Survey reports that AI combining behavioral, genetic, and imaging data with GAN-based augmentation may improve autism spectrum disorder screening, but limited data, class imbalance, and scarce external clinical validation remain

The survey organizes AI-based ASD screening research by data modality into behavioral, genetic, environmental, neuroimaging, physiological, and clinical categories, and discusses approaches for combining information across sources. Compared with prior reviews centered on a single modality or algorithm, it places heterogeneous data sources and multimodal fusion strategies within one framework. This is a narrative synthesis; the abstract lists the modalities and the scope of fusion discussion without reporting sample sizes or effect sizes for individual studies.

The survey specifically examines the role of GANs and related generative approaches in synthetic-data generation and data augmentation to address limited and imbalanced datasets. It treats generative models not merely as generic augmentation but as a targeted response to the data bottleneck in ASD screening. Based on synthesis of existing literature; the text also lists GAN training instability and synthetic-data quality as open issues, indicating the evidence in this direction is not yet uniform.

The survey reviews multimodal fusion strategies, multi-input neural networks, and emerging deep learning approaches in the context of ASD screening. It links architecture-level multi-input design with data-level multimodal integration rather than discussing model structure in isolation. A review-level synthesis; the text provides no specific performance figures or comparison tables.

The survey explicitly lists the field's main challenges: limited and heterogeneous datasets, class imbalance, multimodal data integration, GAN training instability, synthetic-data quality, privacy and security, model interpretability, generalizability, and limited external clinical validation. It presents these issues together as a research agenda pointing toward reliable, explainable, privacy-preserving, and clinically validated multimodal ASD screening systems. The challenge list comes from summarizing reviewed literature and is stated qualitatively, without a quantitative meta-analysis.

The survey is aimed at researchers, algorithm developers, and clinical collaborators interested in ASD screening, and applies to settings where behavioral, genetic, environmental, neuroimaging, physiological, and clinical data need to be combined and where options for fusion strategies and generative data augmentation must be mapped. It also serves as a reference for research directions that take reliability, explainability, privacy preservation, and clinical validation as design goals.

Because the reading scope is incomplete and only the abstract is available, the specific studies cited, dataset sizes, model performance, and comparison results cannot be verified, so the above synthesis stays at the level of scope and themes. Readers should still watch whether multimodal fusion gains hold up in real clinical workflows, whether GAN-generated data introduce bias, and how external clinical validation would be conducted; the text frames these as open questions rather than settled conclusions.