Brain-region specific autism prediction from electroencephalogram signals using graph convolution neural network

Abstract

BACKGROUND:

Brain variations are responsible for developmental impairments, including autism spectrum disorder (ASD). EEG signals efficiently detect neurological conditions by revealing crucial information about brain function abnormalities.

OBJECTIVE:

This study aims to utilize EEG data collected from both autistic and typically developing children to investigate the potential of a Graph Convolutional Neural Network (GCNN) in predicting ASD based on neurological abnormalities revealed through EEG signals.

METHODS:

In this study, EEG data were gathered from eight autistic children and eight typically developing children diagnosed using the Childhood Autism Rating Scale at the Central Institute of Psychiatry, Ranchi. EEG recording was done using a HydroCel GSN with 257 channels, and 71 channels with 10-10 international equivalents were utilized. Electrodes were divided into 12 brain regions. A GCNN was introduced for ASD prediction, preceded by autoregressive and spectral feature extraction.

RESULTS:

The anterior-frontal brain region, crucial for cognitive functions like emotion, memory, and social interaction, proved most predictive of ASD, achieving 87.07% accuracy. This underscores the suitability of the GCNN method for EEG-based ASD detection.

CONCLUSION:

The detailed dataset collected enhances understanding of the neurological basis of ASD, benefiting healthcare practitioners involved in ASD diagnosis.

Keywords

Autism graph convolution neural network electroencephalogram deep learning brain region

1. Introduction

Autism Spectrum Disorder (ASD) stems from brain variations and can manifest from early childhood to old age, often resembling genetic disorders [1]. ASD is believed to have multiple causes impacting individuals’ development [2]. Early diagnosis is critical, as ASD affects approximately 1 in 160 children globally and is on the rise [3]. Diagnostic assessments, like screening tests, aim to identify core ASD symptoms such as social disengagement, communication challenges, and restricted behaviours. The diversity in ASD presentations can lead to both over- and under-diagnosis by specialists unfamiliar with the full spectrum [4].

EEG signals offer valuable insights into brain function [5]. Researchers are using neuroscience to diagnose and treat ASD in children, where studying and classifying EEG patterns can aid in diagnosis [6]. Diagnosis is typically carried out by trained medical professionals using standardized tools and involves interactions with parents or guardians to assess the child’s history and behaviour. However, this process can be time-consuming. To speed up diagnosis and enhance accuracy, a smart system is needed. While machine learning can identify key ASD traits, deep learning methods excel in handling complex, large datasets [7, 8].

Deep learning-based methods are increasingly popular for automatic feature extraction and classification, particularly in the context of diagnosing ASD using EEG signals [9, 10, 11]. For instance, Ari et al. [12] achieved a 98.88% accuracy using a combination of EEG signal reduction, sparse encoding, and deep convolutional neural networks. Oh et al. [13] extracted 18 nonlinear EEG features and reached a 98.70% accuracy using feature selection and a support vector machine (SVM). Lavanga et al. [14] explored EEG data from babies under four months old, achieving a 79% AUC for multiscale entropy in predicting ASD or neurocognitive outcomes. Tawhid et al. [1] introduced an image-based autism detection model with a 95.25% accuracy using EEG data converted into spectrogram visuals and SVM classification. Kang et al. [16] used EEG data from 97 children, achieving an 85.44% accuracy with SVM based on eye-tracking tasks. Baygin et al. [17] proposed a feature extraction technique using short-term Fourier transform and ReliefF selection, resulting in a 96.44% accuracy with SVM. Han et al. [18] combined eye-tracking and EEG data using stacked autoencoders, outperforming other approaches.

Menaka et al. [19] explored ASD detection using pre-trained networks and 5 cepstral coefficient features. They creatively employed Cepstral Coefficients to build spectrograms and adapted AlexNet for enhanced classification. Their results showcased a customized AlexNet with Linear Frequency Cepstral Coefficients achieving a remarkable 90% accuracy, underscoring deep learning’s promise in early ASD detection. Tang et al. [20] introduced a hybrid graph neural network for precise ASD diagnosis using resting-state EEG signals. Their model effectively captures unique brain connectivity patterns, achieving high accuracies of 87.12% in single-subject and 85.32% in cross-experiment analyses. These findings underscore Rest-HGCN’s potential as an efficient tool for clinical ASD diagnosis.

Menaka et al. [21] examined EEG data from both control individuals and children with ASD to evaluate Detrended Fluctuation Analysis (DFA) as a tool for signal analysis. By assessing Hurst exponents derived from DFA, they gauged the self-similarity level in the EEG signals, probing DFA’s utility in early ASD detection. Another study [22], utilized a graph theory-based approach to examine small-world networks in ASD and typically developing brains, focusing on responses to Audio and Video Stimuli. Functional connectivity was assessed using coherence measures on graphically generated data. Yang et al. [23] analysed EEG data from ASD and control children watching emotional movie clips. They found significant differences in brain connectivity across all frequencies, shedding light on ASD’s neurobiological basis and informing targeted interventions.

In this study, a Graph Convolutional Neural Network (GCNN) model was proposed for ASD detection [24]. GCNNs excel in analysing complex data structures, such as graphs, commonly used to represent brain connectivity from EEG data. Their primary advantage over traditional neural networks lies in their ability to incorporate graph topology information, allowing for better modelling of relationships among EEG electrodes and signals, resulting in more informative features and accurate predictions. GCNNs are crucial in EEG-based predictions and serve as viable replacements for CNN models. They harness the graph structure to collect node information from neighbours. Variations of GCNN models have been applied in various EEG applications in psychiatry, including seizure detection [25, 26], intention detection [27], epileptic seizure detection [28, 29, 30], emotion detection [31, 32], motor imagery classification [30] and Alzheimer’s prediction [33, 34]. Unlike CNNs, which operate on uniform visual data patterns, GCNNs handle unpredictable, non-Euclidean data patterns with diverse graph frameworks.

Brain region-specific EEG studies are essential for predicting autism, a complex disorder marked by brain function and connectivity abnormalities [35]. EEG, a powerful neuroimaging tool, reveals differences in brain activity and connectivity in autistic individuals, particularly in key brain regions and neural circuits [36]. Focusing on these regions allows researchers to identify autism biomarkers, including activity and connectivity differences, specific neural patterns, and oscillations [37]. For example, gamma frequency differences in the frontal cortex have been observed in autistic individuals [38]. These studies are vital for understanding autism’s neural mechanisms and improving EEG-based prediction and diagnosis. See Table 1 for commonly studied brain regions in EEG research, and Fig. 1 for EEG electrode placement in primary brain regions.

Table 1
Primary brain regions and functionality with respect to EEGs

Brain regions	Functions
Frontal region	Controls character traits, reasoning, emotion, memory, and social interaction.
Temporal regions	Handle auditory and visual processing, short-term memory, speech, and musical rhythm.
Central region	An essential area for motor and sensory processing, often studied in EEG research.
Occipital region	Governs vision.
Parietal region	Involved in object identification, spatial cognition, and sensation/pain perception.

Figure 1.

Distribution of primary brain regions with respect to EEG electrode placement.

This study aimed to identify the most accurate features for diagnosing autism in its early stages and improving predictive accuracy by inputting these features into a GCNN. It also aimed to showcase the disparities in EEG signals between autistic patients and neurotypical controls.

This study aimed to achieve the following objectives:

•

Develop an efficient preprocessing method to identify 71 useful electrodes with their 10-10 international equivalents and generate predictive features for 257 HydroCel GSN EEG recordings.

•

Create an effective GCNN model for the detection of ASD in young children by combining graph representation with CNN, utilizing Pearson correlation adjacency matrices to establish graph connections.

•

The study introduces the novel integration of GCNN with EEG data for predicting ASD. This approach is cutting-edge, leveraging the capabilities of GCNNs to model complex relationships within brain connectivity networks. GCNNs offer a sophisticated way to learn meaningful representations directly from EEG data. Unlike traditional methods, GCNNs can capture subtle patterns or features indicative of ASD, potentially enhancing the accuracy of diagnosis.

•

Apply GCNN with autoregressive and spectral feature extraction techniques. This combination improves both the predictive performance and interpretability of the model, providing a comprehensive understanding of neurological abnormalities associated with ASD.

•

The study emphasizes the predictive relevance of brain regions, such as the five primary regions depicted in Fig. 1 and seven overlapping brain regions (e.g., fronto-polar, anterior-frontal, etc.), to pinpoint the major autism-affected brain regions.

The structure of the remainder of this article is as follows. The comprehensive approach that was followed is provided in Section 2. The results were analysed and then discussed in Section 3. The study is discussed in Section 4. The final thoughts are included in Section 5.

2. Methodology

This section discusses the materials and methods employed in this study. The suggested ASD detection system’s overall architecture is depicted in Fig. 2.

Figure 2.

Proposed workflow of the ASD detection system.

2.1 Dataset collection

This study included EEG data from eight autistic children (six males, two females) and eight control children (five males, three females), aged 6 to 12 years. Autistic participants were patients at the Central Institute of Psychiatry (CIP) in Ranchi, India, while control children were recruited from nearby areas. The procedures involving experiments on human subjects conducted in this research were carried out in accordance with the ethical standards set forth in accordance with the principles outlined in the Declaration of Helsinki of 1964 and its later amendments, or comparable ethical standards. Informed parental consent was obtained, and the study was ethically approved by the Institute Ethics Committee (IEC), CIP, Ranchi (IEC/CIP/2022-23/1710). EEG data were recorded using a 257-channel HydroCel GSN headset following the 10-20 international system, at a sampling rate of 1000 Hz. Resting state EEG recordings were collected for 15 minutes per subject in a soundproof room, with auditory stimuli at 20 Hz, 30 Hz, and 40 Hz presented via headphones to stimulate brain responses. These frequencies have been shown to potentially enhance cognitive function and alertness. The EEG acquisition process is illustrated in Fig. 3.

During EEG recordings, efforts were made to keep each child stable. ASD diagnosis was based on the Childhood Autism Rating Scale (CARS) [39] and the Indian Scale of Assessment of Autism (ISAA) [40], with CARS scores over 30 and ISAA scores over 70 indicating autism. Autistic children showed mild to moderate ASD based on these scores. Additional assessments included the Developmental Screening Test (DST) [41] for social skills and the Vineland Social Maturity [42] scale for the social quotient, which gives the IQ. The DST provided the mental development quotient, indicating a mental retardation score. Descriptive statistics for all subjects are presented in Table 2, while Table 3 contains ASD-related subject-specific statistics and measures, exclusive to the children with ASD.

Table 2
Descriptive statistics of the EEG data collected from autistic children and healthy subjects ( $N=$ 16, Autistic $=$ 8, healthy $=$ 8)

Descriptive statistics
	N	Minimum	Maximum	Mean		Std. deviation
Age	16	6	12	9	.500	2.309
Education	16	2	3	2	.062	0.250
Family history	16	1	2	1	.375	0.500
Head circumference	16	52	55	53	.575	0.986
Age of onset	8	3	6	4		1.069
Developmental quotient	8	44	87	57	.250	15.153
Social quotient	8	44	86	59	.500	16.248
CARS	8	30	41	35		3.818
ISAA	8	75	146	105	.875	22.267

Figure 3.

EEG signal acquisition flow.

Table 3

Descriptive statistics and measures of all participants ( $N=$ 16)

ID	Type	Age	Gender	Family history	Age of onset	Head circumference	Developmental quotient	Social quotient	CARS	ISAA
1	ASD	10	Male	Absent	5	55	87	86	31	86
2	ASD	12	Male	Absent	6	55	73	78	30	75
3	ASD	12	Male	Present	4	54.50	55	64	36.50	112
4	ASD	6	Male	Absent	3	52.50	52	51	37	118
5	ASD	6	Male	Absent	4	53	48	45	41	146
6	ASD	12	Male	Present	4	54.50	55	64	36.50	111
7	ASD	7	Female	Present	3	52.40	44	44	32	110
8	ASD	7	Female	Present	3	52.40	44	44	32	89
9	Control	7	Female	Absent	–	54.50	–	–	–	–
10	Control	8	Female	Absent	–	54	–	–	–	–
11	Control	10	Female	Absent	–	53	–	–	–	–
12	Control	11	Male	Absent	–	52	–	–	–	–
13	Control	12	Male	Absent	–	53	–	–	–	–
14	Control	10	Male	Present	–	54	–	–	–	–
15	Control	12	Male	Absent	–	54	–	–	–	–
16	Control	10	Male	Present	–	53.40	–	–	–	–

2.2 Data pre-processing

The pre-processing pipeline implemented in this study aimed to enhance the quality and reliability of the EEG signals for subsequent analysis. The process was initiated by conducting EEG signal pre-processing using MATLAB 2020a in conjunction with the EEGLAB toolkit [43]. The novel preprocessing algorithm is as follows:

Step 1: Removal of channel 257

The channel 257 was initially used as the reference channel. This removal was essential to prevent potential biases and inaccuracies in the subsequent analyses. Subsequently, the average reference value was computed from the remaining 257 channel electrodes. This computation provided a more stable baseline for the EEG signals, reducing the impact of individual channel variations on the overall signal.

Step 2: Downsampling

To further refine the EEG data, downsampling was performed, reducing the sampling rate from 1000 Hz to 128 Hz. This reduction in sampling rate not only helped to conserve computational resources but also effectively filtered out high-frequency noise, thus enhancing the signal-to-noise ratio. Additionally, a low bandpass filter with a cutoff frequency of 0.5 Hz was applied to further attenuate noise and ensure that only relevant signal components were retained.

Step 3: Noise removal

The EEG recordings are mostly corrupted with unwanted electrical interferences that was addressed using the CleanLine() function. This inbuild function of EEGLAB effectively removes periodic interference caused by power lines and other electrical sources, thus minimizing potential artifacts in the data. To mitigate the impact of artifacts further, a combination of manual inspection and EEGLAB’s automatic artifact removal method was employed. Manual inspection allowed for the identification and removal of obvious artifacts, while EEGLAB’s automatic method provided a systematic approach to detect and eliminate additional artifacts.

Step 4: Data decomposition

In this step the independent component analysis (ICA) was applied for the final data decomposition. ICA is a powerful technique that separates mixed signals into statistically independent components, making it particularly effective for isolating and removing artifact sources from EEG data. By employing ICA, the pre-processing pipeline ensured comprehensive artifact removal, thereby enhancing the integrity and reliability of the EEG data for subsequent analysis.

Step 5: Mapping of 257 HydroCel to 10-10 international standard

The 257 HydroCel electrodes initially used in the EEG recording were based on the 10-20 system, which is a widely accepted standard for electrode placement in EEG research. The 10-20 system divides the scalp into regions based on percentages of the distance between cranial landmarks, with electrodes placed at specific points within these regions. However, the 10-20 system employs a relatively large number of electrodes, which can lead to increased computational complexity and may not be necessary for certain analyses.

To address this issue, the electrode positions were mapped from the 10-20 system to the 10-10 international standard based on a pre-established standard [44]. The 10-10 system is a modification of the 10-20 system that includes fewer electrodes, focusing on key anatomical landmarks on the scalp. By mapping the electrode positions to the 10-10 system, the dimensionality of the EEG data was reduced, simplifying subsequent analysis and improving computational efficiency.

During the mapping process, only electrode positions that conformed to the 10-10 standard were retained for the final analysis. This involved identifying the closest corresponding positions in the 10-10 system for each electrode in the original 10-20 configuration. Electrodes that did not have a counterpart in the 10-10 system were excluded from further analysis.

Table 4
Selected channels of its 10-10 international equivalent for 10-20 HydroCel GSN

10-10	HydroCel	10-10	HydroCel	10-10	HydroCel	10-10	HydroCel
Frontal		Temporal		Frontal polar		Central parietal
F3	36	T3-T7	69	FP1	37	CpZ	81
F4	224	T4-T8	202	FP2	18	CP6	172
F7	47	T5-P7	96	FPZ	26	CP4	164
F8	2	T6-P8	170	Anterior frontal		CP3	66
F2	5	T9	68	AF8	10	Cp2	143
F6	222	T10	210	AF3	34	CP1	79
F9	252	Parietal		AF4	12	Cp5	76
F1	29	P3	87	Afz	20	Temporo-parietal
F10	226	P4	153	Afz	26, 27	TP8	179
F5	48	T5-P7	96	AF7	46	TP9	94
Fz	21	T6-P8	170	Frontal central		TP10	190
Central		Pz	101	FC5	49	Tp7	84
C3	59	P10	169	FC3	42	Parietal occipital
C4	183	P1	88	FC1	24	Poz	119
C6	194	P5	86	FC2	207	PO8	161
C5	64	P6	162	Fcz	15	PO3	109
C1	44	P2	142	FC4	206	PO4	140
C2	185	P9	106	FC6	213	Po7	97
CZ	256	Occipital		Frontal temporal
		O1	116	Ft10	219
		O2	150	Ft9	67
		Oz	126	FT8	211
				FT7	62

As a result of this mapping and selection process, a total of 71 electrodes were retained for the final analysis as given in Table 4. These electrodes were distributed across the scalp according to the 10-10 international standard, ensuring comprehensive coverage of key anatomical regions while reducing the computational burden associated with a larger electrode set. This approach allowed for more focused and efficient analysis of the EEG data, facilitating the extraction of meaningful insights from the recorded neural signals.

Step 6: Division into brain regions

Understanding and treating ASD is a complex endeavour due to the intricate nature of the condition, which involves multiple brain regions and their interactions. To address this complexity, it’s essential to focus on specific brain regions that are known to be implicated in ASD pathology. By dividing the EEG electrodes into distinct regions, researchers can target their analysis on these specific areas, allowing for a more nuanced examination of the neural activity associated with ASD. The division of electrodes into 12 regions, namely, five primary (frontal, temporal, central, occipital, parietal) and seven overlapping regions (fronto-polar, anterior-frontal, etc.) serves several purposes:

Figure 4.

Steps followed for EEG pre-processing in EEGLAB.

•

Localization: Dividing electrodes into 12 regions enables precise localization of brain activity, essential for understanding ASD’s neural mechanisms.

•

Network Identification: Overlapping regions help identify altered functional networks, shedding light on disrupted brain circuits in ASD.

•

Feature Extraction: Organized regions facilitate feature extraction, aiding in the identification of spatially localized biomarkers and neural activity patterns.

•

Comprehensive Analysis: Analysing EEG data from multiple regions enables a comprehensive examination of ASD, uncovering variations in neural activity and connectivity crucial for understanding the disorder’s complexity.

Table 4 lists the brain regions and corresponding electrodes, providing a reference for the spatial organization of the EEG data. Figure 5 displays the positions of the electrodes on the scalp, illustrating their distribution across different regions. By extracting features from each of these regions and analysing them collectively, researchers can gain a more nuanced understanding of the neural correlates of ASD and develop more targeted interventions for individuals affected by the condition.

Figure 5.

257 channel HydroCel GSN with color-coded selected regions based on its 10-10 international equivalent. F-Frontal, T-Temporal, C-Central, O-Occipital, P-Parietal, FP-Fronto-polar, AF-Anterior-frontal, FC-Fronto-central, FT-Fronto-temporal, CP-Central-parietal, TP-Temporo-parietal, PO-Parietal-occipital.

2.3 Feature extraction

Feature extraction is a critical step in analysing EEG signals due to their intricate nature and susceptibility to noise. By extracting relevant features, we can reduce the dimensionality of the data while retaining essential information, making subsequent analysis more manageable and insightful.

We applied two main techniques for feature extraction: autoregressive (AR) modelling and spectral analysis, offering several advantages:

•
AR models are particularly useful for capturing temporal dependencies within EEG signals. They characterize the relationship between each data point and its preceding values, providing insights into how brain activity evolves over time. This temporal perspective is invaluable for tracking dynamic changes in neural activity associated with cognitive processes, stimuli, or neurological conditions.
•
Spectral analysis is instrumental in unveiling the frequency content of EEG signals. Different frequency bands correspond to distinct brain states and functions. For instance, delta waves are prominent during deep sleep, while alpha rhythms dominate during relaxed wakefulness. Spectral features allow us to characterize these frequency-specific patterns, aiding in the identification of abnormalities or deviations from normal brain activity.

Both AR and spectral features offer interpretability, which is crucial for understanding EEG findings in a clinical or research context. By interpreting these features, we can gain insights into underlying neural processes and cognitive functions. Additionally, the interpretability of these features facilitates the detection of different consciousness states and neurological disorders characterized by specific EEG signatures.
2.3.1 Autoregressive features

AR features are commonly used in EEG signal analysis [45]. They are statistical models for time series data like EEG signals, capturing temporal dependencies to predict future values based on past ones. AR models extract features reflecting underlying brain activity, facilitating classification of consciousness states (e.g., waking, sleeping, coma) and detection of abnormal brain activity hinting at neurological disorders.

The simplest definition of an AR model is a linear regression of the most recent data in a series against one or more previous observations. The AR model is typically expressed as follows:

$\displaystyle z(t)=\mathop{\sum}\limits_{j=1}^{s}\phi_{j}z({t-1})+\lambda_{t}$ (1)

where the time series to be modelled is denoted by $z(t)$ , and the coefficients of the model are denoted as $\phi_{j}$ . $\lambda_{t}$ denotes the white noise and $s$ denotes the order of the model. The time is represented by $t$ [45].

When evaluating performance, selecting the AR model order is crucial. Different methods exist to determine this, each with pros and cons. Higher orders provide better estimation but increase computational costs. In our study, we opted for an order of three.

2.3.2 Spectral features

In contrast, spectral features are derived from EEG signal frequency content, often extracted using Fast Fourier Transform (FFT) methods [46]. These techniques assess EEG recordings mathematically, capturing signal characteristics through a power spectral density (PSD) approximation. Four frequency bands were obtained as features: $\beta$ (14–30 Hz), $\theta$ (4–7.9 Hz), $\alpha$ (8–13.9 Hz), and $\delta$ (0.1–3.9 Hz), each linked to distinct brain activity.

The PSD was obtained from the nonparametrically calculated autocorrelation waveform. One strategy for reaching this objective is Welch’s strategy. The data sequence underwent windowing, which produced updated periodograms. The following data sequence is given for $x_{j}(n)$ :

$\displaystyle x_{k}(n)=x({n+kD})$ (2)

where $n$ is the sequence $0,1,2,\ldots,R-1$ and $k$ is the sequence $0,1,2,\ldots,S-1$ . Sequence $k$ begins with the variable $k D$ . $R$ , which has length $2S$ , is used to represent the data segment [46]. Following the average of the revised periodograms, the Welch power spectrum, $P^{T}$ [46], is given as follows:

$\displaystyle P^{T}(f)=\frac{1}{RT}\left|{\mathop{\sum}\limits_{n=0}^{R-1}x_{i% }(n)t(n)e^{-k2\pi fn}}\right|^{2}$ (3)

Where $t(n)$ is the windowing function and $T$ is the power normalisation factor.

2.4 Proposed graph convolution neural network (GCNN)

GCNN model was applied after the feature extraction phase, by utilizing the autoregressive and spectral features, to enhance model interpretability and effectiveness. While deep learning methods can extract features autonomously, prior feature extraction provides a structured representation of EEG data. Autoregressive and spectral features capture specific EEG characteristics relevant to ASD, aligning with domain knowledge. This preprocessing enriches the input representation, guiding the GCN to capture subtle ASD-related patterns and potentially improving generalization and performance. Thus, the feature extraction phase is crucial for optimizing the GCN’s efficacy in ASD prediction.

GCNNs differ from traditional CNNs by performing convolutional operations on graphs instead of regular grids. This allows the model to capture relationships between nodes and edges, enhancing prediction accuracy for graph data. The GCNN model aimed to understand EEG activity by considering both functional and structural pairings of brain areas. To provide input, a graph structure was constructed using EEG features and adjacency matrices. Graph convolutions were applied to generate node-level embeddings, followed by node-level averaging to create graph-level embeddings. Two graph convolution layers were used, and a fully connected network made classification predictions. Figure 6 outlines the GCNN model workflow, and subsequent sections define its unique characteristics mathematically.

Figure 6.

Proposed graph convolution neural network for autism detection.

2.4.1 Basic graph structures

EEG recordings are often depicted as 2D mappings with the EEG nodes and time acting as the two dimensions. Previous studies used CNNs on unprocessed 2D mappings without considering the relationships between different EEG nodes. Because each EEG node represents the region of the brain from where the sensory signal is obtained, describing the connections between various EEG nodes is crucial for effective EEG investigation. EEG electrode features were utilised instead of EEG electrodes.

Each graph representation comprises three components: the edges (i.e., nodes connection), the graph signals and the nodes (i.e., the EEG channels, which are the sets of features extracted from the EEG signals recorded at the nodes). $G=\{{V,E,A}\}$ is used to denote the graph’s structure, with $V$ denoting the nodes of the graph. In addition, $|V|=N$ where $N$ is the overall number of nodes. $E$ denotes the graph’s edges and $A\in\mathbb{R}^{N\times N}$ stands for the adjacency matrix that defines the connections between the nodes.

2.4.2 Adjacency matrix

The link between two nodes is typically represented in the standard adjacency matrix by 0 or 1; however, this depiction is too basic to accurately portray node relationships. Therefore, the absolute Pearson correlation coefficient was calculated to build an adjacency matrix to illustrate the relationship between the nodes of the graph. The absolute Pearson correlation coefficient $P_{yz}$ [47] between the features $y$ and $z$ is given by:

$\displaystyle P_{yz}=\textit{abs}\left[{\frac{\sigma_{yz}}{\sigma_{y}\sigma_{z% }}}\right]$ (4)

where $\sigma_{yz}$ signifies the covariance and $\sigma_{y}$ , $\sigma_{z}$ denotes the standard deviation of the population. Next, the adjacency matrix $A_{yz}$ [48] is calculated by subtracting the absolute Pearson correlation coefficient from the eye matrix $(I)$ :

$\displaystyle A_{yz}=P_{yz}-I$ (5)

2.4.3 Spectral graph filtering

Spectral graph filtering is a mathematical operation used in GCNNs to process graph-structured data. The spectral graph filtering operation applies a filter to the graph data in the frequency domain, which can be calculated using the graph Laplacian matrix. The Laplacian matrix is a matrix that captures the connectivity structure of the graph, and the spectral decomposition of this matrix can be used to define a frequency domain representation of the graph signals. In this study, the graph Laplacian matrix was used for the graph representation. The graph Laplacian matrix $L_{yz}$ [48] is calculated using the formula below:

$\displaystyle L_{yz}=D-A_{yz}$ (6)

Where, $D$ is the degree matrix that records the degree of each vertex, or the amounts of edges that connect each vertex to other vertex and is calculated as $D_{y,y}=\mathop{\sum}\limits_{z}A_{y,z}$ [48].

A diagonalization of the real symmetric matrix $L$ is

$\displaystyle L_{yz}=V\Lambda V^{T}$ (7)

where eigenvalues of $L_{yz}$ make up the diagonal matrix $\Lambda$ , and $V=[{v_{1},v_{2},v_{3},\ldots v_{n}}]$ is a pack of orthogonal bases of $L_{yz}$ [48].

The inversion of graph data $y$ and the Fourier transform [48] can be stated as:

$\displaystyle\tilde{y}=V^{T}y$ (8) $\displaystyle y=V\tilde{y}$ (9)

The following is the definition of the convolutional operation between graph data $x$ and convolutional kernel $k$ :

$\displaystyle k\ast y=V({({V^{T}k})\odot({V^{T}y})})=V\tilde{k}V^{T}y$ (10)

Where, $k$ is a diagonal matrix with $n$ spectral convolution factors and $\odot$ denotes Hadamard product operation [48].

Considering that the input and output graph data are $y$ and $z$ , respectively, and that the graph filter is $g(\cdot)$ , after filtering $y$ can then be expressed as follows:

$\displaystyle z=g({L_{yz}})y=g({V\Lambda V^{T}})y=Vg(\Lambda)V^{T}y$ (11)

Where, $g({\Lambda})$ is given as,

$\displaystyle g({\Lambda})=\left[{{\begin{array}[]{*{20}c}{g({\xi_{1}})}&% \ldots&0\\ \vdots&\ddots&\vdots\\ 0&\ldots&{g({\xi_{N}})}\\ \end{array}}}\right]$ (12)

Where, $\xi$ is the eigen value of the Laplacian matrix [48]. We may obtain $g(L)$ with K-order Chebyshev polynomials [48] as,

$\displaystyle g({\Lambda})=\mathop{\sum}\limits_{m=0}^{m-1}\theta_{m}T_{m}(% \tilde{\Lambda})$ (13)

Where, $\theta_{m}$ is the coefficient of Chebyshev polynomials and $T_{m}(.)$ is the calculation method [48].

2.4.4 Graph convolutions layer

The process entails utilizing a graph convolution layer to perform convolutions on graph signals and extract features from visualizations, employing filters rooted in spectral graph theory. Two convolution layers are superimposed on the graph, updating node signals through averaging with connected nodes using trained graph convolutional filters. Each layer generates an updated graph with a consistent structure. To enhance graph signals, information from both signals and the adjacency matrix is integrated using each layer’s convolutional filter. The algorithm for the proposed GCNN model is summarized in Algorithm 1.

Algorithm 1: Proposed graph convolution neural network
Input: Input EEG features, corresponding labels of class, the maximum number of iterations $m$ , learning rate $l$ , Chebyshev polynomials $c$ , halt iteration cut-off $s$ ;
Output: The required GCNN variables;
for $j=1,2,3\ldots n$ do
Compute the adjacency matrix $A_{yz}$ using Eq. (5);
Compute the Laplacian matrix using Eq. (6);
Compute the Chebyshev polynomials ( $c=1,2,3\ldots,c-1$ ) using Eq. (13);
Calculate $\mathop{\sum}\limits_{c=0}^{c-1}\theta_{c}T_{c}({\widetilde{L_{yz}}})y$
End
epoch $=$ 0;
while Loss $>s\|\|$ epoch $<m$ do
Compute output results of convolution;
Regularize output using ReLU;
Compute fully connected layer results;
Compute loss;
Update the parameters of graph convolution;
epoch $=$ epoch $+$ 1;
end while

3. Results

In this study, we aimed to automatically detect ASD using EEG signals. For this purpose, a novel pre-processing step was developed to reduce the computation and complexity of the system. A 257-channel GSN HydroCel headset was used to record EEG data. EEGLAB was used to pre-process the EEG signals. Statistical analyses were performed using SPSS software. Python 3.8.10 was used to implement the GCNN classifier, which was tested and trained on an NVIDIA GeForce GTX 1050 GPU. Brain region-specific prediction was carried out to determine the presence of ASD. The dataset was split between a training group and a testing group in an 80:20 ratio.

Table 5
Hyperparameters used in the proposed GCNN model

Parameter	Value
Number of epochs	1000
Batch size	10
Evaluation frequency	10
Building block
Filter	chebyshev5
Activation function	ReLU
Pool	mpool1
Architecture
Number of graph convolution filters	[64, 128]
Polynomial orders	[2, 2]
Pooling sizes	[2, 2]
Output dimensionality of fully connected layers	[4]
Optimization
Regularization	0.000001
Dropout	0.50
Learning rate	0.000001
Decay rate	1 i.e., no decay
momentum	0 i.e., Adam optimizer

The parameters for the GCNN model in all the experiments are shown in Table 5. The chosen hyperparameters for the GCNN model were carefully selected to optimize model performance. The learning rate was set to 0.000001, which is a small value suitable for fine-tuning the model parameters while avoiding large oscillations in the loss function during training. This value was chosen empirically through experimentation to balance learning speed and stability. The model was trained for 1000 epochs, indicating the number of passes through the entire dataset during training. This value was chosen to ensure sufficient iterations for the model to converge to an optimal solution without overfitting. The Adam optimizer was used with a momentum of 0, which means standard Adam optimization without momentum adjustment. Adam is known for its effectiveness in training deep neural networks and was chosen for its robust performance in optimizing GCNN parameters. These hyperparameters were tuned iteratively based on model performance on a validation set. Evaluation occurred every 10 epochs, allowing for the monitoring of model progress and potential overfitting. Regularization with a regularization rate of 0.000001 and a dropout rate of 0.50 were applied to prevent overfitting by penalizing large parameter values and randomly dropping units during training, respectively. These values were chosen to strike a balance between preventing overfitting and preserving model capacity. The architecture of the GCNN included two graph convolutional layers with filters based on Chebyshev polynomials of order 2, followed by max pooling layers with a pool size of 2. The output dimensionality of the fully connected layers was set to 4 to match the desired output dimensionality. To sum up, these hyperparameters were selected based on empirical testing and domain knowledge to ensure effective training and optimal performance of the GCNN model for the specific task at hand.

Brain region-specific prediction was carried out to determine which region was most capable of identifying ASD. Autoregressive and spectral features were extracted for each brain region, and separate adjacency matrices and graph structures were constructed for each brain region. The GCNN was trained and tested for each brain region separately.

Five performance metrics were used to assess the outcomes, namely, accuracy, precision, recall, F1-score, Area Under the Curve (AUC) an its corresponding Receiver Operating Curve (ROC). The definitions of these metrics are beyond the scope of this study; therefore, the details can be found in [49].

Table 6

Representation of classification performance across brain regions using a GCNN based on testing data

	Region	Accuracy	Precision	Recall	F1score	AUC
1	Frontal	70.384	74.793	66.058	70.155	70.630
2	Temporal	71.042	68.543	79.007	73.404	70.949
3	Central	69.615	66.049	81.679	73.037	69.521
4	Parietal	73.269	75.190	72.693	73.921	73.294
5	Occipital	72.692	72.789	77.536	75.087	72.374
6	Anterior-frontal	87.065	82.624	92.828	87.429	87.238
7	Fronto-central	67.692	69.852	68.840	69.343	67.617
8	Central-parietal	60.000	61.129	66.909	63.888	59.576
9	Parietal-occipital	75.628	88.461	60.526	71.875	76.079
10	Fronto-temporal	49.420	56.383	19.343	28.804	51.269
11	Temporo-parietal	51.063	66.666	13.284	22.153	52.983
12	Frontal polar	76.878	91.351	61.904	73.799	77.700

From the test data, Table 6 displays the values of the performance measures obtained from the GCNN classifier applied across various brain areas. Comparing the results from Table 6, the anterior-frontal region has the best predictive capability among all brain regions, with 87.07% accuracy. The frontal-polar and parietal-occipital regions were the second- and third-best-performing brain regions, with accuracies of 76.88% and 75.63%, respectively.

Figure 7.

Validation accuracy and training loss graph of the GCNN model for brain regions (a) anterior-frontal (b) frontal-polar (c) parietal-occipital. Where the x-axis represents the epochs over time. The blue curve represents validation accuracy over successive epochs based on a validation dataset, whereas the green curve represents training loss over successive epochs based on training data.

The validation accuracy and training loss graphs in Fig. 7 depict the learning progress and generalization ability of the GCNN model for the anterior-frontal, frontal-polar, and parietal-occipital brain regions. The x-axis represents the epochs that indicates the progression of time or iteration during a training phase. Plotting the validation accuracy against the training loss curve is a common practice in machine learning, representing the model’s performance on unseen data. Higher validation accuracy indicates better generalization and prediction accuracy on new data. On the other hand, training loss reflects the model’s prediction errors during training, with the goal of minimizing it. It can be observed from the figure that the training loss is close to zero, which is an ideal situation for a good-performing classifier. Furthermore, Fig. 7 highlights a concurrent increase in validation accuracy as the training loss diminishes. This observed pattern is indicative of the model’s ability to generalize well to unseen data, as evidenced by its improved performance on the validation set.

Figure 8.

ROC curve obtained for the GCNN model across brain regions.

Figure 8 shows the ROCs obtained from the GCNN model for different brain regions. The ROC employs the probability of informing us of a model’s capacity to distinguish between classes. The threshold used is 0.5. The model performance is considered perfect if the ROC value is found to be close to 1. A model with poor performance is said to have an ROC value of 0.5 or below. Figure 8 shows that the fronto-temporal and temporo-parietal regions were the worst-performing models with ROC near 0.5.

4. Discussion

Artificial intelligence (AI) is transforming the prediction and diagnosis of psychological disorders and neurological conditions, leveraging diverse datasets including genetic, neuroimaging, and clinical data. Developing an effective AI-based tool for ASD screening and diagnosis in children is imperative. While many AI methods utilize EEG data for ASD diagnosis, their effectiveness requires enhancement. This study introduced a brain region-specific GCNN classifier to improve understanding of ASD-related disparities in brain activity and enhance the diagnostic process. The aim was to identify predictive neural brain regions. By utilizing region-specific electrodes, dimensionality and complexity were reduced. Previous disease prediction studies mainly focused on the five basic brain regions, with limited exploration of cross-region predictions. This research is pioneering, being the first to incorporate cross-region ASD prediction.

Like our research, Avdakovic et al. [50], utilized global wavelet power spectrum to predict epilepsy using EEG. Findings confirm that GWS enables straightforward identification of epileptic EEG signals, revealing substantial differences in signal components between patients with epilepsy and healthy subjects, with dominant values observed in the delta and theta frequency bands for epileptic patients. Another study on meningitis [51] highlights the potential of AI in diagnosing a serious neurological condition, just as research on autism prediction underscores the importance of leveraging technology to address complex neurological disorders and enhance clinical decision-making.

Psychometric testing remains the primary method for ASD prediction and diagnosis [52, 53], with early screening instruments well-documented in the literature [54]. However, these tests are not suitable for newborns due to inherent limitations. Recent research has explored the effectiveness of biological assessments in early screening [55]. Early developmental stages are crucial for identifying brain structure and function anomalies in ASD patients. Promising physiological and neuroimaging modalities, including Functional magnetic resonance imaging (fMRI) [56, 57], an EEG [58] and Functional near-infrared spectroscopy (fNIRS) [59, 60], offer high predictive accuracy as screening methods for children.

To our knowledge, no previous research has conducted overlapping brain region-specific analysis using EEG signals for ASD prediction. Understanding both structural and functional brain changes in individuals

Table 7
Comparison between the current work with previous works

Year	Dataset	Age (years)	Channels/ Brain regions	Assessment tool	Feature extraction	Classifier	Test accuracy (%)
2019 [69]	Own	4–8	2 (C3, C4)	Not mentioned	Topological features	k-nearest neighbour (kNN)	Up to 81%
2020 [16]	Own	3–6	62	Psychoeducational profile (Third Edition), DSM-5	Power spectrum	SVM	85.44%
2020 [70]	Own	3–12	2 (C3, C4)	CARS	Linear and Nonlinear Features	kNN	72.77%
2021 [71]	Own	11–18	15 (frontal, temporal, parietal, occipital)	Autism Diagnostic Interview-Revised (ADI-R)	Statistical and discrete wavelet transform features	Neural Network	84.66%
2022 [18]	Own	3–6	62	DSM-5	Relative power energy, entropy, brain network	Stacked denoising autoencoders	81.11%
2022 [72]	Own	6–32	6 (occipital)	ADOS	Univariate and multivariate	Elastic net logistic regression	54%
					features	Linear SVC	64%
2022 [73]	Own	18–55	64	DSM-IV, DSM-5	Spectral, criticality and functional connectivity	SVM, logistic regression, random forest	56%
2023 [19]	Own	5–7	8 (Occipital, temporal, frontal)	DSM V	Cepstral coefficients’ spectrograms	AlexNet with Linear Frequency Cepstral Coefficients	85.1%
Proposed work	Own	6–12	71 (5 primary and 12 overlapping regions)	CARS, ISAA	Autoregressive and spectral features	GCNN	87.06%

Table 8

Comparison of current work to previous work using GCNN as predictor with fMRI as modality

Year	Modality	Dataset	Test accuracy
2018 [34], 2021 [75], 2022 [66], 2022 [67],	fMRI	ABIDE	70.4%, 73.7%, 69.38%, 75.86%, 82.20%, 70.10%
2022 [78], 2022 [79]			respectively
Proposed method	EEG	Own	87.06%

with ASD is crucial. Previous studies have reported abnormal frontal and temporal lobe growth, altered grey matter, white matter, and amygdala volume in young children with ASD, consistent with current findings [61]. Research in individuals of all ages with ASD has shown both increases and decreases in cortical thickness [62]. For instance, Piven et al. observed reduced cortical thickness in various brain regions, while Hardan et al. found thicker cortical layers in individuals with ASD [63]. Voxel-based morphometry studies also reported differences in grey and white matter volume in frontal, parietal, and temporal lobes in ASD participants [64]. Our study highlights the significance of the anterior-frontal, frontal-polar, and parietal-occipital brain regions, responsible for personality development, emotions, locomotion, speech, and cognition, which are often underdeveloped in autistic children.

The Autism Diagnostic Observation Schedule [65], Autism Diagnostic Interview-Revised [66], and Diagnostic and Statistical Manual of Mental Disorders [67] are commonly employed by psychiatrists for ASD diagnosis. Predicting autism frequently relies on functional fMRI and questionnaire-based methods. Notably, only 4% of studies, as per Hosseinzadeh et al.’s review [68], considered EEG as a sensor for ASD prediction. The majority of approaches (54%) involved wearable sensors, smartphones, and smartwatches. Table 7 provides a comparison of our study’s results with existing research.

GCNNs have gained popularity, particularly in medical applications like autism prediction, where fMRI is predominantly used [74]. These models leverage fMRI data, representing it as a graph with nodes denoting brain regions and edges indicating functional connections. Table 8 compares our study, which uses EEG data, with previous research utilizing fMRI and GCNNs for ASD prediction. Notably, the GCNN with EEG data outperforms its fMRI counterpart in terms of accuracy.

It’s essential to recognize biological variations with clinical significance in understanding ASD dynamics [2]. Studies have linked ASD to various anatomical and functional features, observed histologically and through neuroimaging, compared to typically developing children [80]. Combining multiple neuroimaging modalities has shown greater diagnostic sensitivity, and future research should emphasize this approach. Additionally, meta-regression indicated that EEG exhibited higher sensitivity in ASD classification [81]. Theorized network architecture in ASD includes increased short-range connections between frontal and parietal/temporal lobes and reduced long-range coherence, especially between frontal and occipital lobes [82].

This study has few drawbacks. Integrating young children with ASD into research can be challenging due to concerns about privacy and stigma. Enhancing prediction accuracy may involve using a different GCNN variant. ASD studies face the limitation of overlapping symptoms with other mental disorders, leading to potential inaccuracies in predictions. Accurate diagnosis is challenging due to symptom overlap, and errors in diagnosis have significant consequences. To address these limitations, future research could explore multiclass classification approaches that encompass various mental disorders, allowing for a more nuanced understanding of symptomatology and improving diagnostic accuracy. Additionally, efforts to mitigate recruitment challenges and ensure participant privacy could involve community engagement initiatives, partnerships with relevant organizations, and the implementation of rigorous privacy protocols. While this study has made valuable contributions to ASD prediction using EEG data, acknowledging and addressing its limitations are essential for advancing the field and ensuring the validity and reliability of findings.

While the participant sample size may have been small, it’s essential to recognize that the EEG data collected during the 15-minute recording sessions provided a substantial amount of data for analysis, even after preprocessing. However, the study’s main limitation is the small sample size, which deserves attention. A limited number of participants restricts the generalizability of findings beyond the studied group. This constraint also reduces statistical power, potentially limiting the reliability of conclusions. Furthermore, the small sample size may hinder the detection of subtle patterns in the data, limiting the depth of insights gained.

Even though the study’s findings may provide valuable insights into ASD prediction using EEG data within the sampled population, extrapolating these conclusions to a wider population may be tenuous due to the small sample size. This limitation could potentially undermine the robustness and reliability of the findings when applied to different demographics or clinical settings. To address this concern, future research should aim to replicate the study with larger and more diverse samples to enhance the generalizability of the findings. Additionally, conducting multicenter studies involving multiple institutions and collaborating with research consortia could facilitate the recruitment of larger and more representative samples, thereby improving the external validity of the findings.

5. Conclusion

ASD is a widespread neurodevelopmental disorder, often accompanied by intellectual disabilities in 10% to 33% of cases [83]. Timely diagnosis is essential, but current methods have limitations, leading to delayed diagnoses. While tools like MRI, fMRI, and EEG have been explored, their lack of specificity, sensitivity, and associated risks hinders routine use. Our study focuses on using EEG signals for ASD detection using GCNN.

We introduce novel pre-processing techniques to streamline complexity and resource usage. Employing channels with 10-10 international equivalents, we pioneered brain region-specific classification. Feature extraction involved autoregressive and spectral features. Our proposed two-layer GCNN network leverages convolutional neural networks and graph representation for ASD prediction. GCNNs excel in analysing non-Euclidean structures like graphs, making them ideal for studying brain connectivity data. Our research represents a groundbreaking exploration of brain region-specific ASD detection using EEG signals.

EEG has long been recognized as a valuable tool in ASD research; however, the brain-region-specific approach represents a significant advancement, providing deeper insights into the neural mechanisms underlying the disorder. Through the combination of EEG’s strengths with a targeted focus on specific brain regions, this approach holds promise for advancing our understanding of ASD and enhancing diagnostic and therapeutic outcomes for individuals affected by the condition. Additionally, by pinpointing the brain regions most predictive of ASD, this approach facilitates the development of devices capable of selectively capturing EEG information from electrodes belonging to the most predictive brain-region, streamlining diagnostic processes, and enhancing the effectiveness of interventions.

The findings of this study align with existing theories on brain network abnormalities in ASD. One prevalent theory suggests that ASD is characterized by atypical connectivity patterns within and between brain regions, leading to impaired cognitive functions such as emotion processing, memory, and social interaction [84]. The study’s focus on the anterior-frontal brain region, known for its involvement in higher-order cognitive functions and social cognition, reinforces the notion that abnormalities in this area are associated with ASD. The high predictive accuracy achieved using EEG signals from this region further supports the idea that disruptions in neural networks involved in social and cognitive processing are central to ASD pathology.

Leveraging sensory-based data, such as EEG and fNIRS, for ASD detection remains an underexplored area, offering ample opportunities for further research. Given the rising prevalence of ASD and the growing need for ASD prediction models, future work should explore combining neuroimaging and physiological signals with deep learning to achieve high prediction accuracy. In summary, leveraging advanced technologies for early ASD identification enables highly precise classification, even in infants. Early diagnosis facilitates timely therapeutic interventions with potential long-term benefits.

Funding

We declare that this research work was supported by DST PURSE 2022.

Ethical approval

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Informed consent

Informed consent was obtained from the parents of the children.

Data availability

The data will be available on request.

Footnotes

Acknowledgments

We would like to acknowledge Dr Sanjay Munda and Dr Ashwani Garg for their continuous support over the work.

Conflict of interest

The authors report that there is no conflict of interest.

References

Sealey

Hughes

Sriskanda

Guest

Gibson

Johnson-Williams

Pace

Bagasra

. Environmental factors in the development of autism spectrum disorders. Environment International. 2016 Mar 1; 88: 288-98.

Lord

Elsabbagh

Baird

Veenstra-Vanderweele

. Autism spectrum disorder. The Lancet. 2018 Aug 11; 392(10146): 508-20.

Onaolapo

. Global data on autism spectrum disorders prevalence: A review of facts, fallacies and limitations. Univers J Clin Med. 2017; 5(2): 14-23.

Kabot

Masi

Segal

. Advances in the diagnosis and treatment of autism spectrum disorders. Prof Psychol Res Pract. 2003; 34(1): 26-33.

Subha

Joseph

Acharya

Lim

. EEG signal analysis: A survey. Journal of Medical Systems. 2010 Apr; 34: 195-212.

Gurau

Bosl

Newton

. How useful is electroencephalography in the diagnosis of autism spectrum disorders and the delineation of subtypes: A systematic review. Frontiers in Psychiatry. 2017 Jul 12; 8: 121.

Chattaraj

Chakraborty

Koner

Gupta

. Machine learning in ASD: An intensive study of automated disease prediction system. Agile Software Development: Trends, Challenges and Applications. 2023 Feb 17: 345-62.

Chakraborty

Mali

. An overview of biomedical image analysis from the deep learning perspective. Research Anthology on Improving Medical Imaging Techniques for Analysis and Intervention. 2023; 43-59.

Janiesch, Christian, Patrick Zschech, and Kai Heinrich. Machine learning and deep learning. Electronic Markets. 2021; 31(3): 685-695.

10.

Das

Zomorrodi

Mirjalili

Kirkovski

Blumberger

Rajji

Desarkar

. Machine learning approaches for electroencephalography and magnetoencephalography analyses in autism spectrum disorder: A systematic review. Progress in Neuro-Psychopharmacology and Biological Psychiatry. 2022 Dec 24: 110705.

11.

Mehta

Menaka

Karthik

Thanga Aarthy

. Analysis of Connectivity Model to Study the Neurophysiological Process for Autism Detection. InInternational Virtual Conference on Industry 4.0: Select Proceedings of IVCI4. 0 2020 2021. pp. 191-200. Springer Singapore.

12.

Ari

Sobahi

Alçin

ÖF

Sengur

Acharya

. Accurate detection of autism using Douglas-Peucker algorithm, sparse coding based feature mapping and convolutional neural network techniques with EEG signals. Computers in Biology and Medicine. 2022 Apr 1; 143: 105311.

13.

Jahmunah

Arunkumar

Abdulhay

Gururajan

Adib

Ciaccio

Cheong

Acharya

. A novel automated autism spectrum disorder detection system. Complex & Intelligent Systems. 2021 Oct; 7(5): 2399-413.

14.

Lavanga

De Ridder

Kotulska

Moavero

Curatolo

Weschke

Riney

Feucht

Krsek

Nabbout

Jansen

. Results of quantitative EEG analysis are associated with autism spectrum disorder and development abnormalities in infants with tuberous sclerosis complex. Biomedical Signal Processing and Control. 2021 Jul 1; 68: 102658.

15.

Tawhid

Siuly

Wang

. Diagnosis of autism spectrum disorder from EEG using a time-frequency spectrogram image-based approach. Electronics Letters. 2020 Dec; 56(25): 1372-5.

16.

Kang

Han

Song

Niu

. The identification of children with autism spectrum disorder by SVM approach on EEG and eye-tracking data. Computers in Biology and Medicine. 2020 May 1; 120: 103722.

17.

Baygin

Dogan

Tuncer

Barua

Faust

Arunkumar

Abdulhay

Palmer

Acharya

. Automated ASD detection using hybrid deep lightweight features extracted from EEG signals. Computers in Biology and Medicine. 2021 Jul 1; 134: 104548.

18.

Han

Jiang

Ouyang

. A multimodal approach for identifying autism spectrum disorders in children. IEEE Transactions on Neural Systems and Rehabilitation Engineering. 2022 Jul 19; 30: 2003-11.

19.

Menaka

Karthik

Saranya

Niranjan

Kabilan

. An improved AlexNet model and cepstral coefficient-based classification of autism using EEG. Clinical EEG and Neuroscience. 2024 Jan; 55(1): 43-51.

20.

Tang

Zhang

Chen

Yang

Chen

Gao

. A hybrid graph network model for ASD diagnosis based on resting-state EEG signals. Brain Research Bulletin. 2024 Jan 1; 206: 110826.

21.

Radhakrishnan

Ramamurthy

Kothandaraman

Madaan

Machavaram

. Investigating EEG signals of autistic individuals using detrended fluctuation analysis. Traitement du Signal. 2021 Oct 1; 38(5).

22.

Menaka

Karthik

Parthasarathy

Manideep

Varsha

. Coherence Analysis in the Brain Network of ASD Children using Connectivity Model and Graph Theory. Journal of Scientific and Industrial Research. 2022 Sep 1: 940-51.

23.

Yang

Chen

Cai

. A preliminary study of brain connectivity in autism using affection-evoked movie clips. In 2021 IEEE International Conference on Engineering, Technology & Education (TALE). IEEE. 2021 Dec 5. pp. 1135-1138.

24.

Zhang

Tong

Maciejewski

. Graph convolutional networks: A comprehensive review. Computational Social Networks. 2019 Dec; 6(1): 1-23.

25.

Raeisi

Khazaei

Croce

Tamburro

Comani

Zappasodi

. A graph convolutional neural network for the automated detection of seizures in the neonatal eeg. Computer Methods and Programs in Biomedicine. 2022 Jun 10: 106950.

26.

Meng

Deng

. Electrical status epilepticus during sleep electroencephalogram waveform identification and analysis based on a graph convolutional neural network. Biomedical Signal Processing and Control. 2022 Aug 1; 77: 103788.

27.

Zhang

Yao

Chen

Wang

Haghighi

Sullivan

. A graph-based hierarchical attention model for movement intention detection from EEG signals. IEEE Transactions on Neural Systems and Rehabilitation Engineering. 2019 Sep 24; 27(11): 2247-53.

28.

Wang

Shi

Cheng

Wei

Chen

Zhou

. A Spatiotemporal Graph Attention Network Based on Synchronization for Epileptic Seizure Prediction. IEEE Journal of Biomedical and Health Informatics. 2022 Nov 10.

29.

Dissanayake

Fernando

Denman

Sridharan

Fookes

. Geometric Deep Learning for Subject Independent Epileptic Seizure Prediction Using Scalp EEG Signals. IEEE Journal of Biomedical and Health Informatics. 2021 Jul 27; 26(2): 527-38.

30.

Sun

Zhang

. Adaptive spatiotemporal graph convolutional networks for motor imagery classification. IEEE Signal Processing Letters. 2021 Jan 8; 28: 219-23.

31.

Jia

Zhang

. CR-GCN: Channel-Relationships-Based Graph Convolutional Network for EEG Emotion Recognition. Brain Sciences. 2022 Jul 26; 12(8): 987.

32.

Jin

Chen

. EEG-based Emotion Recognition Using Graph Convolutional Network with Learnable Electrode Relations. In 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE. 2021 Nov 1. pp. 5953-5957.

33.

Zhu

Yuan

Zhu

. Interpretable learning based dynamic graph convolutional networks for alzheimer’s disease analysis. Information Fusion. 2022 Jan 1; 77: 53-61.

34.

Parisot

Ktena

Ferrante

Lee

Guerrero

Glocker

Rueckert

. Disease prediction using graph convolutional networks: Application to autism spectrum disorder and Alzheimer’s disease. Medical Image Analysis. 2018 Aug 1; 48: 117-30.

35.

Wang

Barstein

Ethridge

Mosconi

Takarae

Sweeney

. Resting state EEG abnormalities in autism spectrum disorders. Journal of Neurodevelopmental Disorders. 2013 Dec; 5: 1-4.

36.

Just

Cherkassky

Keller

Minshew

. Cortical activation and synchronization during sentence comprehension in high-functioning autism: Evidence of underconnectivity. Brain. 2004 Aug 1; 127(8): 1811-21.

37.

Travers

Adluru

Ennis

Tromp

Destiche

Doran

Bigler

Lange

Lainhart

Alexander

. Diffusion tensor imaging in autism spectrum disorder: A review. Autism Research. 2012 Oct; 5(5): 289-313.

38.

Wilson

Rojas

Reite

Teale

Rogers

. Children and adolescents with autism exhibit reduced MEG steady-state gamma responses. Biological Psychiatry. 2007 Aug 1; 62(3): 192-7.

39.

Schopler

Reichler

Renner

. The childhood autism rating scale (CARS). Los Angeles: WPS. 2010.

40.

Patra

Arun

. Use of Indian scale for assessment of autism in child guidance clinic: An experience. Indian Journal of Psychological Medicine. 2011 Jul; 33(2): 217-9.

41.

Bharath

. Developmental Screening Test (DST). Swayamsiddha. Mysore, Karnataka, India. 1998.

42.

Malin

Raj

. Vineland social maturity scale and manual, Indian adaptation. Nagpur: SWAYAMSIDDHA-Prakashana. Saraswathipuram, Mysore, Karnataka, India. 1992.

43.

Delorme

Makeig

. EEGLAB: An open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. Journal of Neuroscience Methods. 2004 Mar 15; 134(1): 9-21.

44.

Luu

Ferree

. Determination of the HydroCel Geodesic Sensor Nets’ average electrode positions and their 10–10 international equivalents. Inc, Technical Note. 2005 Oct; 1-1.

45.

Ning

Bronzino

. Autoregressive and bispectral analysis techniques: EEG applications. IEEE Engineering in Medicine and Biology Magazine. 1990 Mar; 9(1): 47-50.

46.

Chen

. FFT-based deep feature learning method for EEG classification. Biomedical Signal Processing and Control. 2021 Apr 1; 66: 102492.

47.

Šverko

Vrankić

Vlahini

Rogelj

. Complex Pearson correlation coefficient for EEG connectivity analysis. Sensors. 2022 Feb 14; 22(4): 1477.

48.

Zhong

Wang

Miao

. EEG-based emotion recognition using regularized graph neural networks. IEEE Transactions on Affective Computing. 2020 May 11; 13(3): 1290-301.

49.

Erickson

Kitamura

. Magician’scorner: 9. Performance metrics for machine learning models. Radiology: Artificial Intelligence. 2021 May; 3(3).

50.

Avdakovic

Omerhodzic

Badnjevic

Boskovic

. Diagnosis of epilepsy from EEG signals using global wavelet power spectrum. In 6th European Conference of the International Federation for Medical and Biological Engineering: MBEC 2014, 7–11 September 2014, Dubrovnik, Croatia 2015. pp. 481-484. Springer International Publishing.

51.

Šeho

Šutković

Tabak

Tahirović

Smajović

Bečić

Deumić

Bećirović

Pokvić

Badnjević

. Using artificial intelligence in diagnostics of meningitis. IFAC-PapersOnLine. 2022 Jan 1; 55(4): 56-61.

52.

Thabtah

Peebles

. Early autism screening: A comprehensive review. International Journal of Environmental Research and Public Health. 2019 Sep; 16(18): 3502.

53.

Levy

Wolfe

Coury

Duby

Farmer

Schor

Van Cleave

Warren

. Screening tools for autism spectrum disorder in primary care: A systematic evidence review. Pediatrics. 2020 Apr; 145(Supplement_1): S47-59.

54.

Pialoux

Goyard

Lesourd

. Screening tools for frailty in primary health care: A systematic review. Geriatrics & Gerontology International. 2012 Apr; 12(2): 189-97.

55.

Alonso-Esteban

Alcantud-Marín

. Screening, diagnosis and early intervention in autism spectrum disorders. Children. 2022 Jan 25; 9(2): 153.

56.

Rafiee

Rezvani Habibabadi

Motaghi

Yousem

. Brain mri in autism spectrum disorder: Narrative review and recent advances. Journal of Magnetic Resonance Imaging. 2022 Jun; 55(6): 1613-24.

57.

Moridian

Ghassemi

Jafari

Salloum-Asfar

Sadeghi

Khodatars

Shoeibi

Khosravi

Ling

Subasi

Abdulla

. Automatic autism spectrum disorder detection using artificial intelligence methods with MRI neuroimaging: A review. arXiv preprint arXiv:2206.11233. 2022 Jun 20.

58.

Valizadeh

Moassefi

Nakhostin-Ansari

Oskoie

Some’eh

Aghajani

Torbati

Ghorbani

Aghajani

Asl

Mirzamohammadi

. Accuracy of machine learning algorithms for the diagnosis of autism spectrum disorder based on cerebral sMRI, rs-fMRI, and EEG: protocols for three systematic reviews and meta-analyse. medRxiv. 2021 Jan 1.

59.

Conti

Scaffei

Bosetti

Marchi

Costanzo

Dell’Oste

Mazziotti

Dell’Osso

Carmassi

Muratori

Baroncelli

. Looking for “fNIRS Signature” in Autism Spectrum: A Systematic Review Starting From Preschoolers. Frontiers in Neuroscience. 2022; 16.

60.

Clairmont

Wang

Tariq

Sherman

Zhao

Kong

. The value of brain imaging and electrophysiological testing for early screening of autism spectrum disorder: A systematic review. Frontiers in Neuroscience. 2022 Feb 3; 15: 812946.

61.

Amaral

Schumann

Nordahl

. Neuroanatomy of autism. Trends in Neurosciences. 2008 Mar 1; 31(3): 137-45.

62.

Piven

Bailey

Ranson

Arndt

. An MRI study of the corpus callosum in autism. American Journal of Psychiatry. 1997 Aug 8; 154(3): 1051-6.

63.

Hardan

Muddasani

Vemulapalli

Keshavan

Minshew

. An MRI study of increased cortical thickness in autism. American Journal of Psychiatry. 2006 Jul; 163(7): 1290-2.

64.

Chen

Jiao

Herskovits

. Structural MRI in autism spectrum disorder. Pediatric Research. 2011 May; 69(8): 63-8.

65.

Luyster

Gotham

Guthrie

Coffing

Petrak

Pierce

Bishop

Esler

Hus

Oti

Richler

. The Autism Diagnostic Observation Schedule – Toddler Module: A new module of a standardized diagnostic measure for autism spectrum disorders. Journal of Autism and Developmental Disorders. 2009 Sep; 39(9): 1305-20.

66.

Rutter

Le Couteur

Lord

. Autism diagnostic interview-revised. Los Angeles, CA: Western Psychological Services. 2003; 29(2003): 30.

67.

Vahia

. Diagnostic and statistical manual of mental disorders 5: A quick glance. Indian Journal of Psychiatry. 2013 Jul 1; 55(3): 220-3.

68.

Hosseinzadeh

Koohpayehzadeh

Bali

Rad

Souri

Mazaherinezhad

Rezapour

Bohlouli

. A review on diagnostic autism spectrum disorder approaches based on the Internet of Things and Machine Learning. The Journal of Supercomputing. 2021 Mar; 77(3): 2590-608.

69.

Bajestani

Behrooz

Khani

Nouri-Baygi

Mollaei

. Diagnosis of autism spectrum disorder based on complex network features. Computer Methods and Programs in Biomedicine. 2019 Aug 1; 177: 277-83.

70.

Abdolzadegan

Moattar

Ghoshuni

. A robust method for early diagnosis of autism spectrum disorder from EEG signals based on feature selection and DBSCAN method. Biocybernetics and Biomedical Engineering. 2020 Jan 1; 40(1): 482-93.

71.

Wadhera

Kakkar

. Social cognition and functional brain network in autism spectrum disorder: Insights from EEG graph-theoretic measures. Biomedical Signal Processing and Control. 2021 May 1; 67: 102556.

72.

Garcés

Baumeister

Mason

Chatham

Holiga

Dukart

Jones

Banaschewski

Baron-Cohen

Bölte

Buitelaar

. Resting state EEG power spectrum and functional connectivity in autism: A cross-sectional analysis. Molecular Autism. 2022 Dec; 13(1): 1-6.

73.

Weiland

Konvalinka

Mansvelder

Andersen

Smit

Begeer

Linkenkaer-Hansen

. Intellectually able adults with autism spectrum disorder show typical resting-state EEG activity. Scientific Reports. 2022 Nov 8; 12(1): 19016.

74.

Vieira

Pamplona

Fachinello

Silva

Foss

Salmon

. On the prediction of human intelligence from neuroimaging: A systematic review of methods and reporting. Intelligence. 2022 Jul 1; 93: 101654.

75.

Cao

Yang

Qin

Zhu

Chen

Wang

Liu

. Using DeepGCN to identify the autism spectrum disorder from multi-site resting-state data. Biomedical Signal Processing and Control. 2021 Sep 1; 70: 103015.

76.

Wen

Cao

Bao

Yang

Zheng

Zaiane

. MVS-GCN: A prior brain structure learning-guided multi-view graph convolution network for autism spectrum disorder diagnosis. Computers in Biology and Medicine. 2022 Mar 1; 142: 105239.

77.

Wang

Liu

Xiang

Wang

Chen

Chong

. MAGE: Automatic diagnosis of autism spectrum disorders using multi-atlas graph convolutional networks and ensemble learning. Neurocomputing. 2022 Jan 16; 469: 346-53.

78.

Huang

Chung

. Disease prediction with edge-variational graph convolutional networks. Medical Image Analysis. 2022 Apr 1; 77: 102375.

79.

Wen

Cao

Liu

, R. Zaiane O Yang

. Exploring interpretable graph convolutional networks for autism spectrum disorder diagnosis. International Journal of Computer Assisted Radiology and Surgery. 2022 Nov 4: 1-1.

80.

Hashem

Nisar

Bhat

Yadav

Azeem

Bagga

Fakhro

Reddy

Frenneaux

Haris

. Genetics of structural and functional brain changes in autism spectrum disorder. Translational Psychiatry. 2020 Jul 13; 10(1): 1-7.

81.

Song

Topriceanu

Ilie-Ablachim

Kinali

Bisdas

. Machine learning with neuroimaging data to identify autism spectrum disorder: A systematic review and meta-analysis. Neuroradiology. 2021 Dec; 63(12): 2057-72.

82.

Ouyang

Kang

Detre

Roberts

Huang

. Short-range connections in the developmental connectome during typical and atypical brain maturation. Neuroscience & Biobehavioral Reviews. 2017 Dec 1; 83: 109-22.

83.

Azeem

Palomba

Shi

Wang

. Machine learning techniques for code smell detection: A systematic literature review and meta-analysis. Inf Softw Technol. 2019; 108: 115-138.

84.

Just

Cherkassky

Keller

Minshew

. Cortical activation and synchronization during sentence comprehension in high-functioning autism: Evidence of underconnectivity. Brain. 2004 Aug 1; 127(8): 1811-21.

Brain-region specific autism prediction from electroencephalogram signals using graph convolution neural network

Abstract

BACKGROUND:

OBJECTIVE:

METHODS:

RESULTS:

CONCLUSION:

Keywords

1. Introduction

Table 1 Primary brain regions and functionality with respect to EEGs

Table 2 Descriptive statistics of the EEG data collected from autistic children and healthy subjects ( N = 16, Autistic = 8, healthy = 8)

Step 1: Removal of channel 257

Step 2: Downsampling

Step 3: Noise removal

Step 4: Data decomposition

Step 5: Mapping of 257 HydroCel to 10-10 international standard

Table 4 Selected channels of its 10-10 international equivalent for 10-20 HydroCel GSN

Step 6: Division into brain regions

2.4.2 Adjacency matrix

3. Results

Table 5 Hyperparameters used in the proposed GCNN model

Table 7 Comparison between the current work with previous works

Funding

Ethical approval

Informed consent

Data availability

Footnotes

Acknowledgments

Conflict of interest

References

Table 1
Primary brain regions and functionality with respect to EEGs

Table 2
Descriptive statistics of the EEG data collected from autistic children and healthy subjects ( $N=$ 16, Autistic $=$ 8, healthy $=$ 8)

Table 4
Selected channels of its 10-10 international equivalent for 10-20 HydroCel GSN

Table 5
Hyperparameters used in the proposed GCNN model

Table 7
Comparison between the current work with previous works