A computational neuroimaging study has introduced MBF-Net, an interpretable graph neural network framework designed to characterize brain network organization across autism spectrum disorder, attention-deficit/hyperactivity disorder, and their co-occurring clinical presentation. Published in the Journal of King Saud University Computer and Information Sciences, the investigation combines resting-state functional magnetic resonance imaging data, frequency-specific time-series decompositions, inter-regional phase synchronization, and demographic embeddings to investigate complex neurodevelopmental alterations. In an exploratory three-class classification task distinguishing between individuals with ADHD alone, ASD alone, and comorbid ASD and ADHD across a multi-site cohort of 334 participants, the proposed model achieved an overall accuracy of 73.69 percent, a macro-averaged F1 score of 69.32 percent, and a macro-area under the receiver operating characteristic curve of 81.32 percent under five repeated stratified ten-fold cross-validations. While these quantitative findings demonstrate the feasibility of automated connectome modeling, the authors emphasize that the system provides exploratory cohort stratification and hypothesis-generating biological insights rather than a standalone, clinically validated diagnostic instrument.
Autism spectrum disorder and attention-deficit/hyperactivity disorder represent two of the most prevalent neurodevelopmental conditions emerging in early childhood, affecting approximately one to two percent and five to six percent of the general population, respectively. In everyday clinical environments, establishing an accurate differential diagnosis presents considerable difficulties due to shared neuropsychological, behavioral, neuroimaging, and genetic features that obscure clear diagnostic boundaries. Moreover, clinical investigations report that ADHD co-occurs in roughly forty to seventy percent of children diagnosed with ASD, making it one of the most frequent psychiatric comorbidities in this group. When clinicians fail to identify overlapping presentations, patients frequently encounter misdiagnosis, diagnostic delays, or suboptimal intervention strategies that can exacerbate behavioral challenges and social impairments over time. Functional magnetic resonance imaging provides a noninvasive window into these conditions by recording blood-oxygen-level-dependent hemodynamic signals across distributed regions, but conventional connectomic models frequently struggle to isolate subtle, frequency-dependent network alterations across overlapping diagnostic categories.
To assemble the experimental cohorts required to train and evaluate the deep learning architecture, the investigative team curated resting-state neuroimaging scans and clinical phenotypic records from three major open-access repositories: the ADHD-200 consortium, ABIDE I, and ABIDE II. For initial binary benchmarking against neurotypical development, the researchers selected 299 individuals with ADHD and 358 healthy controls across four acquisition sites within ADHD-200, alongside 403 individuals with ASD and 468 healthy controls across international centers within ABIDE I. For the more challenging three-class comorbidity investigation, the study constructed a unified cohort of 334 subjects, comprising 147 individuals with ADHD alone from the New York University site of ADHD-200, 90 individuals with ASD alone from seven ABIDE II acquisition sites, and 97 individuals with documented comorbid ASD and ADHD confirmed through explicit diagnostic metadata across the same seven ABIDE II locations. The resulting three-class cohort exhibited an average age of 11.07 years in the ADHD group, 13.83 years in the ASD group, and 10.46 years in the comorbid group, without missing age or sex entries.
All raw neuroimaging datasets underwent standardized preprocessing using the FMRIB Software Library toolbox, incorporating brain extraction, rigid-body head motion correction, spatial smoothing with a six-millimeter full-width at half-maximum Gaussian kernel, temporal band-pass filtering, and affine spatial normalization to the Montreal Neurological Institute standard space. Standardized mean time series were subsequently extracted across 116 anatomical regions of interest parcellated according to the Automated Anatomical Labeling 3 atlas. Recognizing that spontaneous low-frequency oscillations govern critical cognitive, attentional, and default-mode processes, the researchers decomposed each regional time series into four distinct slow-wave frequency bands: SLOW-2, SLOW-3, SLOW-4, and SLOW-5. For every anatomical region and frequency band, the framework extracted four statistical descriptors—mean activity, standard deviation, skewness, and kurtosis—alongside local dynamical structure captured by the reconstruction weight variance of five-nearest-neighbor Local Linear Embedding. This decomposition produced twenty cross-band node features per anatomical structure, forming an informative node feature matrix for subsequent graph computation.
To represent inter-regional communication across the functional connectome, the researchers quantified pairwise phase synchronization by calculating phase-locking values across regional time series. For each pair of brain regions, analytic signals were generated via the Hilbert transform to extract instantaneous phase trajectories over time. The resulting phase-locking values, bounded strictly between zero and one, measured the consistency of phase differences across the entire scanning run. In parallel, the framework evaluated the cosine similarity between the twenty-dimensional multi-band feature vectors corresponding to each regional pair. Brain graphs were subsequently constructed by linearly combining phase-locking values and node-feature cosine similarities using equal fixed weights of 0.5. To suppress noisy or biologically spurious inter-regional associations and maintain a sparse network topology suitable for message passing, edges were retained only when their combined similarity score exceeded an empirical threshold of 0.3, with a fully connected fallback implemented in the event of complete edge elimination.
Because raw phase-locking value matrices across all four frequency bands yielded a high-dimensional feature vector of 26,680 unique inter-regional connections per subject, dimensionality reduction was essential to avoid severe statistical overfitting. The authors implemented a nested recursive feature elimination protocol paired with an L2-regularized RidgeClassifier. Crucially, to prevent data leakage and overly optimistic performance estimates, this supervised feature ranking and selection pipeline was executed strictly and exclusively on the training subjects within each outer cross-validation fold, without exposing held-out test data to ranking calculations or hyperparameter optimization. The iterative elimination operated with a batch step factor of 0.2, discarding the lowest twenty percent of contributing connection weights at each round down to a minimum candidate threshold. Candidate feature subsets underwent fine refinement using inner stratified cross-validation, ultimately retaining fold-specific subsets ranging between 1,247 and 2,615 highly discriminative connectivity features for downstream classification.
The structural core of the MBF-Net framework is its second-stage graph reasoning submodule, termed the Selective-Node Dynamics Network. To address the pronounced spatial heterogeneity of neurodevelopmental conditions, this module avoids uniform information propagation by coupling hierarchical graph transformations with local convolutional aggregation. The submodule first routes the 116 anatomical node representations through a Graph Transformer U-Net featuring three encoding layers that combine TransformerConv operations with TopKPooling. Each pooling step halves the node count while preserving a 128-dimensional hidden representation, condensing whole-brain topology down to fourteen core nodes before progressively unpooling and reconstructing full spatial resolution through residual skip connections. In a parallel branch, three stacked residual graph convolutional blocks normalize and filter local neighborhood features across the batch dimension. A parameterized node-wise gating mechanism then dynamically computes element-wise weights to blend global topological context from the U-Net with local neighborhood representations from the convolutional pathway.
Downstream decision learning in MBF-Net relies on a multimodal fusion architecture that combines graph-level representations with connectivity features and demographic profiles. Regional node embeddings are collapsed through global average pooling into a compact graph vector, which is mapped alongside the selected phase-locking features into a shared 64-dimensional attention space. Simultaneously, a Variational Autoencoder pretrained strictly on training-fold data encodes demographic covariates—including standardized age, binary sex, and a categorical one-hot site vector—into a 128-dimensional stochastic phenotypic embedding. A multimodal attention mechanism evaluates the relative predictive contribution of the graph topology, selected connectivity weights, and phenotypic embeddings, projecting their dimension-matched weighted sum into a two-layer classification multilayer perceptron regularized by dropout. The training objective optimizes a compound loss function that combines class-weighted Focal Loss to counter sample imbalance, a graph smoothness penalty that preserves representation similarity across connected anatomical nodes, and an adversarial loss designed to suppress scanner-specific biases.
Because multi-center neuroimaging repositories inherently incorporate scanner hardware differences, sequence protocol discrepancies, and variable signal-to-noise ratios, the researchers deployed a learning-based site-adversarial debiasing branch. A lightweight site discriminator was connected to the concatenated graph and connectivity embeddings via a gradient reversal layer, leaving the explicit phenotypic site embedding untouched for covariate-aware classification. The reversal parameter followed a scheduled ramp that remained at zero during early training epochs and scaled linearly between epochs 40 and 160. During training, tracking curves demonstrated that site-prediction accuracy dropped steadily from roughly 84 percent down to 54 percent as the adversarial penalty saturated, while disease classification accuracy stabilized near 74 percent. This divergence verified that the internal imaging representations became substantially less site-specific without sacrificing disease-relevant diagnostic signatures, although residual coupling remained present due to the structural composition of the multi-site cohorts.
Comprehensive performance benchmarks demonstrated competitive classification capabilities across multiple experimental setups. In the exploratory three-class comorbidity task, MBF-Net secured the highest mean classification accuracy and macro-AUC among all tested machine learning and graph neural network baselines, though BrainGNN achieved a higher mean macro-F1 score of 70.58 percent compared to 69.32 percent for MBF-Net. Detailed class-wise evaluation revealed that the comorbid subgroup yielded the lowest sensitivity, achieving a recall of 62.9 percent, with misclassifications predominantly assigned to ASD-only and ADHD-only classes, highlighting minority-class boundary overlap as an ongoing technical hurdle.
Post hoc neurobiological analyses provided detailed visualizations of abnormal connectivity patterns and localized functional alterations across clinical groups. Benjamini–Hochberg false discovery rate correction identified significant group-level differences in the amplitude of low-frequency fluctuations across the slow-wave spectrum. Relative to single-disorder groups, participants with comorbid ASD and ADHD demonstrated marked activity alterations in the right posterior cerebellum, right inferior frontal gyrus, right pallidum, left inferior parietal lobule, and right fusiform gyrus. Concurrently, overlapping increases in spontaneous activity emerged within the right thalamus and left postcentral gyrus, suggesting potential compensatory involvement in sensory integration and attentional control. Group-level phase-locking value comparisons against neurotypical controls highlighted distributed disruptions across frontostriatal, dorsal attention, and default mode networks in ADHD, impaired social-language and temporal-occipital integration in ASD, and cross-system synergistic dysconnectivity spanning emotional memory, executive control, and sensorimotor networks in comorbid individuals.
Despite these encouraging modeling advancements, the authors delineate critical structural limitations that constrain current clinical translation. Foremost among these is an inescapable dataset-site-label coupling within the three-class cohort: all 147 individuals with ADHD alone originated from a single site within ADHD-200, whereas all ASD-only and comorbid participants were drawn from seven ABIDE II locations. To directly probe this confounding factor, the team performed a supplementary within-database binary analysis restricted to ABIDE II subjects with ASD alone versus comorbid ASD and ADHD. In this setting, pooled ten-fold cross-validation accuracy dropped to 67.8 percent, and leave-one-site-out evaluation across the seven sites yielded an unweighted mean accuracy of 64.0 percent. These supplementary results confirm that cross-dataset composition partially elevated the three-class performance metrics, underscoring that current evidence supports exploratory characterization of comorbidity far more reliably than universal, site-independent differential diagnosis until prospective, balanced multi-site cohorts become accessible.
Subject of Research: Multimodal graph neural network modeling of multiband resting-state fMRI connectivity in ASD and ADHD comorbidity
Article Title: Multiband functional dysconnectivity patterns for exploratory identification of ASD–ADHD comorbidity
Article References: Ye, H., Han, Q., Niu, L., Dong, X., Dong, H., Xue, Y., Zhao, J., & Kuang, Z. (2026). Multiband functional dysconnectivity patterns for exploratory identification of ASD–ADHD comorbidity. Journal of King Saud University Computer and Information Sciences, 38(9), Article 973. https://doi.org/10.1007/s44443-026-01345-7
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01345-7
Keywords: MBF-Net, Autism Spectrum Disorder, Attention-Deficit/Hyperactivity Disorder, Comorbidity, Resting-State fMRI, Phase-Locking Value, Graph Neural Networks, GTU-Net, SND-Net, Multiband Analysis, Amplitude of Low-Frequency Fluctuations, ABIDE II, ADHD-200, Site-Adversarial Training, Functional Connectome, Variational Autoencoder, Recursive Feature Elimination
Cite Scienmag News
Cassandra Pierce. (November 21, 2026). Multiband Graph Neural Network Explores Brain Connectivity in ASD-ADHD Comorbidity. Scienmag. https://scienmag.com/multiband-graph-neural-network-explores-brain-connectivity-in-asd-adhd-comorbidity/
Cassandra Pierce. "Multiband Graph Neural Network Explores Brain Connectivity in ASD-ADHD Comorbidity." Scienmag, 21 November 2026, https://scienmag.com/multiband-graph-neural-network-explores-brain-connectivity-in-asd-adhd-comorbidity/. Accessed 21 November 2026.
Cassandra Pierce. "Multiband Graph Neural Network Explores Brain Connectivity in ASD-ADHD Comorbidity." Scienmag. November 21, 2026. https://scienmag.com/multiband-graph-neural-network-explores-brain-connectivity-in-asd-adhd-comorbidity/








