Comprehensive Autism Spectrum Disorder Analysis: ML and DL Models in Multimodal Datasets

2.1 ASD signs and indicators

People with ASD may struggle with having enjoyable distractions, dullness, as well as social situations. They may also struggle to focus in class or obey instructions. These characteristics could make daily tasks challenging. It is significant to remember that some people may display comparable symptoms even in the absence of an autism spectrum condition.

Interaction and Social Communication Skills:

• Does not recognize their name by the age of nine months.

• Refuses to make eye contact.

• Does not exhibit emotional expressions (happy, sad, enraged, etc.) by the age of nine months.

• By the time they are 12 months old, they seldom ever make gestures (like waving goodbye).

• By the time they are 15 months old, do not play well with others (for instance, display to you something they find enjoyable).

• Shows no signs of emotional sensitivity at 24 months old.

• By 36 months of age, neither initiates play nor engages in interactions with other kids.

• Cannot complete the task in 60 months.

Limited or Continual Interests or Behaviors: People across the spectrum could exhibit odd hobbies or behaviours. These differentiate ASD from conditions characterized solely by problems through social interaction and communication as shown in Figure 1. The following are some examples of pastimes and pursuits that may be limited or repetitive as a result of the autism spectrum condition:

• Overreacts to little variations in the way his or her toys or other belongings are arranged.

• Concentrates on specific portions of objects (wheels, for example).

• Uses the same words or phrases repeatedly (Echolalia).

• Animated by little changes.

• Has peculiar sensory responses to sounds, tastes, scents, and touches in their environment.

1111.png

Figure 1. ASD symptoms

There are other symptoms as well, such as irregular sleep and eating schedules, digestive issues including delayed language development, delayed motor development, delayed cognitive or learning development, constipation, anxiety, stress, excessive or lack of fear, etc. [7].

2.2 Reasons and potential hazards

Although the precise causes of ASD are unknown, research has demonstrated that a person's development can be significantly influenced by both genetic and environmental factors, which may lead to the disorder. ASD has been connected to several circumstances, including having an older parent, a sibling with the illness, a particular genetic disorder (e.g., Down syndrome, fragile X syndrome), or extremely low birth weight.

2.3 Diagnosis approaches of ASD

Healthcare providers can evaluate a patient's likelihood of having ASD by keeping an eye on their growth and behavior. By the time a child is two years old, an accurate diagnosis of ASD can usually be made. It's critical to receive a diagnosis as soon as possible [8].

Diagnoses in Early Childhood: There are two stages in the diagnosis process for young children:

During healthy-child visits, general development screening: A paediatrician or other early childhood health care provider should examine every child. At the ages of nine, eighteen, twenty-four, or thirty months, all children should have a developmental delay screening. Additional screening may be done on a child if they are highly susceptible to developmental problems or ASD. Children with aging parents, ASD-related behavior, an ASD family member, Low birth weight, and genetic disorders are regarded as high-risk factors [9].

The physician may combine inquiries regarding the child's conduct with the results of an ASD assessment, and clinical findings of the child to assess the child.

Further Diagnostic Assessment: Children with ASD must receive an early and precise diagnosis because it will highlight their unique skills and challenges. Using this information, parents can then choose the educational and behavioral therapy programs and services that will best support their child's success.

The diagnostic evaluation will most likely consist of neurological and medical testing, a cognitive skills assessment, a language skills assessment, observation of the child's behavior, and other methods.

Diagnosis of adolescents and older children: Parents and teachers of school-age children and adolescents frequently notice the first indications of ASD. Following an initial evaluation, the student may be recommended for further testing by the department of special education at the school, their regular physician, or an expert in ASD.

Using subtleties in language can make it challenging to read silent signs like facial expressions, body language, and voice tone. Witter, humour, and metaphor may be difficult for older children and teenagers. They might also have trouble connecting with kids their age.

Adult diagnosis: Adults with ASD have far more diagnostic challenges than children do. Among the indications and manifestations in adults is ADHD, ASD may also be present in people with anxiety disorders or ADHD.

Adults who exhibit symptoms that align with ASD should consult a doctor to receive a diagnosis. As part of the examination, other family members or caregivers could be questioned to find out additional details regarding the person's early stages of development. This would subsequently enable a precise diagnosis.

Therapeutic measures: Treatment for ASD should start as soon as it is practical after diagnosis. It's critical to start ASD treatment early due to appropriate resources and assistance, people may decrease their difficulties and learn new skills while making the most of their strengths. Working closely with a healthcare provider can help to find the ideal mixture of services and treatments.

Pharmaceutical: A Pharmaceutical may be prescribed by a doctor to address particular symptoms. An individual with ASD may experience a reduction in various issues such as irritability, aggression, hyperactivity, repetitive behavior, anxiety, and sadness when taking medication [10].

Interventions in behavior, psychology, and education for a child with ASD may benefit from the recommendation of a medical professional who focuses on offering therapeutic services that are behavioral, psychological, educational, or skill-building. Siblings and other family members are often involved in these programs as well as caregivers. These programs may be helpful for individuals with ASD who wish to: Create social, language, and interpersonal relations and build on their advantages.

2.4 Both DL and ML

ML: ML is a sub-field of AI that mimics human decision-making in machines, through the use of intricate computational algorithms, which seek to "train" machines to analyze big datasets in a fast, accurate, and effective manner. ML is typically divided into three major categories [11]. The subject of this paper is ML algorithms, both supervised and unsupervised.

Supervised learning: Another name for this is predictive learning. In this instance, a labelled dataset is used to train an ML algorithm containing the required inputs as well as outputs. During training, the algorithm picks up notes that will aid in its prediction-making.

Unsupervised learning: Unsupervised learning algorithms aren't trained on labelled datasets. Without human assistance, they must extract insights and hidden patterns from the data that has been provided.

Reinforcement learning: Another name for reinforcement learning is reward-based learning. Here, the algorithm picks up new skills by interacting with the outside world, acting in certain ways, and observing the outcomes. Every right move will earn the algorithm a reward, and every wrong move will cost it a penalty [12].

DL: Natural language processing relies on Artificial Neural Networks (ANNs). DL techniques outperform ML algorithms and data analysis techniques in numerous situations. By including more neurons and layers, DL evolved from the application of neural networks. The middle layers take in data, process it through various functions, and then output the results to the layers below. The middle layers are referred to as hidden layers, while the first and last layers are called input and output layers, respectively. The layers in the DL architecture include input, convolutional, activation, fully connected, sequence, normalization, drop-out, pooling, combination, and output layers [13].

Metrics for evaluating the performance of models: Numerous metrics are available for evaluating the model's performance. It is important to select metrics carefully when evaluating the model's performance because:

• All model comparisons and measurements are determined by metrics.

• Ultimately, the selected metric will greatly influence the relative importance of various attributes.

Confusion matrix, recall, F1-score, AUC, specificity, accuracy, and precision were the most often utilized metrics for assessing model performance [14].

2.5 Determining data inclusion parameters

This study thoroughly examined how ML techniques aid in the prediction of ASD. We created a keyword that is closely associated with ASD detection by employing the phrases below to widen the search parameters: (Text-based analysis OR Image analysis (Brain MRI OR Facial Expressions OR Gesture and Postures) OR Video analysis) AND (autism identification OR intervention OR detection OR diagnosis classification) AND (ML OR AI OR DL). We have compiled data from articles published by IEEE, Google Scholar, Springer, Scopus, and Elsevier between 2018 to 2023.

As part of this survey, we examine the effectiveness of various ML and DL algorithms for diagnosing ASD disease by examining prior research conducted by researchers. An ML model is assessed by employing multiple metrics, including sensitivity, specificity, accuracy, and f1-score. The dataset and the ML model influenced the model's overall performance, regardless of the metric that was used.

2.6 Taxonomy

This research has shown that the current methods for detecting ASD have been thoroughly studied.

Our main contributions to this survey:

• A thorough analysis of the ML and DL models utilized in ASD is listed in this survey.

• In this survey, the main performance evaluation metrics for ASD are examined, and the effectiveness of well-known ASD algorithms in DL and ML models is examined.

• The ASD detection techniques used in recent years are detailed in this survey.

• This is the first comprehensive analysis that we are aware of ML techniques for ASD.

This breakdown represents the distribution of various methods or data types used in autism-related research as shown in Figure 2. Brain MRI comprises the largest percentage, representing approximately 49% of the methodologies utilized. Following that, text-based analysis makes up around 21%. Facial expression data accounts for roughly 19% of the methodologies. Gesture and posture recognition constitute a smaller portion, approximately 2%, while video analysis represents about 9%.

2.jpg

Figure 2. Analysis of ASD dataset

3.jpg

Figure 3. Taxonomy of this review

The percentages illustrate the relative prominence or utilization of different data types or techniques in studies related to autism. Brain MRI holds the most significant share among the listed methods, while Gesture and Posture Recognition have the smallest share in the overall analysis of autism-related information or research.

It encompasses a comprehensive framework categorizing crucial elements into four main divisions. As shown in Figure 3 firstly, it partitions datasets into three primary categories: text, image, and video data. Under image data, further outlines involve brain MRI, facial expressions, and gesture/posture data. Secondly, the taxonomy includes the source of data divided into two groups: vision-based (mobile, web, and device cameras) and source-based (motion sensors and RF sensors). Thirdly, it outlines feature extraction, categorizing it into specific behavioral traits and brain activity patterns. Lastly, the taxonomy encompasses learning models, classifying them into ML models (SVM, RF, KNN, LR, RL, clustering analysis) and DL models (CNN, RNN, LSTM). This taxonomy provides a structured and comprehensive framework for understanding the diverse components involved in autism detection methodologies, facilitating a systematic analysis of the field's research landscape.

4.1 The role of ML and DL in brain MRI

Numerous studies suggest that children and adolescents diagnosed with autism tend to exhibit an enlarged hippocampus, the brain region accountable for memory formation and storage. However, it remains uncertain whether this variation persists as individuals transition into adolescence and adulthood [25, 26].

The dimensions of the amygdala appear to vary between individuals with and without autism. However, research conducted by various laboratories has produced conflicting findings on this matter. According to researchers, Individuals diagnosed with autism may exhibit a smaller amygdala compared to those without autism. It's also suggested that their amygdala might be smaller specifically when they also experience anxiety [27]. According to research, as shown in Figure 4, children with autism exhibit early-stage amygdala enlargement with the difference gradually decreasing [28].

A review of imaging studies revealed that the cerebellum, the brain structure located at the base of the skull, has less brain tissue in autistic individuals [29]. Researchers know that the cerebellum plays a role in both social interaction and cognition, in addition to its long-held belief that it primarily controls movement. People with and without autism have different patterns of thickness in the outer layer of the brain. This variation corresponds to changes in a single neuronal sub-type during development.

Raja and Kannimuthu [30] proposed the context of ASD diagnosis and neuroimaging methods reveals a predominant reliance on fMRI techniques as shown in Table 2, often with limited dataset sizes, yielding high accuracy but limited generalizability. While traditional supervised ML algorithms like SVM show promise in handling unstructured data, they are hindered by limited training data. In contrast, DL approaches, particularly Generative Adversarial Networks (GANs), are gaining traction due to their capacity to automatically learn patterns and generate new data examples. However, their application in ASD prediction is relatively unexplored. This paper introduces a new conditional generative adversarial network (cGAN) for ASD prediction, demonstrating superior accuracy compared to traditional methods, with a 74% improvement, and remarkable efficiency, requiring only around 10 minutes for training even on large datasets. This paper underscores the potential of DL, specifically cGANs, in ASD diagnosis and rehabilitation.

Sabegh et al. [31] studied the realm of diagnosing diverse mental illnesses, particularly ASD, and the existing reliance on behavioral symptom observation presents challenges, especially in children. This study seeks to enhance diagnostic accuracy through the integration of sophisticated and expandable ML techniques, specifically DL networks. They are leveraging resting-state fMRI information from the ABIDE1 repository encompassing seventeen different imaging locations and an extensive preprocessing pipeline is implemented, involving registration of data on atlases, extraction of brain region average time series, and computation of correlation matrices. The method of selecting chi-square features is then employed to identify the most critical features within these matrices. An innovative CNN architecture, featuring Convolutional layers with two dimensions is proposed for data analysis and classification. This model's evaluation encompasses three distinct experiments based on different datasets and atlases. Remarkably, the highest accuracy achieved in these experiments stands at 73.53%, surpassing previous attempts at categorizing ASD against standard controls. This research introduces a pioneering CNN-based model capable of automatic ASD classification, demonstrating strong classification performance. Unlike prior studies that focused on limited data sites, this work capitalizes on data from all sites, thereby enhancing the generalizability and potential clinical utility of the proposed model.

4.png

Figure 4. Parts of the brain affected by autism

Ulaganathan et al. [32] present a novel approach for ASD classification using deep RL methods, namely deep Q learning network (DQN) and Spinal Net, with the training of hyperparameters facilitated by a unique training optimizer, the Driving Training Political Optimizer (DTPO). The process commences with image acquisition from a dataset, followed by adaptive Wiener filtering and ROI extraction. Subsequently, the core region is identified using the DTPO. The classification hinges on these extracted features. Furthermore, the study's classification performance is evaluated, and a fusion process between DQN and Spinal Net, aided by Czarnowski similarity, results in ASD classification. The assessment encompasses various metrics, including impressive values of 0.907, 0.958, 0.936, 0.536, 0.732, 0.488, and 0.409, respectively, for accuracy, sensitivity, specificity, mean-squared error, root-mean-squared error, R, and mean absolute error. This research highlights the potential of deep RL networks in effectively categorizing ASD, offering a possible path toward improvement in early diagnosis and intervention.

Lei et al. [33] studied that the accurate diagnosis of ASD is essential to ensure optimal prognosis and treatment, and Functional Brain Networks (FBNs) built from fMRI data have gained popularity as a diagnostic tool. However, existing model-driven approaches for FBN construction often fail to record any non-linear connections that might exist between the diagnostic labels and the data. Furthermore, the conventional separation of FBN construction and disease classification in most studies leads to substantial inter-subject variability in the calculated FBNs and lowers the next group's statistical power comparisons. To get beyond these restrictions, this study introduces a novel approach known as the deep unrolling-based spatial constraint representation (DUSCR) method. It combines it with a convolutional classifier to produce an end-to-end architecture for the identification of ASD. This method utilizes a proximal gradient descent algorithm to solve the Spatial Constraint Representation (SCR) and unrolls using the deep unrolling algorithm, it is incorporated into deep networks. Putting things into categories is achieved through a convolutional prototype learning model. The method's efficiency was assessed on the ABIDE I dataset, demonstrating an important enhancement in classification accuracy and model performance, thus offering a promising solution for improving ASD diagnosis and understanding the complex relationships within functional brain networks.

RethikumariAmma and Ranjana [6] aims to facilitate early ASD diagnosis by leveraging neuroimaging functional images. This method employs a DL approach to discern whether individuals with ASD, obtain strong characteristics from these neuroimaging data. It evaluates the effectiveness of images with prior processing in categorizing neuronal patterns, with the ultimate goal of enabling early diagnosis for long-term health management. Functional connectivity analysis is used to identify crucial brain areas using the box neighbourhood search algorithm as its foundation. To diagnose ASD, a Deep Neuro-Fuzzy Network (DNFN) is utilized, and Feedback-Henry Gas Optimization (FHGO) is used to train the DNFN. The results of this FHGO-DNFN approach demonstrate exceptional performance, achieving a high accuracy rate of 93.3%, along with a sensitivity of 94.7% and specificity of 91.4%. This promising methodology can potentially enhance early ASD detection, thereby advancing clinical practices and improving the long-term well-being of individuals with ASD.

Sherkatghanad et al. [34] focus on automating the detection of ASD by harnessing the power of CNNs and a brain imaging dataset. We achieved this by identifying ASD patients within the ABIDE, a multi-site dataset that includes fMRI resting-state data. Our approach leveraged operational interconnectedness patterns to group individuals with ASD and control subjects effectively. Our Results from experiments show that our model accomplishes a 70.22% accuracy rate utilizing the brain's CC400 functional parcellation atlas and the ABIDE I dataset. Notably, our CNN model is computationally more efficient than state-of-the-art methods because it uses fewer parameters. This model, now poised for further testing with larger datasets, holds the capacity to act as a valuable tool regarding the preliminary screening of people with ASD.

Nogay and Adeli [35] introduced an automatic diagnostic structural MRI-based model for ASD. The model comprises two essential stages: the initial preprocessing stage, and the subsequent diagnostic phase. The preprocessing stage entails the removal of unclear images, followed by the application of the Canny Edge Detection (CED) algorithm to delineate image edges, cropping the images to the required dimensions, and augmenting the dataset by upscaling the images while preserving essential image characteristics. Notably, data augmentation is applied meticulously to avoid any bias introduced into the data, ensuring its integrity for both ASD and TD groups. Phase two involves the application of the grid search optimization (GSO) algorithm to fine-tune the deep CNN by optimizing hyperparameters. Remarkably, this approach achieves a diagnostic success rate of 100% for ASD, as evidenced by rigorous five-fold cross-validation. The model's robustness is further validated by comparisons with current research and extensively utilized pre-trained models, establishing its superiority in ASD diagnosis based on MRI. This groundbreaking research offers a promising method for accurate and reliable automated ASD diagnosis, which is based on a substantial leap forward within the domain of neuroimaging and diagnosis of ASD.

Bhandage et al. [36] studied that ASD diagnosis is achieved through a novel approach based on the Adam War Strategy Optimization (AWSO) Deep Belief Network (DBN). The algorithm used by AWSO is developed by integrating the War Strategy Optimization (WAO) and Adam optimizer, presenting a simpler yet highly effective approach that outperforms existing methods. The preprocessing stage involves extraction of the Region of Interest (ROI) and anisotropic diffusion to eliminate image noise. Furthermore, the Algorithm for Box Neighbourhood Search that Uses Functional Connectivity is employed to extract crucial brain regions, enhancing ASD classification performance. The classification itself is carried out through the DBN, with the algorithm used by AWSO optimizing the DBN's learning process. The AWSO-DBN analysis was conducted on ABIDE-I as well as ABIDE-II datasets, with the AWSO-DBN demonstrating exceptional work, particularly using the dataset ABIDE-I, where it attained a high degree of sensitivity (0.930), accuracy (0.924), and specificity (0.935). These results showcase the promise of the AWSO-DBN algorithm as a powerful instrument for accurate ASD classification, underscoring its potential influence on the field of diagnosing ASD.

Parui et al. [37] studied a novel approach for constructing functional connectivity networks using rs-fMRI data. By extracting time series data from fMRI and calculating correlation matrices that represent the interactions among various brain regions, the study leverages different brain atlases. The proposed technique incorporates the idea of majority voting based on atlas-specific outcomes, resulting in a commendable ASD detection accuracy of 84.79%. This approach underscores the potential of AI-driven neuroimaging techniques in advancing our understanding of ASD and enhancing diagnostic accuracy, thus supporting early intervention and assistance for people with ASD.

Yin et al. [38] suggested using a semi-supervised autoencoder (AE) approach introduced for the diagnosis of autism, utilizing functional connectivity (FC) patterns derived via fMRI data in the resting state. By combining unsupervised AE training utilizing networks for supervised classification, the proposed semi-supervised learning framework facilitates simultaneous training of an autoencoder and a neural network-based classifier. This approach, in contrast to training these components separately, allows the latent feature representation to be tailored to the goal of classification, resulting in enhanced diagnostic performance for autism. On the ABIDE I database, cross-validation is used to evaluate the model, demonstrating improved classification performance. The results highlight the utility of the suggested framework for semi-supervised learning for integrating unlabelled fMRI information, leading to improved classification accuracy and feature learning in the context of ASD diagnosis.

Ahammed et al. [39] explore the application of ML for ASD identification, focusing on overcoming limitations associated with using fMRI and large datasets. The proposed innovative architecture that uses the Bag-of-Features model as a basis, presents a comprehensive approach to ASD classification. This approach involves preprocessing images, extracting speeded-up robust features (SURF), creating a visual vocabulary through K-Means clustering, and encoding Bag-of-Features using quantization and coding methods. The study highlights the preference for the SVM as the ASD classifier. The evaluation includes three datasets, encompassing ABIDE fMRI pre-processed images and the subject's face images. The experimental results indicate that SVM performs much better when using the Bag-of-Feature approach, which achieves the highest accuracy of 81% and specificity of 86%. This suggests that ML classifiers using extractors for bags of features have the potential to reinforce ASD diagnosis, particularly in clinical and medical contexts, in contrast to alternative cutting-edge techniques.

Itani and Thanou [40] discuss ongoing research in ASD and the application of network science and contemporary ML to gain an improved comprehension of the neuropathology of ASD and the emergence of diagnostic aids. This research specifically focuses on classifying individuals with ASD and neurotypical incorporating an understanding of the structure and function of the brain. The brain is represented as a graph, with rs-fMRI signals mapped to the graph's nodes. Graph Signal Processing (GSP) tools are applied to analyse the signals' frequency composition, creating discriminative features by expanding the Fukunaga-Koontz transformation. These features are used to instruct a decision tree for categorization, resulting in an interpretable diagnostic model that outperforms existing methods, as demonstrated in the ABIDE gathering. Predictive marker analysis highlights the role of the temporal and frontal lobes in the diagnosis of ASD, aligning with prior discoveries in the literature on neuroscience. This research emphasizes the importance of considering structural and functional brain information for gaining insights into the complexity of ASD neuropathology. Additionally, the study presents test accuracies achieved by various methods with standard deviation intervals for different training set sizes.

Zhan et al. [41] studied the field of psychiatric disorders, particularly ASD, Obsessive-Compulsive Disorder (OCD), and ADHD, there exists an ongoing debate regarding precise diagnosis as well as the potential convergence of their anatomical foundations. A recent study employed non-invasive neuroimaging techniques in both human subjects and primates that are not humans to investigate neural markers linked to DSM-5 diagnoses and quantitative indicators of the severity of symptoms. Diagnostic classifiers were built using resting-state functional connectivity data from both wild-type and methyl-CpG binding protein 2 (MECP2) transgenic monkeys, which were then applied to four human datasets (OCD local institutional database: N=186; ADHD-200 sample: N=776; ABIDE-I: N=1,112; ABIDE-II: N=1,114). This innovative approach revealed nine key brain regions that are primarily found in the frontal and temporal cortices, which were instrumental in informing diagnostic classifications for ASD and OCD, although not for ADHD. Notably, models based on the functional connections of specific brain regions predicted symptom severity scores in individuals with ASD and OCD. These results hold significant promise regarding the development of diagnostic indicators for OCD and ASD, as well as for gauging the severity of symptoms. Moreover, they suggest potential avenues for enhancing the precision and effectiveness of clinical evaluations in the field of psychiatric disorders by leveraging machine-learning models.

Liu et al. [42] studied the context of ASD, characterized by social deficits and repetitive behaviours, the absence of reliable biomarkers has been a longstanding challenge. Numerous efforts have been dedicated to identifying biomarkers using resting-state fMRI. Nevertheless, a substantial amount of information is lost in feature selection because these studies frequently ignore strong group relationships. To resolve this problem, a novel approach is proposed for ASD diagnosis, utilizing the elastic net method utilizing rs-fMRI data. The key advantage of the method of elastic networks is its capacity to bypass the need for upfront feature selection, saving time and enhancing algorithm efficiency. Experimental outcomes conducted on the open database ABIDE affirm the efficacy as well as the utility of the suggested technique, marking a promising step toward addressing the biomarker challenge in ASD diagnosis.

Mostafa et al. [43] tackled the challenging problem of diagnosing ASD by introducing a novel approach that relies on brain network-based features extracted from functional MRI data. By defining 264 raw brain features using network eigenvalues and centralities and employing feature selection techniques, they reduced the dimensionality to 64 discriminative features. Training ML models, particularly Linear Discriminant Analysis, utilizing the ABIDE dataset, achieved a noteworthy classification precision of 77.7%. This outperformed existing methods and shows promise for improving based on ASD diagnosis neuroimaging information, addressing the critical need for more effective diagnostic tools in the field. Further validation and replication studies are essential to confirm the robustness of these findings.

Wang et al. [44] focus on addressing the challenge of identifying biomarkers for accurate diagnosis of ASD using fMRI information from several sites. Given the variability across sites in multi-site data, the research introduces a multi-site adaption framework based on low-rank representation decomposition (maLRR). This framework aims to create a standard low-rank representation for information from various websites, thus reducing variations in data distributions. The low-rank representation method adapts the data to a shared space by considering one site as the target domain and the others as source domains. This adaptation process minimizes the differences between the source and target domains. The results from evaluations on multi-site fMRI data that is both real and artificial suggest that the proposed technique outperforms numerous cutting-edge domain adaptation techniques, offering promise for more accurate ASD identification and early intervention.

Fredo et al. [45] aimed to distinguish between TD individuals and those with ASD by analyzing fMRI. They used data from 320 participants for training and 80 for validation, sourced from the ABIDE-I, and ABIDE-II. The fMRI images were pre-processed through a standard pipeline, and Functional Connectivity (FC) matrices were computed 237 regions of interest (ROIs) were used in the cortex, subcortex, and cerebellum. To reduce the FC matrix's dimensionality, they employed RF with conditions and assessed classification accuracy by employing random forests at every dimension. The findings showed that, within this dataset, RF achieved a maximum accuracy of 65% with 143 characteristics. The region that contributed the greatest features was the Cingulo-Opercular Task Control (COTC) region. associated with precise categorization and communication between the dorsal attention network and COTC played a crucial role in distinguishing ASD and TD participants. This research sheds light on the potential of fMRI-based connectivity patterns in aiding the classification of ASD.

Heinsfeld et al. [46] studied DL techniques to recognize individuals with ASD from a substantial brain imaging dataset obtained from the ABIDE database. The neurodevelopmental disorder known as ASD is typified by repetitive behaviours and social deficits, with a significant impact on child populations. The research focused on the brain's functional connectivity patterns to objectively distinguish ASD participants from typically developing controls and to gain insights into the neural patterns underlying this classification. The results represented an advancement in the field by achieving that 70% of ASD patients in the dataset could be identified accurately. Notably, the findings revealed a difference in brain activity between the anterior and posterior regions of the brain, which aligns with existing empirical evidence of disruptions in anterior-posterior brain connectivity in ASD. The study further identified the specific brain areas that played a crucial role in distinguishing ASD from customarily creating controls through their DL methods, shedding light on the neural mechanisms associated with ASD.

Vandewouw et al. [47] focus has recently shifted towards identifying subgroups among Children and teenagers suffering from neurodevelopmental disorders by leveraging measures of brain function, utilizing data from two large and independent networks: The Healthy Brain Network (HBN) and the Province of Ontario Neurodevelopmental Network (POND). Researchers’ studies have explored the potential for data-driven clustering analyses to uncover subgroups characterized by common functional brain features, independent of traditional diagnostic categories. The results revealed subgroups exhibiting shared biological characteristics, yet often deviating significantly regarding intelligence, hyperactivity, and impulsivity traits. These findings challenge the conventional diagnostic boundaries, suggesting that a more nuanced understanding of neurodevelopmental conditions may emerge through the exploration of brain connectivity patterns, potentially advancing our comprehension of the underlying complexities of these disorders.

Gao et al. [48] proposed the realm of ASD prediction and neuroimaging, existing research predominantly focuses on fMRI methods, which are primarily applied to individuals older than 5 years of age for diagnosis. This approach faces limitations when dealing with infants due to the challenges of fMRI applications. Consequently, a growing emphasis is on leveraging structural MRI to enable early ASD prediction, particularly around 24 months of age. This study introduces an automated prediction framework, employing an infant-specific pipeline, I BEAT V2.0 Cloud, to generate Partitioning and dividing maps from baby MRI scans. The framework utilizes feature extraction from paired maps using a CNN and Siamese network to discern if the matched individuals exhibit similar or different ASD statuses. The research demonstrates that integrating maps with segmentation and parcellation enhances the ASD prediction's sensitivity, specificity, and overall accuracy as confirmed by two datasets employing distinct imaging procedures and devices through analysis of the receiver operating characteristic. Comparative assessments against cutting-edge techniques emphasize how much more reliable and effective the suggested strategy is. Moreover, the study employs attention maps to pinpoint subject-specific autism effects, validating the predictive outcomes. These collective findings underscore the practicality of the unified system for predicting ASD in its early stages using sMRI, particularly when compared to models trained solely on T1w images which exhibited 76.9% sensitivity, 81.5% specificity, and 80.6% accuracy.

Conti et al. [49], brain structural differences among individuals with ASD, Childhood Apraxia of Speech (CAS), and TD individuals were explored using MRI measures and predictive Machine Learning techniques. The study aimed to identify distinct patterns among these conditions, without providing specific findings and reported an AUC of 73%. The authors suggested that future research should involve a larger sample size to understand better how ASD and comorbid CAS cases differ from pure disorders. Additionally, the study emphasized the importance of utilizing advanced imaging techniques and larger cohorts to validate and refine the observed patterns, facilitating more accurate diagnostic tools for ASD and CAS.

Saad and Islam [50] present an automated method for classifying individuals Using Diffusion Tensor Imaging (DTI) and graph theory-based features, ASD and TD brains were compared as mentioned in Table 2. The DTI is employed to study microstructural changes in the brain by monitoring the flow of water in white matter trails. Connectivity matrices are created from DTI data, and a graph theory-based analysis extracts relevant features. SVM and Principal Component Analysis (PCA) are utilized for classification, and Linear Discriminant Analysis (LDA) is employed to lower noise in the features. This study achieved classification accuracies of 75.00%, 62.50%, and 56.25% with 2, 7, and 10 PCA features using SVM, and 64.58%, 60.42%, and 58.33% using LDA. Notably, using fewer features led to better accuracy, especially with the same test samples. This approach leverages DTI and graph theory to provide a promising method for distinguishing between ASD and TD brains.

Table 2. Parametric Assessment of ASD Detection Methods using Image-Based (Brain MRI) Dataset

Ke et al. [51] address the need for efficient processing of the complex and abundant MRI data used in medical diagnostics. They propose a novel approach that involves applying the use of adaptive independent subspace analysis (AISA) to find meaningful EEG activity in the data from MRI. The outcomes obtained through AISA are then subjected to texture analysis of images, encompassing first-order statistics, Grayscale run-length matrix, Grayscale size-zone matrix, Grayscale co-occurrence matrix, and surrounding Features of a grayscale difference matrix. These features are transformed by employing t-distributed stochastic neighbour embedding (t-SNE) into a 2D space. Subsequently, the authors employ a 10-fold cross-validation, the KNN, achieving a notable 94.7% precision and a high f-score of 0.9356 when classifying real ASD MRI data. The study successfully demonstrates the efficacy of this method in distinguishing between normal and autistic brain MRI slices in the tissue of the brain. upcoming study plans include expanding the analysis to encompass visual moments and morphological elements to further enhance accuracy.

4.1.1 Observation

Based on the above study in ASD detection future research concerning brain MRI datasets for ASD analysis, several pivotal improvements can be explored to enhance the accuracy, reliability, as well as applicability of ML models

Future advancements in brain MRI dataset analysis for ASD should prioritize three crucial areas. Firstly, there should be a concentrated effort on the integration of multiple MRI modalities into a unified framework, combining structural, functional, and diffusion MRI data along with other relevant information like genetics or behavioral assessments. Employing fusion methodologies, particularly DL-based multimodal architectures, can unlock intricate patterns across diverse data sources, potentially improving diagnostic accuracy and unveiling a more comprehensive understanding of ASD-related brain alterations. Secondly, it's imperative to establish model robustness across diverse datasets and populations by validating models on larger, more heterogeneous datasets encompassing different demographics and imaging protocols. Ensuring model generalizability and transferability is critical for their clinical applicability. Lastly, focusing on advancements in interpretability and explainability techniques specific to brain MRI datasets will enable a better understanding of model predictions, enhancing trust and facilitating the translation of ML-based insights into actionable clinical information for ASD diagnosis and treatment decision-making.

4.2 The role of ML and DL in facial expression recognition

Facial expressions play a vital part in social communication and within the context of autism as shown in Table 3. Face identification involves challenges in non-verbal communication, which can extend to facial expressions. Some individuals with autism might have difficulty maintaining typical eye contact or displaying socially expected facial expressions in response to emotions. This doesn't mean they lack emotions, but rather that their expressions may differ from what is considered the norm. Understanding these variations in facial expressions is crucial for building effective communication and empathy with individuals on the autism spectrum, as it allows for a better interpretation of their emotions and intentions. The phases of emotion recognition are face feature extraction and feature categorization [52].

Singh et al. [53] studied ASD as a developmental state that profoundly impacts a person's perception, communication, as well as behaviour, resulting from underlying modifications to the brain. Onset typically occurs before the age of three and may continue throughout a person's life. Individuals with ASD may exhibit higher rates of self-harming behaviour compared to those without the disorder, underscoring the importance of early detection and intervention. To address this need, ML methods like DT, RF, SVM, and NB, as well as DL approaches like VGG16, Dense Net, and Alex Net, can be leveraged to analyse behaviour-based questionnaires and even images for ASD identification. The suggested approach emphasizes the use of transfer learning-based DL algorithms for early ASD detection. The ASD identification process consists of multiple phases, including data collection/acquisition, pre-processing, data augmentation, extraction of features, and classification. This comprehensive strategy has the potential to enhance ASD early diagnosis, which is essential for prompt intervention and assistance.

Ahmed et al. [54] concentrated on the potential of utilizing facial characteristics for the diagnosis of ASD. A neurodevelopmental disorder called ASD can impact the physical appearance of an individual's face, leading to distinct patterns in autistic children. The research aims to facilitate the diagnosis process for families and psychiatrists through a user-friendly web application. This application leverages DL techniques, specifically a CNN with transfer learning, and is built using the Flask framework. The study employs pre-trained models for classification like InceptionV3, Xception, and MobileNet. 3,014 facial images total are included in the dataset; 1,507 of these are from children with autism and 1,507 from nonautistic children, obtained from a Kaggle dataset that is openly accessible. The results indicate high classification accuracy: Mobile Net achieves 95% accuracy, while Xception achieves 94% and InceptionV3 0.89%. This research highlights the potential of using DL-based facial analysis as an accessible and accurate tool for diagnosing autism, offering a promising avenue for early intervention and support.

Mujeeb Rahman and Subashini [55] delve into the possibility of using static characteristics taken from facial photographs of children to serve the role of a biomarker for distinguishing those with ASD from generally developing kids. The research employs using five pre-trained CNN models as feature extractors-MobileNet, Xception, EfficientNetB0, EfficientNetB1, and EfficientNetB2. These features are then used as input for a DNN binary classifier, aiming to accurately identify autism in children. The dataset used for training comprises facial photos of kids with and without autism, obtained from a publicly available source. Among the CNN models, Xception stands out as the top performer, achieving an 88% NPV, 88.46% sensitivity, and 96.63% AUC. EfficientNetB0, on the other hand, consistently predicted a 59% score for both groups with and without autism at a 95% confidence level. These findings highlight the potential of utilizing facial characteristics and DL for the accurate diagnosis of ASD in kids, which can significantly improve early intervention and support.

Narala et al. [56] proposed a model to facilitate the detection of ASD by predicting whether a person, based on their facial images, is either normally developing or autistic. This study utilizes the Autism Image data dataset, consisting of a test set of 300 facial images and a training set of 2530 facial images for model evaluation. This method is constructed using the Efficient Net CNN, achieving an accuracy level of 88%. These findings showcase the potential of utilizing facial characteristics and DL techniques to distinguish between children who are typically developing and those who have ASD, providing a valuable tool for early diagnosis and intervention.

Bhargavi et al. [57] address the critical issue of ASD, which significantly affects social, linguistic, and communication skills. Early person identification with ASD, especially in children, is crucial for implementing timely therapeutic strategies. Individuals with ASD frequently display unique facial features and behaviours that can aid in their recognition. The proposed method leverages training datasets and DL models to train, test, and evaluate a system for accurate autism face recognition, with the potential for integration into a monitoring robot. The study compares the effectiveness of the therapy bot in conjunction with three different transfer learning architectures: VGG16, Mobile Net, and Resnet50. Notably, VGG16 surpasses other transfer learning methods, achieving 97.66% accuracy on real-time images, demonstrating its effectiveness in the context of autism face recognition and monitoring through neural networks.

Tao and Shyu [58] address the challenging problem of diagnosing ASD, a condition that can manifest with diverse symptoms and severities affecting language and behaviour. The study introduces a novel approach called SP-ASDNet, which combines CNN and LSTM networks. This hybrid method analyses the gaze path of a viewer on a particular image to determine whether the viewer has ASD or is TD. The proposed SP-ASDNet was evaluated in the 2019 Saliency4ASD grand challenge, achieving an accuracy of 74.22% during the validation phase. The integration of deep neural networks, particularly CNNs, and LSTMs, for visual data analysis regarding the diagnosis of ASD, represents a viable option for enhancing the accuracy of detection and comprehension of this complex neurodevelopmental disorder.

Kang et al. [59] aimed to recognize children with autism by characteristics from the EEG and eye-tracking modalities data-and using ML methods, specifically SVM. The research included 97 children aged 3 to 6, who underwent rEEG data recording during resting state and eye tracking tests involving both strangers of one race and another race face stimuli. Power spectrum examination was applied to EEG data, while in the eye-tracking data, Areas of Interest (AOI) were chosen for gaze analysis. Feature selection was performed using the Minimum Redundancy Maximum Relevance (MRMR) technique, and SVM classifiers were employed to categorize children with autism and typically developing kids. The study demonstrated that combining data from both modalities resulted in a classification accuracy of up to 85.44%, with an AUC of 0.93 in the case of 32 features chosen. Although the sample was limited to children aged 3 to 6, these findings suggest that the integration of EEG and eye-tracking data via an ML technique may serve as a useful instrument for identifying kids with ASD and enhancing the procedure for diagnosis.

Guo [60] addresses the possibility of creating an automated technique for examining visual cues indicative of ASD by comparing images taken by individuals utilizing ASD to those captured by individuals in a variety of situations without ASD. Individuals with ASD often exhibit abnormal focus on social cues gaze patterns, and facial expressions. Motivated by previous studies that relied on a manual examination of the images, this research aims to automate the process of identifying visual cues associated with ASD. The challenge lies in determining how to characterize these photos effectively for automated separation. The study proposes several features to characterize observable behaviours related to ASD, and through experimental validation, it achieves a prediction accuracy of 85.8 percent. This work represents a pioneering effort in automatically analysing photos taken by individuals with ASD, offering the promising potential for enhancing the understanding and diagnosis of ASD through visual cues.

Lee and Yoo [61] as shown in Table 3 have explored various neural network architectures, including ResNet-18 and Mobile Net, to improve FER performance. Recent studies have introduced innovative strategies such as divide-and-conquer learning, which leverages confusion matrices to group similar facial expressions and retrain models, leading to substantial accuracy improvements across different datasets, including thermal and RGB datasets like Tufts, RWTH, RAF, and FER2013. These advancements reflect a continued effort to enhance the accuracy and applicability of FER systems, with potential applications spanning human-computer interaction, emotion analysis, and affective computing.

4.2.1 Observation

Based on the above study in ASD detection, future work will involve using facial expression datasets for ASD identification, there are several key areas for improvement and exploration to enhance the accuracy, robustness, as well as applicability of ML methods.

Future improvements in facial expression dataset analysis for ASD detection should centre on three critical avenues. Firstly, integrating multimodal data sources alongside facial expressions, such as voice patterns, physiological signals, or behavioral cues, could offer a more comprehensive understanding of ASD-related characteristics. By merging these different data types, ML models could capture a wider range of features, potentially enhancing diagnostic accuracy. Secondly, expanding and diversifying datasets is pivotal for enhancing model generalizability. Collecting larger and more diverse datasets that encompass variations in age, ethnicity, cultural backgrounds, and emotions across different developmental stages would aid in constructing more robust and inclusive models. Additionally, longitudinal data could provide insights into the evolving nature of facial expressions associated with ASD, enabling more accurate and timely diagnoses. Lastly, focusing on interpretability and explainability within facial expression analysis models is crucial. Research efforts should aim to elucidate the rationale behind the model's decision-making processes, making the predictions transparent and understandable. Techniques that highlight the crucial facial features contributing to the model's classifications can foster trust and comprehension among clinicians, researchers, and stakeholders, ultimately advancing ASD diagnostic accuracy and clinical relevance.

5.png

Figure 5. Autism gestures and postures

Table 3. Parametric assessment of ASD detection methods using image-based (facial expression recognition) dataset

4.3 The role of ML and DL in gesture and postures

Body language encompasses a variety of non-verbal cues, including posture, gestures, eye movements, and facial expressions. These are significant indicators of a person's inner emotional and mental state. Many individuals with autism may display typical postures and gestures as shown in Figure 5, such as repetitive hand-flapping, rocking, or unusual body movements. These movements are often seen as self-soothing behaviours individuals with autism may struggle with non-verbal communication, including maintaining eye contact using appropriate facial expressions, which can result in different gestures and postures when interacting with others. Better communication and acceptance of people with autism require an understanding of and respect for distinctive postures and gestures [62].

Yoo et al. [63] address the significance of pointing gestures in separating kids with ASD from those who have TD. As shown in Table 4, the study highlights the absence of datasets created especially with kids' pointing gestures in mind, which results in performance degradation when applying conventional supervised CNNs due to domain shift. To mitigate this issue, the authors suggest an end-to-end learning strategy that uses Self-Supervised Regularization (SSR) for domain-generalized pointing gesture recognition. They validate their approach by developing an ASD diagnostic tool based on Social Interaction-Inducing Content (SIIC) and creating an ASD-Pointing dataset containing 40 TD and ASD children. After conducting several tests, they were able to attain a 72.5% ASD screening accuracy, underscoring the pivotal role of pointing ability in distinguishing between ASD and TD children.

Table 4. Parametric assessment of ASD detection methods using image-based (Gesture and posture) dataset

4.3.1 Observation

Based on the above study in ASD detection future research focusing on gesture and posture datasets for distinguishing between TD and ASD, several key avenues for improvement can be explored.

In future work, enhancements to the SIIC-based tests should consider incorporating a preliminary "warming-up" phase to potentially improve the ASD diagnostic system's performance in recognizing pointing gestures. This additional phase could involve introductory interactions or preparatory activities before capturing the pointing gestures, potentially aiding in acclimatizing the participants, especially children, to the testing environment. Furthermore, efforts to augment the ASD-Pointing dataset by expanding the sample size and diversity, encompassing a broader range of pointing gestures exhibited by children with ASD and TD, would likely fortify the model's robustness and generalizability. Additionally, refining the SSR technique and exploring novel methodologies to alleviate domain shift issues in CNNs for pointing gesture recognition could be valuable avenues for improving accuracy in identifying kids with TD and ASD according to their pointing abilities.

4.4 Synthesis of findings from ML/DL studies across different image-based datasets

This section synthesizes the findings from studies utilizing ML/DL algorithms across various image-based datasets for ASD detection, including brain MRI, facial expression recognition, and gesture/posture analysis.

4.4.1 Brain MRI dataset

1). ML/DL Algorithms Used:

• Studies utilized a range of algorithms such as CNNs, SVMs, DQN, DBN, and others.

• Techniques like CNN were commonly applied for feature extraction from brain MRI scans.

2). Performance Metrics:

• CNN-Based Models: Achieved accuracies ranging from 70% to 100%, depending on the specific dataset and preprocessing techniques [32, 37].

• DQN and DBN: Reported accuracies of up to 92.4% with high sensitivity and specificity values [33, 38].

3). Consistency of Results:

• Results varied based on dataset size, preprocessing methods, and model complexity.

• Advanced DL techniques generally outperformed traditional ML approaches in accuracy and robustness.

4). Best Performing Approaches:

•  DQN: Noted for their high accuracy and robust performance in ASD classification using fMRI data [32].

• DBN with Optimization: Achieved competitive accuracy and efficiency in capturing complex relationships in brain imaging data [36].

• CNNs with Functional Connectivity Patterns: Effective in identifying ASD through functional MRI data, showcasing the potential for clinical application [35].

4.4.2 Facial expression recognition dataset

1). ML/DL Algorithms Used:

• Commonly employed algorithms include CNNs (e.g., VGG16, MobileNet), Transfer Learning (e.g., InceptionV3), and combinations thereof.

• These algorithms focus on analyzing facial features to detect ASD-related patterns in expressions.

2). Performance Metrics:

• CNN-Based Models: Achieved accuracies ranging from 86.7% to 97.66% across different studies [55, 59].

• Transfer Learning Models: Demonstrated high accuracy rates (e.g., MobileNet-95%) in detecting ASD through facial expression analysis [54].

3). Consistency of Results:

• Studies consistently reported high accuracies in facial expression recognition for ASD detection.

• Variability existed in dataset sizes and inclusion criteria, influencing performance metrics.

4). Best Performing Approaches:

• Transfer Learning Models: Particularly effective due to their ability to leverage pre-trained networks for feature extraction from facial images [56, 57].

• CNNs with Ensemble Methods: Showcased potential in enhancing classification accuracy through ensemble learning techniques [55].

4.4.3 Gesture and posture analysis dataset

1). ML/DL Algorithms Used:

• Techniques included CNNs, self-supervised regularization (SIIC), and supervised learning approaches for gesture recognition.

• Focus on distinguishing between ASD and neurotypical individuals based on motor skills and posture.

2). Performance Metrics:

• CNN-Based Models: Reported accuracies around 72.5% to 97.75% depending on the specific approach and dataset [63].

• SIIC-Based Methods: Achieved high recall and precision scores in recognizing subtle differences in gesture patterns [63].

3). Consistency of Results:

• Results varied across studies due to differences in methodology and dataset characteristics.

• SIIC-based methods showed promise in improving gesture recognition accuracy and reliability.

4). Best Performing Approaches:

• End-to-end learning Schemes: Such as those utilizing SIIC for domain-generalized pointing gesture recognition [63].

• Partitioned Learning Models: Introduced novel approaches to enhancing gesture and posture analysis accuracy [61].

The synthesis of findings across brain MRI, facial expression recognition, and gesture/posture analysis datasets using ML/DL algorithms highlights significant advancements in ASD detection methodologies. While variability exists in results due to dataset specifics and methodological differences, advanced DL techniques consistently demonstrate superior performance over traditional ML methods. Future research should focus on standardizing protocols, expanding dataset sizes, and exploring ensemble learning strategies to further refine ASD diagnostic models across different image-based datasets.

6.jpg

Figure 6. Workflow of ASD

The comprehensive review underscores the pivotal role of ML, DL, and computational methodologies in transforming the landscape of ASD diagnosis and understanding. Across various studies employing diverse datasets, neuroimaging modalities, and behavioral analyses, the integration of advanced computational techniques has showcased promising advancements. ML and DL models, including CNNs, reinforcement learning, and clustering analyses, have demonstrated high accuracy rates in ASD detection, emphasizing their potential for early diagnosis and tailored therapeutic interventions. Moreover, the amalgamation of multi-modal data, such as facial images, EEG, and handwriting samples, coupled with innovative computational approaches, offers a pathway toward more precise and objective ASD identification. Furthermore, the integration of AI into healthcare, along with the analysis of neuroimaging data, presents a promising avenue for enhanced early prediction, which could lead to personalized interventions and improved clinical practices. Ultimately, all of the studies that have been reviewed demonstrate the revolutionary effects of computational methodologies in advancing ASD diagnosis, potentially revolutionizing clinical practices and significantly benefiting individuals with ASD in the long term.

Our future scope is to improve,

• Develop methodologies that effectively fuse data from various sources like MRI modalities, behavioral cues, and genetic information into a unified framework. Utilize advanced fusion techniques, such as DL-based multimodal architectures, to capture intricate patterns across diverse data sources, thereby enhancing diagnostic accuracy and providing a more comprehensive understanding of ASD-related characteristics.

• Focus on longitudinal studies analyzing ASD-related data over time, such as brain MRI scans or behavioral patterns investigate how these characteristics evolve across different developmental stages, potentially identifying biomarkers or behavioral patterns that signify the onset or progression of ASD.

• Develop advanced NLP techniques specifically tailored to understand the context and nuances of text data related to ASD. Aim for models that can comprehend the subtle semantic meaning of textual information, capturing the intricacies of ASD-related linguistic behaviors across diverse cultural contexts and age groups.

• Collect a more extensive and diverse dataset that encompasses a wider range of gestures and postures displayed by individuals with ASD and TD across various age groups, cultural backgrounds, and developmental stages. This dataset expansion would provide a more comprehensive understanding of subtle variations in gestures and postures indicative of ASD.

8.1 Key findings

(1) High Accuracy of ML/DL Models:

• ML and DL models, including CNNs, reinforcement learning, and clustering analyses, have demonstrated high accuracy rates in ASD detection, emphasizing their potential for early diagnosis and tailored therapeutic interventions.

(2) Multi-Modal Data Integration:

• The amalgamation of multi-modal data, such as facial images, EEG, and handwriting samples, coupled with innovative computational approaches, offers a pathway toward more precise and objective ASD identification.

(3) Enhanced Early Prediction:

• The integration of AI into healthcare, along with the analysis of neuroimaging data, presents a promising avenue for enhanced early prediction, which could lead to personalized interventions and improved clinical practices.

8.2 Limitations

• Data Quality and Availability:

High-quality, large-scale datasets are required to train robust ML/DL models. Current studies often face challenges related to data heterogeneity and limited sample sizes.

• Validation and Generalizability:

Ensuring the accuracy and generalizability of ML/DL models across diverse populations and clinical settings remains a critical challenge.

• Integration into Clinical Workflows:

Seamlessly incorporating ML/DL tools into existing diagnostic workflows without disrupting clinical routines requires careful planning and user-friendly interfaces.

8.3 Potential impact

Ultimately, all of the studies reviewed demonstrate the revolutionary effects of computational methodologies in advancing ASD diagnosis. The potential impact of these advancements includes:

• Improved Early Detection:

ML/DL models can significantly enhance the early detection of ASD, allowing for timely interventions that can improve long-term outcomes.

• Personalized Interventions:

The ability to analyze individual-specific data can lead to more personalized and effective treatment plans.

• Objective and Precise Diagnosis:

Integrating AI into diagnostic pathways can provide more objective and precise ASD identification, reducing the reliance on subjective assessments.