Automated Autism Assessment With Multimodal Data and Ensemble Learning: A Scalable and Consistent Robot-Enhanced...

Area of Science:

• Robotics
• Artificial Intelligence
• Developmental Psychology

Background:

• Autism Spectrum Disorder (ASD) diagnosis and intervention face challenges due to variability in therapeutic practices and the need for scalable solutions.
• Standard Human Therapy (SHT) for ASD can be limited by therapist variability and scalability issues.

Purpose of the Study:

• To present a novel Robot-Enhanced Therapy (RET) framework for automating ASD assessment and intervention.
• To leverage advanced computational methods, including 3D biomarkers and saliency maps, for precise ASD diagnosis.

Main Methods:

• Developed an Adaptive Boosted 3D biomarker approach integrated with Kernel Density Estimation-generated Saliency Maps via majority voting.
• Extracted novel features from body skeleton, head movement, and eye gaze data within the DREAM Dataset (61 children).
• Utilized gaze data to generate pioneering saliency maps, enhancing predictive model performance.

Main Results:

• The 3D biomarker approach achieved 95.59% accuracy and 92.75% F1 score for ASD level prediction.
• The same approach yielded an RMSE of 1.78 and R-squared of 0.74 for Autism Diagnostic Observation Schedule (ADOS) score prediction.
• Incorporating gaze-based saliency maps improved ASD level prediction to 97.36% accuracy and 95.56% F1 score.

Conclusions:

• The RET framework demonstrates significant potential to revolutionize ASD management by offering precise, automated assessment.
• RET provides a scalable and consistent alternative to SHT, mitigating therapist variability.
• Future research should address sample size and model generalizability to further validate RET's impact.