Classification of neurodevelopmental disorders and typical development using deep learning and a portable patch-type electroencephalography device

A Scoping Review of Machine Learning and Deep Learning Methods for Autism Spectrum Disorder Diagnosis and Analysis.

Autism spectrum disorder (ASD) is a heterogeneous neurodevelopmental condition characterized by diverse behavioral, cognitive, sensory, and communication profiles, making early diagnosis and personalized intervention challenging. Recent advances in machine learning (ML) and deep learning (DL) have enabled the development of computational tools for ASD screening, classification, severity assessment, and intervention monitoring. This review synthesizes findings from 50 recent studies that applied ML and DL techniques to ASD-related datasets, including electroencephalography (EEG), eye-tracking, behavioral video, microbiome, voice acoustic, demographic, and multimodal data. The review addresses three key questions: (i) which data modalities and computational approaches are most frequently used, (ii) how diagnostic performance is evaluated across different study designs, and (iii) what methodological challenges limit clinical translation. The literature is organized according to data modality, algorithmic approach, and clinical readiness. Approaches examined include conventional ML methods, convolutional neural networks, graph neural networks, hybrid deep learning architectures, federated learning, explainable artificial intelligence, topological data analysis, and multimodal fusion. The findings suggest that multimodal and graph-based approaches provide a more comprehensive representation of ASD phenotypes than single-modality methods. Explainability and privacy-preserving learning have also emerged as important considerations for clinical deployment. However, many reported high-performance models are based on small sample sizes, repeated use of the ABIDE dataset, class imbalance, single-site validation, or limited external testing, raising concerns regarding generalizability. Beyond diagnostic accuracy, this review evaluates model interpretability, calibration, scalability, validation rigor, and clinical applicability. Overall, the analysis highlights the need for standardized benchmarks, externally validated multimodal datasets, clinically relevant evaluation metrics, and decision-support systems that complement rather than replace expert clinical assessment in ASD diagnosis and management.

Transdiagnostic monoamine-based subtyping for attention-deficit/hyperactivity disorder and autism spectrum disorder via unsupervised machine learning.

Neuroimaging and molecular studies have examined the etiology of attention-deficit/hyperactivity disorder (ADHD) and autism spectrum disorder (ASD). However, their findings remain inconsistent because of within-disorder heterogeneity and cross-disorder phenotypic overlap. We sought to identify monoamine-based subtypes across ADHD and ASD and clarify their brain structural characteristics. In 83 children with ADHD and/or ASD, we applied unsupervised machine learning (NbClust with K-means) to identify neurodevelopmental disorder (NDD) phenotypes using urinary monoamine metabolite (MM) profiles. Behavioral symptoms, cognitive performance, cortical surface area, and gray matter volume (GMV) were evaluated for each NDD phenotype and for 83 typically developing (TD) children as controls. Clustering identified two urinary MM-defined NDD phenotypes: NDD-A (n = 18, including 5 ADHD, 2 ASD, and 11 ADHD + ASD cases), characterized by high levels of 4-hydroxy-3-methoxyphenylglycol, 5-hydroxyindoleacetic acid, and homovanillic acid, and NDD-B (n = 65, including 16 ADHD, 19 ASD, and 30 ADHD + ASD cases), characterized by low levels of these metabolites. Urinary 4-hydroxy-3-methoxyphenylglycol levels correlated positively with social communication difficulties in NDD-A. NDD-B showed significantly lower cognitive control, cognitive flexibility, and inhibitory control than TD. Structurally, compared with TD, NDD-A showed significant surface area enlargement in the isthmus cingulate gyrus, whereas NDD-B exhibited significant GMV reductions primarily in fronto-opercular/orbitofrontal regions, with additional reductions in the superior parietal lobule and supramarginal gyrus. These findings suggest that urinary MM-defined NDD-A and NDD-B phenotypes are associated with distinct cognitive and brain structural characteristics. Such phenotype specificity may provide a novel framework for understanding within-disorder heterogeneity and cross-disorder phenotypic overlaps.

A Novel EEG-Based Topographic Brain Map–Driven Deep Learning Method for Autism Spectrum Disorder Detection in Children

Temporal Deep Learning with Multiscale Principal Component Features for Autism Classification from Electroencephalographic Signals

Autism spectrum disorder (ASD) is a heterogeneous neurodevelopmental condition whose early identification remains challenging because clinical assessment still relies heavily on behavioral observation and expert judgment. Electroencephalography (EEG) offers a noninvasive approach for capturing neural dynamics. Still, EEG-based ASD classification remains difficult because of signal nonstationarity, limited sample sizes, and the risk of subject-identity leakage. This study proposes a comparative temporal deep learning framework for EEG-based ASD classification by evaluating principal component analysis (PCA) and multiscale principal component analysis (MS-PCA) as feature representations combined with a recurrent neural network with bidirectional long short-term memory (RNN-BiLSTM) and a temporal convolutional network with self-attention (TCN-SA). Resting-state eyes-open EEG signals were acquired from only 10 participants, consisting of 5 individuals with ASD and 5 typically developing controls, using a 16-channel acquisition system. The signals were filtered using a fourth-order Butterworth band-pass filter, transformed into PCA or MS-PCA representations, segmented into 4 s windows with 50% overlap, and evaluated using subject-wise 5-fold cross-validation to reduce subject-identity leakage. The results showed that MS-PCA produced higher descriptive performance than PCA in both temporal architectures, with the strongest descriptive result obtained by the MS-PCA + TCN-SA scheme, which achieved a mean accuracy of 97.96 ± 2.37% and balanced precision, recall, F1-score, and specificity. However, the inferential comparison between PCA and MS-PCA did not reach statistical significance at the 0.05 level, and the cohort size was limited to 10 participants. Therefore, these findings should be interpreted as preliminary descriptive evidence within the present cohort rather than evidence of diagnostic readiness or robust clinical applicability. Larger, independent, and demographically diverse EEG datasets with richer clinical characterization are required to confirm the observed trend and evaluate the generalizability of the proposed framework.   

Group-wise sparse coding for the discovery of functional connectivity-based biomarkers in children and adolescents with autism spectrum disorder

Background Autism spectrum disorder (ASD) is currently diagnosed through behavioral observations and evaluations, but there is still a lack of objective and consistent biomarkers. Children and adolescents with ASD exhibit impairments in advanced social, emotional, and cognitive functions, such as a lack of empathy. In general, the concept of empathy encompasses several socio-emotional and cognitive components based on interacting brain circuits. Identification of disease-related biomarkers at the brain network level could provide a crucial avenue for advancing ASD imaging research and improving diagnostic accuracy. Methods This study examined 80 individuals with ASD aged 6–16 years old and 50 matched control subjects, using resting-state functional magnetic resonance imaging and clinical psychological assessment datasets. Specifically, a set of functional brain networks was constructed using dictionary learning and sparse coding (DLSC) in a group-wise manner. Then, the localized common functional brain networks from both the ASD and matched control groups were automatically decomposed into a set of regions of interest (ROIs) for further functional connectivity analyses. Results Using the derived functional connectivity matrix, we investigated three parameters, namely, correlation, partial correlation, and tangent embedding, to differentiate participants with ASD from control subjects. We achieved classification accuracies of 95%, 100%, and 100%, respectively, indicating that the proposed DLSC method could extract representative and characteristic brain ROI atlases for both ASD and control participants. Further analysis of functional connectivity results showed that ASD participants had multiple atypical connections, especially those connecting the left inferior temporal and left inferior parietal regions, which belonged to the temporoparietal junction (TPJ), and connections related to the right insula and anterior cingulum, which belonged to the salience network (SN). Together, our results suggest that individuals with ASD exhibited a lower empathy capability than control subjects. Conclusion Our results suggest that DLSC can effectively extract robust brain ROI atlases. Functional connectomes with high differentiation powers were mainly distributed within the brain networks of SN, social brain networks (SBNs), and the theory of mind (ToM) network (including the TPJ hub). Children and adolescents with ASD exhibited lower empathy capabilities than control subjects, which may be attributed to dysfunctions in the salience and social brain networks. Clinical Trial Registration https://www.chictr.org.cn, identifier ChiCTR-ROC-17012877.

Multiscale deep learning convolutional neural network for ADHD detection using EEG

Attention-Deficit/Hyperactivity Disorder (ADHD) is a common neurodevelopmental condition in children. Electroencephalography (EEG) may support objective assessment, but subject-independent ADHD classification remains challenging because EEG varies across participants and recording segments. We developed the EEG Multiscale Convolutional Network (EEG-MSCNet), a multiscale convolutional architecture for pediatric EEG classification, and evaluated it using subject-level leave-one-subject-out (LOSO) validation on a public EEG dataset. The study also examined preprocessing choices, architectural ablations, channel-region subsets, computational complexity, and repeated-run stability. The final EEG-MSCNet evaluation achieved an Accuracy of 87.6%, F1-score of 87.6%, and area under the receiver operating characteristic curve (AUROC) of 88.2%. Artifact subspace reconstruction (ASR) combined with Fully Automated Statistical Thresholding for EEG artifact Rejection (FASTER) emerged as the strongest preprocessing configuration. The multiscale architecture showed stronger observed Accuracy and F1-score than reduced architectural variants, and the full montage provided the strongest channel configuration. These findings show that subject-level validation, preprocessing, and architectural decisions materially influence EEG-based ADHD classification. EEG-MSCNet contributes both a model and a systematic evaluation framework, supported by source code, configurations, logs, and the public dataset reference for reproducible comparison.