A Meta-Review of Computational Intelligence Techniques for Early Autism Disorder Diagnosis

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition affecting social communication and behavior, with global prevalence rising to 1-1.6%. This scoping review evaluates Machine Learning (ML), Deep Learning (DL), and hybrid models for ASD diagnosis using structural and functional MRI data from the ABIDE (Autism Brain Imaging Data Exchange) database. Following, Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) guidelines, we systematically analyzed 512 articles from Google Scholar, PubMed, and ScienceDirect (2019-2024), ultimately selecting 65 studies meeting inclusion criteria. The review revealed hybrid models dominate the field (66% of studies), outperforming standalone ML (8%) and DL (26%) approaches. Pooled mean accuracies were 76.80% (ML), 80.10% (DL), and 82% (hybrid models), with five hybrid models and one DL model exceeding 95% accuracy. Notably, Convolutional neural network (CNN) based hybrid architectures showed superior performance in classifying ASD vs. neurotypical subjects across Magnetic Resonance Imaging (MRI) (8% of studies), functional MRI (fMRI) (32%), and resting state-fMRI (rs-fMRI) (49%) modalities. Key findings demonstrate that integrating CNN with other algorithms (e.g., Support Vector Machine (SVM), Graph Convolutional Networks (GCN), or attention mechanisms) yields the most reliable discrimination of ASD, combining the feature extraction strengths of DL with the interpretability of traditional ML. These advanced models show particular promise for early diagnosis, with low prediction risk. However, publication bias was detected in DL and hybrid model studies (Egger's test p=0.001), suggesting selective reporting of high-accuracy results. This work highlights hybrid models as the current state-of-the-art for neuroimaging-based ASD classification, though standardization challenges remain regarding dataset heterogeneity and model interpretability. Future research should focus on multi-modal integration, clinical validation, and addressing biases in model development.