Automatic Cry Analysis: Deep Learning for Screening of Autism Spectrum Disorder in Early Childhood

Area of Science:

• Developmental Pediatrics
• Computational Linguistics
• Artificial Intelligence in Healthcare

Background:

• Early detection of Autism Spectrum Disorder (ASD) is crucial for timely intervention and improved developmental trajectories.
• Acoustic cry analysis offers a potential non-invasive biomarker for neurodevelopmental conditions.
• Traditional methods for ASD detection can be time-consuming and may lack objectivity.

Purpose of the Study:

• To identify unique acoustic cry characteristics differentiating Typically Developing (TD) children from those with ASD using Deep Learning (DL).
• To develop and validate a DL model for the early, non-invasive detection of ASD based on cry vocalizations.
• To provide clinicians with an AI-powered tool to support early ASD identification.

Main Methods:

• Utilized a cry dataset comprising 31 children with ASD and 31 TD children (18-54 months).
• Performed statistical analysis on acoustic features: jitter, shimmer, and harmonics-to-noise ratio (HNR).
• Developed a Deep Learning model, specifically a Recursive Convolutional Neural Network (R-CNN), for cry classification.

Main Results:

• ASD cries exhibited statistically significant increases in jitter and shimmer compared to TD cries.
• ASD cries showed a significant decrease in harmonics-to-noise ratio (HNR) relative to TD cries.
• The R-CNN model achieved 90.28% accuracy in distinguishing between ASD and TD cries.

Conclusions:

• Cry vocal biomarkers, analyzed via AI, show promise for early ASD detection.
• Automatic, non-invasive AI tools can empower clinicians, facilitating earlier intervention for at-risk children.
• Improved early detection through AI-supported cry analysis can positively impact long-term developmental outcomes for children with ASD.