Deep learning approach to predict autism spectrum disorder: a systematic review and meta-analysis

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Developmental Pediatrics

Background:

• Deep learning (DL) shows promise for identifying childhood autism spectrum disorder (ASD).
• Existing studies report varying accuracy in DL-based ASD prediction.
• This meta-analysis aims to consolidate findings on DL's classification accuracy for ASD in children.

Purpose of the Study:

• To determine the classification accuracy of deep learning (DL) methods for autism spectrum disorder (ASD) in children.
• To conduct a meta-analysis of existing studies on DL for ASD identification.
• To evaluate the sensitivity, specificity, and AUC of DL techniques in ASD classification.

Main Methods:

• A systematic literature search was conducted across PubMed, EMBASE, Cochrane Library, and Web of Science up to April 16, 2023.
• Eleven predictive trials involving 9495 children were included, with study quality assessed using QUADAS-2.
• Bivariate random-effects models were used to calculate pooled sensitivity, specificity, and AUC.

Main Results:

• The meta-analysis revealed high overall performance for DL in ASD classification: sensitivity 0.95, specificity 0.93, and AUC 0.98.
• Subgroup analysis indicated no significant heterogeneity across different datasets (e.g., Kaggle, ABIDE).
• Specific datasets like ABIDE showed high performance (sensitivity 0.97, specificity 0.97).

Conclusions:

• Deep learning techniques demonstrate satisfactory diagnostic performance for ASD classification.
• Significant heterogeneity among included studies may limit the overall effectiveness of this meta-analysis.
• Further clinical trials are essential to validate the practical application of DL for ASD diagnosis.