Automated ASD detection using hybrid deep lightweight features extracted from EEG signals
Graphical abstract
Introduction
Autism Spectrum Disorder (ASD) is one of the most commonly studied developmental disabilities [1,2]. To meet the clinical diagnostic criteria for ASDs, individuals must have persistent deficits in social communication and social interaction as well as evidence of restricted and repetitive patterns of behavior [3], interests, or activities [4,5]. ASD commonly co-occurs with other neurodevelopmental symptoms including language disorders, intellectual disability, sleep disorders, and epilepsy [6]. It represents complex persistent neurodevelopmental differences that can typically be diagnosed before the age of 3 years [7]. Research on the causes of ASD is ongoing, but cumulative evidence points to ASD being due to a complex interaction of genetic and environmental factors. The heritability of ASD is estimated to be between 40 and 80%, and several hundred individual genes have been linked to autism [8]. A recent surveillance program (the Autism and Developmental Disabilities Monitoring (ADDM) Network) across 11 sites in the United States determined a prevalence of ASD of 18.5 per 1000 children aged 8, that is around one in 54 children. ASD is more commonly diagnosed in males than females. For example, ASD was diagnosed 4.3 times more often in boys than in girls, according to the ADDM study. ASD occurs in all ethnic groups [9]. Specialist doctors usually diagnose it, and several screening tools can be applied by a primary care physician [10]. However, a definitive diagnosis requires a comprehensive developmental evaluation. These evaluations are usually done by a trained developmental pediatrician, child psychologist, or trained allied health specialist [11,12]. In addition, magnetic resonance imaging (MRI) [13], computed tomography (CT) [14], and EEG signals [15] are being explored as objective measures for early ASD diagnosis.
There is currently no cure for ASD. However, it is possible to reduce the impact of this group of conditions. This requires the implementation of comprehensive early interventions to reduce symptoms, improve children's ability to learn and function, and participate in their communities [16]. There is convincing evidence that an early ASD diagnosis is important because younger children will acquire the necessary skills faster, and some of the symptoms seen in ASD will be controlled at an earlier stage through the early implementation of tailored education [17]. In this way, the differences between children with autism and their peers can be reduced over time [18,19].
Structural and functional MRI and (CT) are being explored as a source of objective information during ASD diagnosis, particularly for children with additional neurological symptoms such as atypical regression, severe micro or macrocephaly, epilepsy, or focal abnormalities on a neurological examination. Structural MRI and CT have demonstrated brain enlargement in children with ASD as young as 12 months, with a particular enlargement in the frontal and temporal lobes [20]. Functional MRI has shown alterations in connectivity between brain regions, especially between frontal-posterior regions [21]. However, neither of these imaging modalities is currently recommended for diagnosis or screening [11] due to a lack of precise correlation of neurological findings with clinical or pathophysiological features of ASD. Moreover, both types of imaging are associated with risks. CT is based on exposing the body to potentially harmful ionizing radiation. Therefore using this imaging technology requires careful risk assessment [22,23]. MRI is a less harmful alternative to CT; however, the cost of this imaging modality is also higher, and children often require sedation or a general anesthetic to allow accurate imaging, which carries inherent risks [24].
Children with ASD have an increased risk of epilepsy, ranging from 7 to 23%, with the highest risk in those children with concomitant intellectual disability [25]. Moreover, children with ASD without overt epilepsy commonly have abnormal epileptiform activity on EEG [26]. EEG signals result from passive measurements of the electrical brain activity that pose no risk to the patient. The signal acquisition process is cost-effective compared to MRI and CT [27,28]. The signals contain information about brain functionality [29]. Clinical studies show that they can be used to detect physiological changes that occur early in the disease course [30,31]. However, the level of evidence has not yet reached a stage where standard inclusion of EEG monitoring is recommended for screening or diagnosis of all ASD [11]. EEG signal acquisition is recommended when there is a clinical suspicion of seizures or atypical regression. Overnight EEG is recommended for children with late or atypical language regression due to a potential association with electrical status epilepticus of sleep [11,32].
Extracting information about these physiological changes through feature engineering can help experts to improve the analysis quality and reduce the analysis time [33,34]. To optimize the analysis process even further, current research has moved beyond the information-centric feature engineering to understanding and knowledge extraction through artificial intelligence (AI). AI systems have been used to simplify autism diagnosis. However, these studies are generally carried out with small datasets. In addition, not all EEG channels are considered in these studies. Therefore, some features that may be important for ASD might have been overlooked, and the accuracy rate for automatic diagnosis is low. Deep models have high classification capabilities. Therefore, deep models have been used to solve classification problems based on both images and signals. Recently, spectrogram extraction-based deep classification has reached high performance. In these models, Convolutional neural networks (CNNs) are used as deep learning algorithms.
Automated autism detection is one of the hard problems in machine learning applications. EEG is the most used input for autism detection, and specific models must be presented to translate the used EEG signals. Several datasets and models have been presented by researchers (see Table 1) to overcome this problem. This study aims to provide a general model for automatic ASD detection with high accuracy based on a large dataset. Therefore, an automatic classification model is presented with four phases. These phases are preprocessing, feature engineering, and feature assessment through machine learning (classification). The primary objectives of the proposed automatic autism detection model are to achieve high accuracy with fewer features and show the effectiveness of the proposed hybrid deep lightweight feature generator. Moreover, a big autistic EEG signal dataset has been used to denote the universal performance of the presented model. The proposed model reached 96.44% classification accuracy on a big dataset deploying a conventional classifier (SVM).
Novelties of the fused lightweight deep feature generator and ReliefF2 based model are:
- -
A novel automated ASD detection model is developed using a big database.
- -
A new feature selector is presented (ReliefF2).
- -
Using one dimensional local binary pattern (1D_LBP), and spectrogram extraction, a new signal to image transformation model which is named as 1D_LBP-STFT is presented.
- -
A fused lightweight deep feature generator is presented to use the advantages of three lightweight deep networks incorporated in the feature generator.
The following list highlights the major contributions of the proposed model for automatic autism detection:
- -
A hybrid deep lightweight feature generator is presented to extract salient features for classification. To deploy this model, an image conversion (spectrogram) generation model is used by deploying 1D_LBP. The 1D_LBP is an effective feature generator used for one-dimensional signals to extract the discriminative features. Then, these 1D_LBP features are fed as input to STFT to generate spectrogram images of the EEG signals. The presented transformation model is named as 1D_LBP-STFT spectrogram image generator. The resulting spectrogram images are fed to the deep lightweight feature generator. That generator uses pre-trained MobilNetV2, SqueezeNet, and ShuffleNet to accomplish the feature extraction.
- -
ReliefF2 selector is an improved version of the ReliefF feature selector which uses the ReliefF algorithm more effectively. The algorithm is used to choose the top 500 features from the deep lightweight feature generator.
To illustrate the advantages offered by the model developed in this study, we have structured the remainder of the paper as follows. In the next section, we review relevant literature and present research gaps for automated autism detection. Section 2 introduces the dataset used in the study and the methods employed to create the automated autism detection model. The findings obtained from the proposed model are shared in the experimental results section. Section 4 provides a comparison between the proposed method and other relevant scientific work. The conclusions are given in the last section.
In this literature review, we present the methods used for automated autism detection using EEG. Table 1 provides a summary of studies that investigated automatic diagnosis of ASD using EEG signals.
It can be noted from Table 1 that most of the studies have used smaller datasets. Hence, in this work, we propose an ensemble deep feature generator based on autistic EEG detection which yielded high-quality features with low computational complexity. When compared with the reviewed studies, it delivered the highest classification performance.
Section snippets
Material and methods
The proposed model consists of four steps: preprocessing, hybrid deep lightweight feature generation, feature selection, and classification. A snapshot (schematic demonstration) of the presented model is shown in Fig. 1. Details of these steps are given in the subsections below. This section provides an overview of the hybrid deep lightweight feature generator and ReliefF2 selector-based autism detection model. The proposed model was designed based on a big EEG signal dataset with 64 channels.
Experimental results
In this study, EEG signals were used for automated autism detection. For this purpose, a method based on hybrid lightweight deep feature generation has been developed. The proposed model is implemented using MATLAB (R2020b). The computer used in the study has two Intel Xeon E5-2697 processors. In addition, this computer has 256 GB RAM, 2.70 GHz, and Windows Server 2019 operating system. The features obtained by the lightweight deep feature generation method were classified using the MATLAB
Discussion
There is a pressing need to improve screening and early diagnosis of ASD to improve clinical and neurodevelopmental outcomes. EEG signals hold the promise of being a reliable information source for computer aided ASD diagnosis. However, standard methods have limitations, and EEG is currently not recommended as a standard diagnostic tool for ASD. In this study, a new method is proposed to assist in the automatic detection of ASD. For this purpose, a big EEG signal dataset was used. First, the
Conclusions
ASD is a common neurodevelopmental disorder that occurs in all regions of the world. It has debilitating effects on individuals, families, and communities. A recent systematic review estimated that lifetime costs for individuals with ASD are between $2.4 to $3.2 million. Although it is well established that an early diagnosis is highly desirable to optimize developmental and health outcomes, current screening tools have limitations. As a consequence, many children are diagnosed too late for
Declaration of competing interest
There is no conflict of interest in this work.
References (73)
- et al.
EEG-based multi-feature fusion assessment for autism
J. Clin. Neurosci.
(2018) Outcome of comprehensive psycho-educational interventions for young children with autism
Res. Dev. Disabil.
(2009)- et al.
Predictors of decoding for children with autism spectrum disorder in comparison to their peers
Research in Autism Spectrum Disorders
(2017) - et al.
Computed tomography and radiation: understanding the issues
J. Am. Coll. Radiol.
(2004) - et al.
Computed tomography before transfer to a level I pediatric trauma center risks duplication with associated increased radiation exposure
J. Pediatr. Surg.
(2008) - et al.
Imaging in rheumatoid arthritis–status and recent advances for magnetic resonance imaging, ultrasonography, computed tomography and conventional radiography
Best Pract. Res. Clin. Rheumatol.
(2008) - et al.
Clinical characteristics of children and young adults with co-occurring autism spectrum disorder and epilepsy
Epilepsy Behav.
(2015) - et al.
Interictal epileptiform activity and autism
Epilepsy Behav.
(2015) - et al.
Brain computer interface: control signals review
Neurocomputing
(2017) - et al.
Fuzzy synchronization likelihood-wavelet methodology for diagnosis of autism spectrum disorder
J. Neurosci. Methods
(2012)
Electroencephalography (EEG) signal processing for epilepsy and autism spectrum disorder diagnosis
Biocybernetics and Biomedical Engineering
Deep learning for healthcare applications based on physiological signals: a review
Comput. Methods Progr. Biomed.
Comparison of decision tree algorithms for EMG signal classification using DWT
Biomed. Signal Process Contr.
Feature selection using particle swarm optimization-based logistic regression model
Chemometr. Intell. Lab. Syst.
A generalized mean distance-based k-nearest neighbor classifier
Expert Syst. Appl.
Diagnosis of autism through EEG processed by advanced computational algorithms: a pilot study
Comput. Methods Progr. Biomed.
Automated identification for autism severity level: EEG analysis using empirical mode decomposition and second order difference plot
Behav. Brain Res.
Diagnosis of autism spectrum disorder based on complex network features
Comput. Methods Progr. Biomed.
A robust method for early diagnosis of autism spectrum disorder from EEG signals based on feature selection and DBSCAN method
Biocybernetics and Biomedical Engineering
Automated diagnosis of autism: in search of a mathematical marker
Rev. Neurosci.
Autism: cause factors, early diagnosis and therapies
Rev. Neurosci.
Diagnostic and Statistical Manual of Mental Disorders
Lancet
Autism spectrum disorder as early neurodevelopmental disorder: evidence from the brain imaging abnormalities in 2–3 years old toddlers
J. Autism Dev. Disord.
Genetic causes and modifiers of autism spectrum disorder
Front. Cell. Neurosci.
Prevalence of autism spectrum disorder among children aged 8 years—autism and developmental disabilities monitoring network, 11 sites, United States, 2016
MMWR Surveillance Summaries
Detection of early warning signs in autism spectrum disorders: a systematic review
Children
Identification, evaluation, and management of children with autism spectrum disorder
Pediatrics
Diagnostic and Statistical Manual of Mental Disorders (DSM-5®)
Functional magnetic resonance imaging of autism spectrum disorders
Dialogues Clin. Neurosci.
EEG analytics for early detection of autism spectrum disorder: a data-driven approach
Sci. Rep.
Why early diagnosis of autism in children is a good thing
The Conversation
Clinical impact of early diagnosis of autism on the prognosis and parent–child relationships
Psychol. Res. Behav. Manag.
Does quality of life differ for children with autism spectrum disorder and intellectual disability compared to peers without autism?
J. Autism Dev. Disord.
Early identification of autism spectrum disorder: recommendations for practice and research
Pediatrics
Cited by (76)
Grid-tuned ensemble models for 2D spectrogram-based autism classification
2024, Biomedical Signal Processing and ControlIdentification of autism spectrum disorder based on electroencephalography: A systematic review
2024, Computers in Biology and MedicineA comprehensive analysis towards exploring the promises of AI-related approaches in autism research
2024, Computers in Biology and MedicineDeep learning with image-based autism spectrum disorder analysis: A systematic review
2024, Engineering Applications of Artificial IntelligenceAutism spectrum disorder detection and classification using chaotic optimization based Bi-GRU network: An weighted average ensemble model
2023, Expert Systems with Applications