ASD: Clinical and Computational Perspectives
- Autism Spectrum Disorder (ASD) is a neurodevelopmental condition marked by social communication impairments, repetitive behaviors, and a wide range of clinical presentations.
- Current research integrates questionnaire-based screening with EEG, neuroimaging, and multimodal computational pipelines to enhance early detection and support biomarker discovery.
- Innovative methods leveraging speech, gaze, video analyses, and machine learning address diagnostic challenges while mitigating issues like data imbalance and privacy constraints.
Searching arXiv for recent ASD-related papers to ground the article. Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by difficulties with social interaction, communication, and repetitive or atypical behaviors, with marked heterogeneity across severity, age, and modality of presentation. Contemporary research on ASD diagnosis emphasizes early detection, the absence of robust objective biomarkers, and the consequent proliferation of computational pipelines spanning questionnaire screening, speech and voice analysis, gaze and interactional behavior, EEG, structural MRI, and resting-state fMRI; in unrelated machine-listening literature, the same acronym “ASD” also denotes anomalous sound detection (Rasul et al., 2023, Jayawardana et al., 2019, Zheng et al., 2024, Zhang et al., 2024).
1. Clinical characterization and heterogeneity
ASD is described in the cited literature as a developmental or neurodevelopmental disorder associated with social communication impairments, difficulties perceiving and presenting communication cues, restricted interests, and repetitive behaviors. Several sources also stress that symptom expression is heterogeneous, especially in milder forms, and that this heterogeneity complicates diagnosis, subgrouping, and biomarker discovery (Vacca et al., 2024, Chen et al., 2024).
A persistent theme is the absence of a single robust objective marker. Questionnaire surveys and behavioral observations remain central in practice, but these are repeatedly described as subjective, time-consuming, labor-intensive, and prone to false positives or false negatives. Neuroimaging, EEG, gaze, speech, and multimodal behavioral analysis are therefore being investigated as quantitative complements rather than simple replacements (Jayawardana et al., 2019, Zheng et al., 2024, Li et al., 2024).
The literature also emphasizes that ASD variability is not only behavioral but dynamical. Longitudinal gaze studies report user-specific timelines and novelty effects during intervention, while topological analyses of brain-state trajectories report greater heterogeneity across multiple metrics in ASD cohorts. This suggests that ASD is increasingly being operationalized as a spectrum of measurable phenotypes rather than a single invariant signature (Ramnauth et al., 5 Jan 2025, Chen et al., 2024).
2. Questionnaire-based screening and tabular machine learning
Questionnaire-driven screening remains one of the most mature computational settings for ASD detection. A 2023 evaluation used three UCI datasets: a children’s dataset with 292 instances and 21 attributes, an adult dataset with 704 instances and 21 attributes, and a combined dataset with 996 instances and 21 attributes. Eight supervised classifiers were evaluated with hyperparameter optimization and 5-fold cross-validation using accuracy, precision, recall, specificity, F1-score, AUC, kappa, and log loss. On the children’s dataset, SVM and logistic regression each achieved 100% accuracy; on the adult dataset, logistic regression achieved 97.14% accuracy; on the combined dataset, the proposed ANN achieved 94.28% accuracy. The same study also evaluated five clustering algorithms without true labels and found spectral clustering best overall on NMI and ARI for adult and combined data, while a Chi-square test identified difficulty understanding others’ emotions (A4) as the key feature in children and avoidance of physical contact (A9) as paramount in adult and combined data. A GUI tool was also developed for clinician use (Rasul et al., 2023).
A related ANN study used the adult UCI screening dataset after excluding missing data, retaining 704 records with 20 attributes and a 60.2%/19.9%/19.9% training-selection-testing split. The model was a shallow feedforward neural network trained with the Levenberg-Marquardt algorithm and incremental order selection; the optimal structure contained 1 hidden neuron. No data balancing method was applied, but weighted squared error was used to address class imbalance. Reported test performance was 98.38% accuracy with AUC of 1.0, and the confusion matrix showed 61 of 62 actual positives correctly classified and all 78 actual negatives correctly classified (Choudhury et al., 2018).
Taken together, these studies show that highly parameterized models are not the only successful approach in ASD screening. Logistic regression and SVM remain highly competitive in questionnaire-based settings, while ANN variants can be advantageous when architecture and loss design are tightly matched to dataset structure.
3. Speech, voice, gaze, and interactional behavior
Speech and language have become important non-invasive substrates for ASD detection. A study using TalkBank speech transcripts evaluated logistic regression, random forest, SVM, Naive Bayes, and KNN on the Nadig and Eigsti datasets, with more than 50 linguistic and demographic features, mean imputation for missing values, integer encoding for categorical variables, SMOTE for class balancing, and RFE for feature selection. Logistic regression and random forest each achieved accuracy of 0.75 on the Nadig dataset, while random forest achieved 0.80 on the Eigsti dataset. Key retained features included MLU, MLT ratio, proportions of pronouns, conjunctions, negations, and the total number of words spoken by the child (Ramesh et al., 2021).
Acoustic-prosodic modeling has produced stronger performance on smaller curated corpora. On the ASDBank Dutch Asymmetries Corpus SK sub-corpus, comprising 46 children with ASD and 38 with typical development, 36 candidate features were extracted after manual clipping, noise filtering, pre-emphasis, and normalization. T-test-based feature selection within each 5-fold cross-validation cycle retained eight features: ZCR, fundamental frequency, MFCC3, MFCC4, MFCC6, MFCC8, MFCC12, and FD1. Both SVM and random forest achieved accuracy of 0.988, with F-measure approximately 0.989; all selected features had statistically significant between-group differences at (Vacca et al., 2024).
Eye movement studies have advanced a more explicit critique of architecture choice. The DSTS framework for ASD detection from eye movement data argues that eye movement signals mainly contain short-term and localized dependencies, and that stacked attention layers yield only limited benefits in this setting. Across eight real-world datasets, DSTS outperformed both traditional machine learning and more sophisticated deep learning baselines, while adding a Class-aware Representation mechanism and an Imbalance-aware mechanism (Huang et al., 9 Jan 2026).
Video-based behavior analysis has expanded from feature engineering to structured interaction protocols. The Parent-Child Dyads Block-Play protocol generated a dataset of 40 ASD and 89 TD toddlers, with 187 annotated video clips in the working subset. Using skeleton joint data extracted after person detection and pose estimation, the hybrid 2sGCN-AxLSTM model achieved 89.6% accuracy and unweighted average recall of 0.85. The protocol was designed to elicit differences in joint attention, intention, task engagement, and fine/coarse motor behavior, and the reported qualitative observations describe ASD toddlers as needing greater parental help and exhibiting more disrupted or atypical movements (Li et al., 2024).
A complementary benchmark, the Video ASD dataset, contains feature data from 2,467 videos and approximately 1.4 million frames from 108 children reacting to chemo-sensory and other stimuli. It provides 512-dimensional latent vectors per frame from ViT-Base and ConvNext-Base, head pose angles for 625,007 faces, and full-sentence temporal labels for taste and smell videos. Baseline temporal transformer experiments showed that head pose filtering improved performance, with accuracy ranges of approximately 55–78% depending on split and pose range, while also demonstrating overfitting and degradation under batch-wise generalization (Serna-Aguilera et al., 2024).
Gaze can also function as an intervention outcome variable rather than only a diagnostic feature. In a month-long, in-home social robot intervention, spontaneous gaze toward caregivers, mutual gaze, and joint attention increased, with significant week effects and contingency effects after robot gaze shifts. The study also reported that novelty effects dominated the first two weeks and that diagnostic measures such as ADOS, ADI-R, and DAS-II were strong predictors of gaze patterns for both children and caregivers (Ramnauth et al., 5 Jan 2025).
4. EEG and neurophysiological markers
EEG has been proposed as a non-invasive and relatively inexpensive objective measure for ASD classification. One extended study used a 32-channel LiveAmp wireless EEG system sampling at 250 Hz, referenced to FCz, with preprocessing in EEGLAB/MATLAB that included a 1 Hz high-pass filter, CleanLine for line noise, bad channel rejection, Artifact Subspace Reconstruction, interpolation, average re-referencing, and ICA. Feature families included mean, standard deviation, Shannon entropy, band power in delta, theta, alpha, beta, and gamma bands, Morlet wavelet time-frequency power, and magnitude-squared coherence between channels (Jayawardana et al., 2019).
Performance depended strongly on feature type and task. For EEG features during joint attention tasks, the best reported accuracy was 78% with logistic regression using EEG standard deviation plus eye data and no PCA. For EEG coherence during social interaction tasks, using all electrodes, random forest achieved 98% accuracy and JRip 98.06%, while logistic regression achieved 96.6%. Using only 10 selected electrodes, random forest still achieved 97%. For time-frequency decomposition, non-linear models reached above 92% accuracy, and long-term trend analysis with CNN achieved above 90% accuracy for ASD classification. The same source also reported regression of ADOS-2 severity with up to 0.86 and RMSE around 2.9 (Jayawardana et al., 2019).
At the same time, these results are explicitly framed as preliminary. The study notes relatively small cohorts, vulnerability of deep models to the curse of dimensionality, the importance of ecologically valid paradigms, and the need for automated, robust preprocessing pipelines. EEG therefore occupies a dual status in the current literature: it is promising as an objective marker, but its translational stability remains dependent on larger and more diverse validation cohorts.
5. Neuroimaging, connectomics, and biomarker discovery
Neuroimaging research on ASD spans structural MRI, resting-state fMRI, dynamic connectomics, and topological analysis. In structural MRI, a contrastive variational autoencoder was trained on 78 scans from children aged 0.92–4.83 years, separating ASD-specific features from shared features through two encoders and a shared decoder. Random forest classification on the extracted features achieved mean accuracy above 0.97 across cross-validation settings, and transfer learning from ABIDE-I improved performance in the extremely small-sample regime. Neuroanatomical interpretation using representational similarity analysis linked ASD-specific features to the supramarginal gyrus and inferior temporal cortex (Ma et al., 2023).
In rs-fMRI, feature dimensionality is a primary obstacle. A denoising variational autoencoder pipeline on the IMPAC dataset used Power and Craddock atlases to derive more than 30 thousand connectivity features and compressed them into 5 latent Gaussian distributions, corresponding to 10 latent features. With site harmonization, SVM on the DVAE features achieved a 95% confidence interval for accuracy of ; without DVAE, SVM accuracy was 0.70, which falls within that interval. The DVAE pipeline was reported as 7 times shorter in runtime than training classifiers directly on the raw data, and the Power atlas yielded better performance than the Craddock atlas (Zheng et al., 2024).
Dynamic modeling has pushed beyond static connectivity matrices. BrainTWT constructed dynamic brain connectomes from ABIDE with sliding windows, temporal random walks, and Transformer-based dynamic network embedding. On 871 subjects, BrainTWT with temporal dynamics achieved accuracy of and AUC of , outperforming the same framework without temporal dynamics, which yielded accuracy. In parallel, a topological and graph-theoretical analysis of dynamic functional connectivity found that ASD brain-state networks tend to have decreased modularity, higher von Neumann entropy, increased Betti-0 numbers, and decreased Betti-1 numbers, indicating less organized and more variable brain dynamics (Piriyasatit et al., 16 Mar 2025, Chen et al., 2024).
Several ABIDE-based classifiers aim simultaneously at diagnosis and biomarker localization. MADE-for-ASD used a multi-atlas deep ensemble with AAL, CC200, and EZ atlases plus demographic information, achieving 75.20% accuracy on the whole ABIDE I dataset and 96.40% on the NYU subset, with sensitivity of 82.90% and specificity of 69.70% on the whole dataset; F-score ranking highlighted precuneus and anterior cingulate/ventromedial among the top 10 ROIs. ASD-HNet hierarchically extracted ROI-scale, community-scale, and global features and reported 73.57% accuracy, 72.63% sensitivity, and 74.90% specificity on ABIDE-I. ASDFormer used a Transformer with a Mixture of Pooling-Classifier Experts and achieved AUROC of , accuracy of , and sensitivity of , while identifying disruptions involving the DMN, SMN, FPN, and Limbic communities (Liu et al., 2024, Luo et al., 2024, Izadi et al., 19 Aug 2025).
A plausible implication is that neuroimaging-based ASD diagnosis is now less about any single classifier and more about representation design: disentangling shared and ASD-specific structure, compressing very high-dimensional connectivity, modeling temporal evolution, and exposing network-level saliencies that can support biomarker discovery.
6. Datasets, privacy, reproducibility, and open problems
Open and privacy-preserving resources are increasingly central to ASD research. MMASD provides multimodal intervention data from 32 children with ASD, with 1,315 segmented samples from more than 100 hours of recordings. Each sample contains optical flow, 2D skeleton, 3D skeleton, and clinician ASD evaluation scores, including ADOS-2 raw scores and ADOS Comparison Score. The dataset is intended for therapy monitoring, individualized treatment, action quality assessment, interpersonal synchrony estimation, and cognitive status tracking (Li et al., 2023).
The Video ASD benchmark makes a different trade-off: it releases feature representations, head pose angles, and textual annotations rather than raw video, thereby reducing privacy risk while enabling foundation-model benchmarking on a large corpus of sensory-response behavior. Both MMASD and Video ASD illustrate a broader infrastructural shift from isolated private datasets toward benchmarkable, privacy-aware resources (Serna-Aguilera et al., 2024, Li et al., 2023).
Privacy constraints also motivate distributed learning. A federated learning study for multi-aspect ASD detection used behavioral questionnaire data, facial images, and a merged feature representation based on facial landmarks plus behavioral variables. In the federated setting, behavioral prediction reached 70% accuracy, facial prediction 62%, and the merged model 63%; a regular logistic regression model on merged data reached 65%. The scheme keeps data local to screening centers and transmits only model weights to the global aggregator (Shamseddine et al., 2022).
Several limitations recur across the literature: small sample sizes, data imbalance, restricted age ranges, single-language corpora, site effects, variable recording conditions, and domain shift. These limitations are explicitly identified in questionnaire screening, transcript-based speech modeling, acoustic voice analysis, and deep behavioral video analysis, and they help explain why near-perfect results on small or curated datasets coexist with more modest performance on large multi-site neuroimaging benchmarks (Rasul et al., 2023, Ramesh et al., 2021, Vacca et al., 2024, Li et al., 2024).
A plausible implication is that headline accuracy values in ASD research are not directly comparable across modalities or cohorts. Reported values range from 100% on a children’s questionnaire dataset, to 0.988 on a small Dutch voice corpus, to approximately 70–75% on large multi-site rs-fMRI benchmarks. The field’s current trajectory therefore combines three priorities: larger and more diverse datasets, modality-aware architectures rather than one-size-fits-all models, and privacy-preserving data sharing frameworks capable of supporting reproducible biomarker research at scale.