Machine Learning Based Autism Spectrum Disorder Detection from Videos.

Area of Science:

• Developmental Psychology
• Computer Science
• Pediatrics

Background:

• Early diagnosis of Autism Spectrum Disorder (ASD) is critical for effective intervention and improved outcomes.
• Identifying early behavioral markers in infants can significantly aid in timely diagnosis.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) approach for early ASD diagnosis using infant behavioral analysis from videos.
• To identify specific behaviors indicative of ASD in infants aged 6-36 months.

Main Methods:

• A two-stage ML approach was employed, utilizing a dataset of 2000 videos.
• Stage 1: Deep learning models (image-based and facial behavior features) were developed to classify infant behaviors (smile, look face, look object, vocalization).
• Stage 2: Feature selection and resampling techniques were applied to predict ASD diagnosis using statistical behavioral features.

Main Results:

• Behavior classification achieved accuracies of 70% for smile, 68% for look face, 67% for look object, and 53% for vocalization.
• The ML classifier for ASD diagnosis, after feature selection and addressing class imbalance, reached an accuracy of 82%.

Conclusions:

• Machine learning models can effectively identify key infant behaviors associated with ASD.
• This ML approach demonstrates significant potential for accurate, early ASD diagnosis in infants.