Data-Driven Approaches for Autism Detection: A Comprehensive Review of Machine Learning Algorithms and Datasets
Anupama N1, Chandrashekar M Patil1
1Department of Electronics and Communication Engineering, Vidyavardhaka College of Engineering, Mysuru, Visvesvaraya Technological University, Belagavi, Karnataka, India.
Summary
Machine learning and deep learning show promise for autism spectrum disorder (ASD) detection, with multimodal approaches achieving high accuracy. However, studies need larger datasets, reduced bias, and external validation for clinical use.
Area of Science:
- Neuroscience
- Artificial Intelligence
- Medical Diagnostics
Background:
- Autism spectrum disorder (ASD) diagnosis is complex due to symptom variability.
- Machine learning (ML) and deep learning (DL) offer potential for automated ASD screening.
- Existing research spans various data modalities but faces methodological challenges.
Purpose of the Study:
- To systematically review unimodal and multimodal ML/DL approaches for ASD detection.
- To analyze study methodologies, performance, and limitations.
- To identify pathways for clinical translation of ASD detection systems.
Main Methods:
- Systematic review of 59 peer-reviewed studies (2019-2025) on ML/DL for ASD detection.
- Analysis of diverse data modalities: behavioral, neuroimaging, EEG, eye-tracking, speech.
- Evaluation of classical ML (LR, SVM, RF) and DL (CNN, RNN, Transformers) models.
Main Results:
- ML/DL models achieved diagnostic accuracies from 68% to 99%.
- Hybrid multimodal frameworks outperformed unimodal ones (95%-99% accuracy).
- Common flaws identified: small sample sizes, bias, overfitting, lack of external validation.
Conclusions:
- Large, balanced datasets and explainable AI (XAI) are crucial for ASD detection systems.
- Standardization (e.g., BIDS) and regulatory guidelines are needed for clinical translation.
- Multimodal fusion strategies require careful architectural comparison for clinical applicability.
