants without affecting the quality and efficacy of the result. History of mental illness, any genetic or neurobiological disorder, brain injuries, hearing problems, drug or alcohol addictions, and attention deficit disorder diagnosis are all listed as exclusion measures [77]. Inclusion criteria involve selecting participants with any of the dyslexic traits as dyslexics and none of the dyslexic traits as control or normal [9].

  1. E.

    Experiment setup: Following multiple types of research, dyslexia specific brain wave activation patterns are highly projectable during reading and writing practices. Reading and recognizing difficulties in regular and non-sense words is the successful approach for categorizing dyslexia in individuals [74]. Writing domain is also explored as one of the successful tasks differentiating between dyslexics and normal [79]. Hence, most commonly, a combination of them is considered for the experimental setup. One primary concern is the early detection of dyslexia. Late diagnosis of learning disorders causes the development of incompetency, poor self-assurance, low confidence, physiological and emotional disturbances, hassles, anxiety and depression, in the person. Hence, based on the problems concerned with late diagnosis, it becomes mandatory to search the methods dealing with early diagnosis of dyslexia. Early-stage detection of dyslexia at younger ages of a child can provide timely treatments along with implementations of the right ways of learning to cope with their abilities [25, 80]. Therefore, learning disabilities need to be diagnosed in the child before commencing school life to avoid unease, bully, and other social problems faced by a child in school, which affects the child’s future and leads to neurological and personality imbalances in the future child. Most commonly, as depicted in the studies mentioned above, reading and writing traits; or words and spell recognitions; are selected to differentiate between the dyslexic and non-dyslexic child [45, 56]. However, for determining the learning disability impedance in preschool’s children, above mentioned traits cannot be successfully used as experimental traits. Mammarella et al. [54] tried to implement the learning disability detection based on emotions captured from EEG to analyze the symptoms of disorders like autism spectrum disorder (ASD), attention deficit hyperactivity disorder (ADHD) and dyslexia, which can assist in early detection and follow-up clinical evaluations. The research also presented the neuro-physiological Interface of affect (NPIA) application, using data from EEG, where NPIA analyze an individual’s emotion based on Valence (V) and Arousal (A) [69].

However, research can be pursued on all the optimized parameters with further steps such as pre-processing EEG signals to remove unwanted noises, followed by feature extraction and classification using machine learning algorithms.

3.2.2 Pre-processing of signal

All EEG data need to be pre-processed to confiscate artefacts and unwanted noises, where artefacts most commonly ascend due to automatic muscular movements, eye blinking, jaw movements, cheek related movements and other bodily movements. The unsolicited spikes and alterations in the interpretation of the recorded signals created due to these artefacts need to be eliminated (Xiang et al., 2017). Perera et al. [63] used the artefact subspace reconstruction (ASR) technique to filter raw EEG data’s eye blinks and body movements. The technique utilizes sliding-window Principal Component Analysis (PCA) to intercalate high variance signal components exceeding a threshold relative to the covariance of the calibration dataset. Multiple ways have been used in studies like elliptic filters [69] were used to smoothen the raw EEG signals and descend the bandpass filtering for five different bands (delta (1-3 Hz), theta (4-8 Hz), alpha (9-13 Hz), beta (14-25 Hz) and also gamma (26 Hz–40 Hz)).

3.2.3 Feature selection and feature extraction techniques

Data sampling methods like EEG, ECG or fMRI recordings generate a vast amount of data in the medical domain, necessitating the exploitation of dimensionality reduction techniques. Dimensionality reduction techniques are feature analysis techniques, including feature selection and feature extraction, to improve the experiment’s performance and result in inaccuracy [46, 50]. The reduction in attributes and data compression causes some data loss; however, the data lost is of low significance and causes the least variation in the overall dataset. The lost data dominantly contains redundant features and produces a negligible impact on data quality while significantly reducing the computational time. Along with this, the high computation cost of huge dimensional data often raises difficulties in applying the classification algorithms. In the feature selection process, a subset from available features data is selected for the process of the learning algorithms. The best subset has the least number of dimensions, contributing to the learning accuracy. The feature extraction process involves the reduction of the feature space size without losing the original feature space [46, 89]. Hence, the main difference between feature selection and extraction is that the first reduces by selecting a subset of features without transforming them. In contrast, feature extraction reduces dimensionality by computing a transformation of the original features to produce other substantial features. The choice between feature extraction and feature selection methods depends on data types and application area.

  1. A.

    Feature Selection: The two significant categories of feature selection algorithms include filter-based methods and wrapper-based methods. Evaluation and selection of feature subset can be made using filter-based methods, which rely on general statistical characteristics of data. The best suitable feature search, which will fit well for the study, can be made possible using wrapper methods [47]. Frid and Manevitz [22] made feature selection using the ReliefF algorithm. The ReliefF algorithm is based on both filter and wrapper approaches, where no assumption about data distribution can be made, hence also defined as a non-parametric feature algorithm. Rezvani et al. [73] utilized EEG resting-state data of 29 dyslexics and 15 typical readers in grade 3 and calculated weighted connectivity matrices using the phase lag index (PLI) for deriving weighted connectivity graphs. Graphs allowed computation of several local network measures, after which False Discovery Rate (FDR) corrected features were designated as input to the specific classifier.

  1. B.

    Feature Extraction: Suitable feature extraction techniques are required to reduce large EEG signal data to analyze the specific area of investigation. The recorded raw EEG signals are generally in the time domain but need to be transformed first in the frequency domain using Fast Fourier Transform (FFT) [3]. The technique cause decomposition of waveforms as the combination of the sinusoidal waves into the sum of sinusoids of different frequencies. However, the applicability of FFT for non-stationary signals majorly restricts its application towards feature extraction [82]. Fuad et al. [24] applied Short Time Fourier Transform (STFT) and Discrete Wavelet Transform (DWT) for the extraction to overcome the drawback of the FFT. Nevertheless, researchers used both techniques, but DWT was compatible enough to extract useful information from non-stationary signals [104]. Discrete Wavelet Transform is one of the capable methods of analyzing discontinuities in signals caused through varying window sizes. It can extract energy and analyze EEG signals in time and frequency domains. Another advantage is that it provides accurate frequency information at low frequencies and accurate time information at high frequencies [40]. Upadhyay [90] reported about the classification of EEG signals using wavelet transform, where the study illustrated the wavelet decomposition process of EEG signals in Low pass data (h[n]) and High pass data (g[n]) and finally decomposed the EEG signals in 4 levels (D1-D4). Multiple studies explore the application of DWT as one of the feature extraction techniques [4, 31, 52, 59, 100]. Zainuddin et al. [100] suggested that decomposition of EEG signal bands has to be performed into sub-bands of five different frequencies as Delta band, Theta band, Alpha band, Beta band and Gamma band using DWT analysis. Delta (δ) band (0.5 to 4 Hz) is allied with deep sleep; theta (θ) band (4-8 Hz) is associated with drowsiness or dreaming; alpha (α) band (8-13 Hz) specifies relaxation or awareness; beta (β) band (13-30 Hz) displays concentration or active attention; gamma (γ) band (more than 31 Hz) is acquired by concurrent information processing from different brain parts. The dimensionality reduction techniques commonly remove the high pass data (considered noise during EEG analysis). However, Frid and Manevitz [22] utilized that fraction for feature extraction and related classification, as shown in Fig. 4. DWT was carried out on EEG signal in their study for feature extraction, and transformed signals were used to divide the event-related potential (ERP) signal into two complementary parts as Low Pass (LP) data (defined as ‘Approximations’ of the signal) and High Pass (HP) data (defined as ‘Details’ of the signal). The variant sets of temporal features like latency, absolute amplitude, maximal peak, positive area, overall signal energy, and entropy were extracted from the LP part. Statistical and spectral features like mean, standard deviation, skewness ratio was extracted from the HP fraction. Selection of specific features following region of interest were made using feature selection algorithm, followed with classification by SVM [36].

Fig. 4
figure 4

Overview of Feature extraction, Feature selection and classification process [22]

Full size image

Apart from DWT, other multiple feature extraction methods were compared and studied numerously [55]. Soetraprawata and Turnip [81] suggested that independent component analysis (ICA) allows the decomposition of a signal into temporal independent and fixed components and is a reliable technique for feature extraction. Other methods like principal component analysis [88], empirical mode decomposition (Park et al., 2011), autoregressive modelling [98] and many others have been exploited as feature extraction techniques with specific applications. Multiple other variant techniques were explored in studies to extract writing task linked distinguishing features at specific signal locations in EEG data like Frequency content [98] and Power spectrum [52]. Feature extraction using Mel Frequency Cepstral coefficients (MFCCs) techniques have also been exploited in multiple studies for brain signal analysis ([42]; Kamaruddin et al., 2019; [69]).

4 Classification using ML algorithms

Nowadays, the rapidly emerging fields of Machine Learning (ML) and Artificial Intelligence (AI) are unsettling many traditional algorithms and ultimately guarantee to reorganize many aspects of daily life. Machine learning algorithms assist in identifying new patterns from enormous data as EEG generates a large amount of data. Hence, signal data classification can be performed using variant machine learning algorithms. In line with the studies performed on dyslexia, the studies performed using machine learning techniques in the last 15 years are evaluated. The classification involves utilizing features extracted from previous steps like the minimum, maximum, mean, standard deviation, power, energy, average valley amplitude, peak variation, and root mean square, as followed in recent EEG-related studies. Such reform would be especially beneficial for algorithm improvements in EEG signal data classification [39]. Energy, average valley amplitude, peak variation, root mean square and power are some of the features used in recent EEG-related studies [63]. These features were combined to investigate the feasibility of such features to classify the subjects as either controlled or uncontrolled side. The Confusion Matrix (CM) provides the total number of instances rightly and wrongly classified by the classifier with the features like sensitivity, specificity and accuracy [6, 63].

Depending on the requirement, the classification task can be performed using either Support Vector Machine (SVM), Optimum-Path Forest (OPF), Naïve Bayes, k-Nearest Neighbor (KNN) classifiers [75]. The popular machine learning algorithms used in EEG related studies are linear discriminant analysis, SVM and neural networks [75, 78]. Linear discriminant analysis is considered one of the simple classifiers, classifying data by first designing models of probability density functions for the data and then assigning new data points of larger values than others. However, the algorithm cannot be effectively utilized for complex nonlinear EEG classification and does not yield efficient results [51]. In contrast, neural networks and SVM perform well in EEG classification, where neural networks can be used to implement boundaries for nonlinear classification, and SVM can effectively handle both linear and nonlinear classifications [27, 37, 51]. Frid and Breznitz [21] proposed an algorithm for differentiating dyslexic readers from non-dyslexic readers using their EEG recorded data. Support Vector Machine (SVM), the most common ML technique, used feature analysis (maximal peak amplitude, positive area and spectral flatness measure) and their extraction. The ensemble of SVM was used for classification, and majority electives were used for finalizing the results. The major drawback of the algorithm was its restricted sample size, including only young dyslexics, along with the lack of accurate measurements of the algorithm [21]. Zainuddin et al. [99] explored the improved KNN classify rule for identifying EEG based discrepancies between capable dyslexics and normal children. Mahmoodin et al. [53] performed EEG classification using an SVM classifier for determining EEG electrode localizations specific to the dyslexic domains. Zainuddin et al. [100] assessed the performance of the Extreme Learning Machine (ELM) classifier with radial basis function (RBF) kernel in classifying between normal, poor and capable dyslexics, based on EEG signals of their writing data. The study concluded the highest accuracy of 89% and best Receiver operating characteristics (ROC) performance with high specificity and sensitivity. Zainuddin et al. [101] reported the performance comparison between KNN with correlation distance function and ELM classifier with radial basis function in classifying EEG signal (writing domain) of normal, poor and capable dyslexic children. The study concluded the ELM outperformance over KNN with 89% accuracy.

Perera et al. [62] predicted dyslexia using EEG data and applied the SVM classifier model for feature extraction and classification of dyslexics. They also identified the pros and cons of the EEG approach and suggested optimization techniques for better evaluation. Their subsequent study [63] developed a unique EEG pattern specific to dyslexia following classification using Cubic Support Vector Machine, which reveals the great difficulties in writing and typing by dyslexics compared to the normal controls. They also revealed a unique brain wave pattern related to anterio-frontal lobe channels linked with writing tasks and frontal channels linked with typing difficulties in dyslexics. The measurement of SVM classifier output was made using Validation Accuracy (VA), Sensitivity/True Positive Rate (TPR) and Specificity/ True Negative Rate (TNR). Restriction of study to right-handed adults limited the scope of research and opened the gate for further research, including defining more unique brainwave patterns for other domains related to dyslexia. Several machine learning methods like Shannon Entropy Vector and Artificial Neural Network (ANN) classifier have been used for classifying EEG sub-bands during diagnosis of learning disabilities like ASD [18].

MLP (Multi-layer perceptron) is also proposed as one of the classification methods to identify EEG emotions, addictions, behaviour and others in multiple studies [43, 95, 97]. Razi et al. [69] classified EEG signals based on emotions using MLP and NPIA for positive or negative emotion categorization. Selvi and Saravanan [76], comparative performance analysis of different machine learning algorithms (SVM, Naïve Bayes, Decision tree and Neural networks) were made to detect and diagnose dyslexia in children based on intellectual and emotional intelligence. The study concluded the relative competitive performance of SVM and Neural networks concerning their efficiency. However, SVM is the most popular supervised learning technique and produces exceptional results and is capable of creating adequate space division for the accurate placement of new data points, as shown in Fig. 5 [15, 66].

Fig. 5
figure 5

Comparative of various machine learning techniques [15]

Full size image

Rezvani et al. [73] explored the application of machine learning classification on EEG local network features for discriminating between dyslexics and controls. They utilized EEG resting-state data and extracted the local network features, and related FDR corrected groups, which were used to input the classifier (SVM and KNN). Evaluation of classifier performance was made using cross-validation and random shuffling technique to assure the performance pertinence of the classifier. SVM classifier with linear kernel performed the best with 95% accuracy, 96% sensitivity, 93% specificity, and 96% precision compared to KNN. It was observed through the comparison of the popular machine learning-based classification algorithms for EEG signals that SVM is a better choice and outperforms other machine learning tools [91]. SVM, being a supervised learning method, can handle both linear and nonlinear classification and can classify even overlapping and non-separable data sets by mapping onto higher-dimensional spaces using the kernel functions as shown in Fig. 6 [15, 73]. This feature of SVM postulates the reason for its outperformance over others. Along with dyslexia detection, similar successful results have been obtained in classifying other mental tasks, seizure detection, epilepsy diagnosis, vigilance analysis and others.

Fig. 6
figure 6

SVM classification and performance evaluation [73]

Full size image

Along with the research mentioned above, one insight also directs our interest in applications of deep learning algorithms and related hybrids in related EEG classification tasks. As one of the new emerging fields, deep learning algorithms are also exploited well for EEG classification tasks, where convolution neural networks (CNN), recurrent neural networks (RNN), deep belief networks (DBN) yield better classification accuracy as compared to stacked auto-encoders and multi-layer perceptron neural networks [17, 103]. Deep learning algorithms are generally found to be accessed for emotion recognition [41], motor imagery [86], mental workload [35], seizure detection [38], event-related potential detection [10], and sleep scoring [11]. In future research, deep learning techniques and related hybrid models can be researched to detect and diagnose dyslexia.

Apart from EEG classification tasks, other domains can be traced for dyslexia detection like eye scanning, image or video processing, MRI scan data and many others. Considering that researches, below mentioned Table 2, states the different machine learning algorithms employed for the classification data of dyslexics and controls.

Table 2 Dyslexia detection using different machine learning algorithms
Full size table

5 Conclusion

This manuscript focuses on analyzing EEG specific patterns for dyslexia detection using machine learning techniques for the early detection of disorder to minimize the risk involved. The treatment of disorder could be achieved if detected at early stage with proper learning activities, approaches and techniques. EEG potentially provides practical scanning of neurological defects and analysis for dyslexia detection with machine learning techniques. The complexity of neuroscience and dyslexia detection causes high morbidities and growing instances of neurological disorders necessitate the study on early detection of neurological disorders, EEG scan analysis for dyslexia, and analysis to extemporize the classifier’s performance and accuracy in identifying the disabilities/disorders. The extensive data size suggests the need for machine learning techniques for data handling, analysis, and classification of EEG signals. Identifying unique brainwave patterns for dyslexia using EEG requires optimizing input parameters based on channel, domain, and inclusion-exclusion criteria. The manuscript on machines learning techniques for the classification of EEG signals to detect dyslexia have been reviewed to identify the challenges and opportunities in the process, as the early detection of dyslexia could reduce the mortality from 20% to 5%.

The various approaches of machine learning techniques illustrated the challenges in optimizing input data. The results obtained by various researchers suggest that the application of machine learning for detecting dyslexia compliment the findings of existing model. However, with the involvement of machine learning approaches, the efforts has been reduced significantly. The selection of appropriate pre-processing signals to remove artefacts and unwanted noises in EEG signals, followed by feature extraction techniques to classify the data adequately could improve the results considerably. The review finding suggest that PCA is an effective pre-processing technique and provides valuable information about the data, whereas, wrapper-based methods for feature selection and Discrete Wavelet Transform have been identified as the most appropriate feature extraction technique. The review suggest that SVM found as highly successful technique for the classification of data using multiple machine learning algorithms to analyze and classify EEG signals into multiple classes of various disorders. The several combinations and models of neural networks may also reduce the challenges up to greater extent if accorded adequately and could be used as future research directions. The combination of models working on emotion recognition which is not defined by the region or boundary and models working on pattern recognition, could improve the model accuracy and reduce the model constraints. The development of such ensemble models opens multiple directions for future research in the early detection of dyslexia.