Challenges in translating AI-driven ASD/ADHD diagnosis: A methodological systematic review.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Computer Vision

Background:

• Pediatric neurodevelopmental disorders (NDDs), including autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD), present diagnostic challenges.
• Current diagnostic methods for NDDs are subjective, time-consuming, and prone to bias.

Purpose of the Study:

• To systematically review biomedical informatics methodologies employing deep learning and computer vision for objective pediatric NDD diagnosis.
• To identify effective strategies for multimodal feature extraction, deep learning architectures, and data integration.

Main Methods:

• Systematic review of 43 Q1/Q2 studies (2020-2024) from Web of Science and Scopus, adhering to PRISMA guidelines.
• Assessment of methodological quality and bias using the APPRAISE-AI framework.
• Focused on four informatics research questions: multimodal features, DL architectures, performance strategies, and data integration.

Main Results:

• Multimodal fusion and hybrid informatics pipelines were prevalent (38%), outperforming unimodal approaches by integrating facial imaging, EEG, and fMRI.
• Transfer learning and fusion techniques were common; federated learning and explainable AI were underutilized.
• APPRAISE-AI indicated strong clinical relevance (72.8%) and reporting quality (66.1%), but significant gaps in reproducibility (41.0%) and robustness (45.1%).

Conclusions:

• AI-driven biomedical informatics shows promise for reducing diagnostic delays and costs for NDDs.
• Improvements in reproducibility, interpretability, and ethical data integration are crucial for clinical deployment.
• Standardized, privacy-preserving, and auditable frameworks are needed for scalable implementation of AI in NDD diagnostics.